Facts

Speech recognition: facts and numbers

436 facts, each traced to the exact words of its source.

10.5 percent word error rateA 2024 study that fine-tuned Whisper version 3 on about 68 hours of Northern Kurdish (Kurmanji) speech reached a word error rate of 10.5 percent on a specialized test set, using its best fine-tuning strategy.Abdulhady Abas Abdullah et al., arXiv, 2024
A 2025 study found that a substantial amount of the LibriSpeech and Common Voice evaluation sets appear in public large language model pretraining corpora.Yuan Tseng et al., arXiv (Audio and Speech Processing), 2025
A 2025 study of Whisper found that when non-speech sounds were played to the model, a set of hallucinated phrases appeared frequently.Mateusz Barański et al., arXiv (accepted for IEEE ICASSP 2025), 2025
13.88 percent word error rateA 2026 study says published Swiss German speech recognition results are inflated by benchmark contamination, because a Whisper model trained on the test set with no Swiss German data reached 13.88 percent WER.Felix Akeret, arXiv (Computation and Language), 2026
A developer's January 2019 issue in Brave's GitHub tracker says the Web Speech API is disabled in Brave from returning results, so an app cannot tell which speech service to use.bsclifton, GitHub, Brave browser issue tracker, 2019
10 minutesA Gemini live transcription session supports continuous streaming for up to 10 minutes.Google, Google AI for Developers, 2026
1000 word vocabularyA group of scientists recommended in 1971 a five-year program whose target was a system that accepts connected speech from many speakers on a 1000-word task vocabulary.Raj Reddy, NASA Technical Reports Server (NTRS), 1993
960000 samples per minuteA recording at 16,000 samples per second contains 960,000 samples for each minute of audio.Calculated from other facts
8 kHzA recording sampled at 16 kHz can represent sound frequencies up to 8 kHz, half the sampling rate.Calculated from other facts
1.4 percent word error rateAccording to the Whisper paper, the best word error rate on LibriSpeech test-clean fell from 5.3 percent in 2015 to 1.4 percent by 2021.Alec Radford et al., OpenAI (arXiv), 2022
After OpenAI's October 2025 recapitalization, the nonprofit, renamed the OpenAI Foundation, retains control of the for-profit entity, now called OpenAI Group PBC.Converge Digest, Converge Digest, 2025
27 percentAfter the October 2025 restructuring, Microsoft's investment in OpenAI Group PBC represented roughly 27 percent ownership, on an as-converted diluted basis.Converge Digest, Converge Digest, 2025
60 audio minutes per monthAmazon Transcribe's free tier covers up to 60 audio minutes a month, for the first 12 months after signup.Amazon Web Services, 2026
0.006 US dollars per minuteAmazon's pricing page gives a batch transcription price of $0.006 per minute in its worked example for the US East (N. Virginia) region.Amazon Web Services, 2026
Amazon's pricing page says its base price includes speaker diarization, along with custom vocabularies, vocabulary filtering and language identification.Amazon Web Services, 2026
Android's recognizer intent documents a silence length that is the time after the recognizer stops hearing speech before the input is treated as complete and the session ends.Google, Android Developers, n.d.
Android's recognizer intent says its possibly-complete silence length exists to stop the endpointer cutting off the input during very short pauses in the middle of speech.Google, Android Developers, n.d.
Appen defines benchmaxxing as optimizing models specifically to score well on public test sets without equivalent gains in real-world performance.Appen, Appen (press release), 2026
Appen provides a suite of new, private English-language audio datasets for the Open ASR Leaderboard.Appen, Appen (press release), 2026
38 trillion operations per secondApple says the Neural Engine in its M4 chip can reach up to 38 trillion operations per second, faster than the neural processing unit of any AI PC at the time.Apple, Apple Newsroom, 2024
Artificial Analysis defines time to first partial transcript as the time of the first partial transcript shared after the Silero voice activity detector detects end of speech.Artificial Analysis, Artificial Analysis, n.d.
Artificial Analysis says a speed factor above 1 means the service transcribes faster than real time, so 2.0 means 10 minutes of audio is transcribed in 5 minutes.Artificial Analysis, Artificial Analysis, n.d.
Artificial Analysis says streaming speech-to-text performance varies meaningfully across its three test datasets, which differ in audio length, accents, vocabulary and background noise.Artificial Analysis, Artificial Analysis, 2026
8 hours of audioArtificial Analysis tested its streaming speech-to-text models on about 8 hours of audio, the same audio as its offline benchmark.Artificial Analysis, Artificial Analysis, 2026
74 open-source modelsAs of 27 March 2026, the Open ASR Leaderboard listed 86 speech recognition models, and 74 of them were open-source.Hugging Face and collaborators, arXiv, 2025
0.15 US dollars per hour of audioAssemblyAI lists its Universal-2 speech-to-text model at $0.15 per hour of audio on pay as you go.AssemblyAI, 2026
0.21 US dollars per hour of audioAssemblyAI lists its Universal-3.5 Pro model at $0.21 per hour of audio on pay as you go.AssemblyAI, 2026
185 hours of pre-recorded transcriptionAssemblyAI's pricing page says its free tier includes up to 185 hours of pre-recorded transcription and up to 333 hours of streaming transcription.AssemblyAI, 2026
0.36 US dollars per hour of audioAt Amazon's worked-example batch price, one hour of audio costs about $0.36.Calculated from other facts
0.11 US dollarsAt ElevenLabs' list price of 0.22 dollars per hour, a 30-minute recording costs about 0.11 dollars.Calculated from other facts
0.96 US dollars per hour of audioAt Google Cloud's V1 list price with data logging, one hour of audio costs about $0.96 once the first 60 free minutes each month are used.Calculated from other facts
1.44 US dollars per hour of audioAt Google Cloud's V1 list price without data logging, one hour of audio costs about $1.44 once the first 60 free minutes each month are used.Calculated from other facts
0.18 US dollars per hour of audioAt OpenAI's listed price, one hour of audio transcribed with gpt-4o-mini-transcribe costs about $0.18.Calculated from other facts
0.96 US dollars per hour of audioAt the first-tier V2 standard list price, one hour of audio costs about $0.96.Calculated from other facts
0.18 US dollars per hour of audioAt the list price for gpt-4o-mini-transcribe, one hour of audio costs about $0.18.Calculated from other facts
0.36 US dollars per hour of audioAt the list price for gpt-4o-transcribe, one hour of audio costs about $0.36.Calculated from other facts
1.44 US dollars per hour of audioAt the V1 list price without data logging, one hour of audio costs about $1.44.Calculated from other facts
5 audio hours per monthAzure AI Speech's free (F0) tier includes 5 audio hours a month of standard real-time transcription.Microsoft Azure, 2026
60 secondsAzure's REST API for short audio accepts no more than 60 seconds of audio in a request when the audio is sent directly.Microsoft Learn (Azure AI Speech), n.d.
Azure's REST API for short audio does not support batch transcription or custom speech.Microsoft Learn (Azure AI Speech), n.d.
Azure's REST API for short audio requires the audio to be in one of the formats its page lists, which is a narrower set than the Speech SDK accepts.Microsoft Learn (Azure AI Speech), n.d.
Azure's REST API for short audio returns only final results, without partial text while the speaker is still talking.Microsoft Learn (Azure AI Speech), n.d.
Azure's REST speech to text API accepts your Speech resource key in the Ocp-Apim-Subscription-Key header.Microsoft Learn (Azure AI Speech), n.d.
Azure's REST speech to text API needs a language parameter in the URL, or the request can fail with a 4xx error.Microsoft Learn (Azure AI Speech), n.d.
Buzz's README lists live realtime transcription from a microphone among its features, alongside transcription of audio and video files and YouTube links.Chidi Williams, GitHub (chidiwilliams/buzz), 2026
By 2012, the performance of hybrid systems, which pair neural networks with HMMs, had surpassed traditional HMM/GMM systems.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
2500 hoursBy the time of its 2019 paper, Mozilla's Common Voice corpus held 2,500 hours of audio from over 50,000 contributors, across 38 languages collecting data.Rosana Ardila et al., arXiv (LREC 2020), 2020
1011 word vocabularyCarnegie Mellon's Harpy system recognized speech from male and female speakers on a 1,011-word document retrieval task, with a semantic error of 5 percent.Raj Reddy, NASA Technical Reports Server (NTRS), 1993
Deepgram defines end-of-turn latency as the time from when a user stops speaking to when an end-of-turn event is received.Deepgram, Deepgram, n.d.
Deepgram defines transcript latency as how far behind the transcription is from the audio being sent.Deepgram, Deepgram, n.d.
300 milliseconds (vendor claim, streaming latency)Deepgram says its Nova-3 model delivers sub-300 ms streaming latency.Deepgram, Deepgram, n.d.
Deepgram's documentation says its diarization recognizes speaker changes and assigns a speaker to each word in the transcript.Deepgram, n.d.
Deepgram's documentation says to use the diarize_model parameter for speaker diarization in both batch and streaming requests, and that the older diarize parameter is deprecated.Deepgram, n.d.
200 US dollars of free creditDeepgram's pricing page lists a free plan with a $200 credit, followed by pay-as-you-go billing.Deepgram, 2026
Distil-Whisper, Hugging Face's distilled version of Whisper, is MIT licensed, so it can be used for commercial applications.Hugging Face, GitHub (huggingface/distil-whisper), n.d.
119 hoursEarnings-22 is a corpus of 125 files totaling 119 hours of English earnings calls from companies around the world.Miguel Del Rio et al., arXiv (Computation and Language), submitted to Interspeech 2022, 2022
ElevenLabs bills its speech to text service by how long the audio is that you send for transcription.ElevenLabs, 2026
0.39 US dollars per hour of audioElevenLabs lists its Scribe v2 Realtime model at $0.39 per hour of audio.ElevenLabs, 2026
0.22 US dollars per hour of audioElevenLabs lists its Scribe v2 speech-to-text model at $0.22 per hour of audio.ElevenLabs, 2026
0.050 US dollars per hour of audioElevenLabs lists keyterm prompting for its Scribe v2 speech-to-text model at $0.050 per hour of audio, added to the $0.22 per hour base rate.ElevenLabs, 2026
0.070 US dollars per hour of audioElevenLabs lists transcript editing for its Scribe v2 speech-to-text model at $0.070 per hour of audio, on top of the base rate.ElevenLabs, 2026
ElevenLabs says its Scribe speech-to-text API returns structured JSON transcripts with speaker diarization and word-level timestamps.ElevenLabs, 2025
ElevenLabs says its Scribe v2 Medical model is HIPAA-eligible, with business associate agreements and a Zero Retention Mode.ElevenLabs, 2026
ElevenLabs says keyterm prompting is available with Scribe v2, Scribe v2 Medical in batch mode, and Scribe v2 Realtime.ElevenLabs, 2026
ElevenLabs says Scribe outperforms Gemini 2.0 Flash, Whisper Large V3 and Deepgram Nova-3 on the FLEURS and Common Voice benchmarks across 99 languages.ElevenLabs, 2025
98 percent transcription accuracy, vendor claimElevenLabs says Scribe v2 has over 98% transcription accuracy. The pricing page gives no test set or method for that figure.ElevenLabs, 2026
1000 termsElevenLabs says Scribe v2 supports keyterm prompting with up to 1,000 terms.ElevenLabs, 2026
3 GBElevenLabs says that files up to 3 GB in size are supported by its speech to text API.ElevenLabs, 2026
ElevenLabs says that keyterm prompting and transcript editing come at an additional cost, on top of the base transcription rate.ElevenLabs, 2026
ElevenLabs says you create an API key in its dashboard.ElevenLabs, 2026
ElevenLabs' speech to text API has a model_id field that names the transcription model, and its examples set it to scribe_v2.ElevenLabs, 2026
300 speech commandsEspressif says its MultiNet model recognizes up to 300 Chinese or English speech commands, and that developers can add commands without training the model again.Espressif Systems, Espressif Systems, ESP Component Registry, 2026
Espressif says its WakeNet9s wake word model can run on chips that have no PSRAM and no SIMD support, such as the ESP32-C3 and ESP32-C5.Espressif Systems, Espressif Systems, ESP Component Registry, 2026
Espressif says that from ESP-SR v2.6.0, its WakeNet10s wake word model can run on chips without SIMD instructions, including the original ESP32.Espressif Systems, Espressif Systems, ESP Component Registry, 2026
Espressif says the ESP32-S3 has vector instructions that accelerate neural network computing and signal processing workloads.Espressif Systems, 2026
Espressif's esp-sr README lists the ESP32-S3, ESP32-P4 and ESP32-S31 as the supported targets for its MultiNet speech command recognition, and does not list the original ESP32.Espressif Systems, Espressif Systems, ESP Component Registry, 2026
faster-whisper is a reimplementation of OpenAI's Whisper model that runs on the CTranslate2 inference engine.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
4 times faster, at mostfaster-whisper is up to 4 times faster than openai/whisper for the same accuracy, while using less memory.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
For OCR evaluation, character error rate is based on Levenshtein distance: the minimum number of character substitutions, deletions and insertions needed to turn the ground truth text into the OCR output, divided by the number of characters in the ground truth.Nevidu Jayatilleke, arXiv, 2025
10000 hoursGigaSpeech is an English speech recognition corpus with 10,000 hours of high quality labeled audio, covering several domains.Guoguo Chen et al., arXiv, 2021
0.016 US dollars per minuteGoogle Cloud lists Speech-to-Text V2 standard recognition at $0.016 per minute for the first 500,000 minutes each month.Google Cloud, 2026
Google Cloud requires streaming audio to be sent at about the speed at which it is spoken, and it returns an error for content that exceeds the limits.Google Cloud, Google Cloud, n.d.
Google Cloud says its streaming recognition returns results in real time as the audio is processed.Google Cloud, Google Cloud, n.d.
Google Cloud says streaming speech recognition is available through gRPC only.Google Cloud, Google Cloud, n.d.
Google Cloud sends a speech activity end event when it detects that speech has ended, and no such event is sent if the stream closes first.Google Cloud, Google Cloud, n.d.
Google Cloud Speech-to-Text's speaker diarization detects when speakers change and labels each voice with a number, and it tags each word with the speaker number.Google Cloud, 2026
Google Cloud's API reference says interim results are tentative hypotheses that can be returned before the final result, and that they are marked with is_final set to false.Google Cloud, Google Cloud (Ruby client reference), n.d.
300 US dollars in free creditsGoogle Cloud's pricing page advertises a free program with $300 in free credits and more than 20 free products.Google Cloud, 2026
5 minutesGoogle Cloud's quota table gives a streaming recognition request an audio length of about 5 minutes.Google Cloud, Google Cloud, n.d.
480 minutes (asynchronous request audio length)Google Cloud's quota table gives an asynchronous request an audio length of about 480 minutes, against about 1 minute for a synchronous request.Google Cloud, Google Cloud, n.d.
0.016 US dollars per minuteGoogle Cloud's V1 list price for standard recognition with data logging is $0.016 per minute after the first 60 minutes each month.Google Cloud, 2026
0.024 US dollars per minuteGoogle Cloud's V1 list price for standard recognition without data logging is $0.024 per minute after the first 60 minutes each month.Google Cloud, 2026
60 free minutes per month per accountGoogle Cloud's V1 price table lists 0 to 60 minutes of standard recognition each month as free, per account.Google Cloud, 2026
500 milliseconds (minimum voice activity timeout)Google Cloud's voice activity timeouts must be greater than 500 ms and less than 60 seconds.Google Cloud, Google Cloud, n.d.
Google describes Gemini 3.5 Transcribe as its speech-to-text model, with automatic language detection, speaker diarization, word-level timestamps and custom vocabulary biasing.Google, Google AI for Developers, 2026
0.003 US dollars per minuteGoogle lists dynamic batch recognition for standard models at $0.003 per minute, with lower rates in the same row.Google Cloud, 2026
Google said in 2019 that its new on-device recognizer design offered the same accuracy as its traditional server-based models in a much smaller model.Johan Schalkwyk, Google Research Blog, 2019
Google said its 2019 on-device Gboard speech recognizer uses a model trained with RNN transducer technology that is compact enough to reside on a phone.Johan Schalkwyk, Google Research Blog, 2019
2 GBGoogle said its traditional server-side recognizer used a search graph of almost 2 GB, too large for a phone, which is why it needed a connection.Johan Schalkwyk, Google Research Blog, 2019
12 hours (maximum)Google says a notebook on the free version of Colab can run for at most 12 hours, depending on availability and usage.Google Research, n.d.
Google says a standard API key can be used with any API that accepts API keys, unless the key has API restrictions.Google Cloud, 2026
Google says Cloud Speech-to-Text does not log customer audio or transcripts unless the customer opts into data logging.Google Cloud, 2026
Google says Colab provides free of charge access to computing resources, including GPUs and TPUs.Google Research, n.d.
Google says Colab resources are not guaranteed and not unlimited, and that usage limits sometimes change.Google Research, n.d.
Google says its RNN-T recognizer continuously processes input audio and streams output symbols, a property that suits speech dictation, and in its implementation the symbols are characters.Johan Schalkwyk, Google Research Blog, 2019
1 minute of audio per synchronous requestGoogle says synchronous recognition requests process audio of 1 minute or less.Google Cloud, 2026
Google says that in the free version of Colab, access to expensive resources such as GPUs is heavily restricted.Google Research, n.d.
Google says that not all Google Cloud APIs accept API keys to authorize usage, so you must check the documentation of the service you want to call.Google Cloud, n.d.
Google's API key page says to add at least one restriction when you create an API key.Google Cloud, 2026
Google's API key page tells you to start in the Google Cloud console, on the Credentials page, to create an API key.Google Cloud, 2026
32 tokens per second of audioGoogle's audio guide says Gemini models use 32 tokens for each second of audio, which is 1,920 tokens per minute.Google, Google AI for Developers, 2026
Google's Cloud Speech-to-Text documentation says the service detects words in an audio clip by comparing the input with one of many machine learning models.Google Cloud, 2026
Google's documentation says speaker diarization is not available for every language, and tells you to check its language support page.Google Cloud, 2026
Google's general Gemini audio guide says that for dedicated speech-to-text models with real-time transcription, you should use the Google Cloud Speech-to-Text API.Google, Google AI for Developers, 2026
Google's help says the default text-to-speech engine choices vary by device, and can include Google's engine, the device maker's engine and third-party engines downloaded from Google Play.Google, Google Accessibility Help, n.d.
Google's local-file example sets the language to en-US and the model to chirp_3.Google Cloud, 2026
25 audio tokens per secondGoogle's pricing notes estimate the cost of Gemini 3.5 Transcribe at 25 audio tokens per second of input.Google, Google AI for Developers, 2026
Google's Python example for synchronous recognition of a local file imports the Speech-to-Text v2 client.Google Cloud, 2026
Google's Python reference gives pip install google-cloud-speech as the install command on Mac and Linux.Google Cloud, 2026
Google's quickstart asks you to check that billing is enabled for your Google Cloud project before you start.Google Cloud, 2026
Google's quickstart says to initialize the gcloud CLI with the command gcloud init.Google Cloud, 2026
Google's quickstart tells you to enable the Speech-to-Text APIs in your project if they are not already enabled.Google Cloud, 2026
Google's Speech-to-Text authentication page says the REST interface can be called with gcloud CLI credentials or Application Default Credentials.Google Cloud, n.d.
0.004 US dollars per minuteGoogle's V2 standard table lists $0.004 per minute for monthly volumes of 2,000,000 minutes and above.Google Cloud, 2026
Graves and colleagues wrote that recurrent networks needed pre-segmented training data and post-processing to produce label sequences, and presented a method that trains them to label unsegmented sequences directly.Alex Graves et al., Proceedings of the 23rd International Conference on Machine Learning (ICML 2006), pp. 369-376, 2006
Graves wrote in 2012 that recurrent networks traditionally required a pre-defined alignment between input and output sequences, and introduced an end-to-end transduction system that removes that need.Alex Graves, arXiv (ICML 2012 Workshop on Representation Learning), 2012
17.7 percent test set errorGraves, Mohamed and Hinton report that deep long short-term memory recurrent networks, trained end-to-end with suitable regularisation, reached a 17.7% test set error on the TIMIT phoneme recognition benchmark.Alex Graves et al., arXiv (accepted for ICASSP 2013), 2013
Guthrie and Mackersie write that the most popular way to set the presentation level for a word recognition test is a set number of decibels above the speech recognition threshold.Leslie A. Guthrie and Carol L. Mackersie, Journal of the American Academy of Audiology, 2009
He and colleagues report that their RNN transducer can outperform a conventional CTC-based model in both latency and accuracy across a number of evaluation categories.Yanzhang He et al., arXiv (ICASSP 2019), 2019
17,000 ASR modelsHugging Face says its Hub contains more than 17,000 automatic speech recognition models, and that you can test them in your browser.Hugging Face, Hugging Face, n.d.
Hugging Face's Audio Course describes speech recognition, also called automatic speech recognition (ASR) or speech-to-text (STT), as one of the most popular and exciting spoken language processing tasks.Hugging Face, n.d.
Hugging Face's transformers documentation shows how to use its pipeline function across several tasks, and speech recognition is one of them.Hugging Face, Hugging Face, n.d.
200 transistors per wordIBM says earlier speech recognition machines required as many as 200 transistors per word.IBM, IBM, 2026
1961 yearIBM says engineer William C. Dersch introduced its experimental Shoebox machine at a press conference in 1961.IBM, IBM, 2026
1962 yearIBM says Shoebox appeared on the Today show and at the 1962 Seattle World's Fair after its 1961 press conference.IBM, IBM, 2026
6 command wordsIBM says Shoebox was capable of understanding the numbers zero through nine and six command words, including plus, minus, total and subtotal.IBM, IBM, 2026
31 transistorsIBM says Shoebox's circuitry required only 31 transistors.IBM, IBM, 2026
500 minutes a monthIBM's Lite plan for Watson Speech to Text includes 500 minutes of free speech recognition a month, with 38 pre-trained speech models.IBM, 2026
10.3 percent word error rateIBM's own English conversational telephone recognizer reached a word error rate of 10.3 percent on the CallHome subset of the Hub5 2000 evaluation.George Saon et al., arXiv (cs.CL), IBM Research, 2017
5.5 percent word error rateIBM's own English conversational telephone recognizer reached a word error rate of 5.5 percent on the Switchboard subset of the Hub5 2000 evaluation.George Saon et al., arXiv (cs.CL), IBM Research, 2017
100 concurrent transcriptionsIBM's Plus plan for Watson Speech to Text includes unlimited minutes a month and 100 concurrent transcriptions.IBM, 2026
IEEE Technology Navigator defines speaker recognition as identifying or verifying a person's identity from the characteristics of their voice, and distinguishes it from speech recognition.IEEE Technology Navigator, n.d.
IEEE Technology Navigator says speaker recognition must contend with spoofing attacks, which can include synthesizing speech with text-to-speech tools.IEEE Technology Navigator, n.d.
IEEE Technology Navigator says that in speaker identification no identity is claimed, and the system searches a database of known speakers for the closest match.IEEE Technology Navigator, n.d.
IEEE Technology Navigator says that in speaker verification the system accepts or rejects a claimed identity, by checking whether the voice matches stored models for the person claimed.IEEE Technology Navigator, n.d.
1952 yearIn 1952, Davis, Biddulph and Balashek of Bell Laboratories built a system that recognized isolated spoken digits from a single speaker.B. H. Juang and Lawrence R. Rabiner, Elsevier Encyclopedia of Language and Linguistics (author copy, UC Santa Barbara), 2005
80 MBIn 2019 Google moved Gboard voice typing onto the phone with an 80 MB speech recognizer that works even when the phone is offline, first on Pixel phones in American English.Johan Schalkwyk, Google Research Blog, 2019
In 2022, von Neumann and colleagues proposed a general framework for computing WER when a recognizer processes recordings with several speakers and outputs several word sequences.Thilo von Neumann et al., arXiv (eess.AS), 2022
12.8 percent word error rateIn a 1997 review of noise experiments, a machine with noise adaptation reached 12.8 percent error at a signal-to-noise ratio of 10 dB, more than ten times the 1 percent human error rate.Richard P. Lippmann, Speech Communication, vol. 22, pp. 1-15, 1997
In a 1997 review of six speech corpora, machine error rates were often more than ten times higher than human error rates on quiet, wideband, read speech.Richard P. Lippmann, Speech Communication, vol. 22, pp. 1-15, 1997
43 percent word error rateIn a 1997 review, an HMM recognizer had a word error rate of 43 percent on the spontaneous telephone conversations of the Switchboard corpus.Richard P. Lippmann, Speech Communication, vol. 22, pp. 1-15, 1997
18.5 percent word error rateIn a 2012 review, a deep neural network reduced the word error rate of a GMM-HMM baseline from 27.4 to 18.5 percent on a Switchboard telephone task.Geoffrey Hinton et al., IEEE Signal Processing Magazine, 2012
In a 2017 comparison on conversational telephone speech, the most frequent substitution, deletion and insertion errors of a recognizer and of a commercial transcription vendor showed a high degree of overlap.Andreas Stolcke and Jasha Droppo, Proceedings of Interspeech 2017, ISCA, 2017
15.8 to 2.2 percent word diarization error rateIn a 2019 Google study, a joint speech recognition and speaker diarization model improved word-level diarization error rate from 15.8 percent to 2.2 percent.Laurent El Shafey et al., Interspeech 2019, 2019
In a 2024 study of children's classroom dialogue, Google Speech-to-Text had the lowest mean error rate of the three engines compared and the lowest variation, which the authors describe as the most robust to noise.Yingbo Ma et al., arXiv, 2024
20.1 percent average word error rateIn a 2024 study of eleven speech recognition services on university lecture recordings, Google's average word error rate on English was 20.1 percent.Korbinian Kuhn et al., arXiv (preprint), 2024
0.142 character error rateIn a 2025 study of 200 verses from Juz 30 of the Quran, Whisper reached an average character error rate of 0.142, which the authors describe as about 85 percent transcription accuracy.Azzam Abdullah et al., Bulletin of Computer Science Research, 2025
6.77 percent word error rateIn a 2026 preprint, the reported word error rate was 6.77 percent for ElevenLabs and 5.05 percent for Whisper large, on the paper's common benchmark table.Marcin Pietroń et al., Pietron et al., 2026
5.33 percent average WERIn a Benchmarklist mirror of the Open ASR Leaderboard, dated May 6, 2026, IBM Granite Speech 4.1 2B had the lowest average word error rate among open models, at 5.33 percent, 0.09 points ahead of Cohere Transcribe 03-2026.Benchmarklist, Benchmarklist, 2026
In a hybrid architecture, a feedforward neural network's outputs are used as the probability estimates for an HMM-based architecture.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
In a June 2025 9to5Mac report, Wispr Flow's intelligence was described as based on a mix of Meta's Llama and OpenAI's Whisper models.9to5Mac, 9to5Mac, 2025
5 dB below uncomfortable loudness levelIn a study of 40 adults with sensorineural hearing loss, Guthrie and Mackersie found that for people with moderately severe or severe losses, presenting words 5 dB below their uncomfortable loudness level gave the highest phoneme recognition scores.Leslie A. Guthrie and Carol L. Mackersie, Journal of the American Academy of Audiology, 2009
5.1 percent word error rate (final transcript)In Artificial Analysis's 2026 streaming benchmark, OpenAI's hosted GPT Realtime Whisper model improved from a 7.5 percent word error rate on the first partial transcript to 5.1 percent on the final transcript.Artificial Analysis, Artificial Analysis, 2026
7.4 percent word error rate (Deepgram Flux, first partial)In Artificial Analysis's 2026 streaming benchmark, the first partial transcript from Deepgram Flux had a word error rate of 7.4 percent, while ElevenLabs Scribe v2 Realtime's had 3.6 percent.Artificial Analysis, Artificial Analysis, 2026
0.019 seconds (time to first partial after end of speech)In Artificial Analysis's June 2026 streaming benchmark, Deepgram Flux had the fastest first partial transcript after end of speech, at 0.019 seconds.Artificial Analysis, Artificial Analysis, 2026
0.020 secondsIn Artificial Analysis's June 2026 streaming benchmark, Deepgram Flux was the fastest service on both the first partial and the first final transcript after end of speech, at 0.020 seconds for the final transcript, with a 7.36 percent word error rate.Artificial Analysis, Artificial Analysis, 2026
0.5 real-time factorIn Microsoft's worked example, a device that processes a one-minute audio file in 30 seconds has a real-time factor of 0.5.Microsoft Learn, n.d.
In openai-whisper, the language is left unset by default, and the program then detects the spoken language itself.OpenAI, GitHub (openai/whisper), 2026
0.6 no-speech probability thresholdIn OpenAI's open-source Whisper command line, the --no_speech_threshold option defaults to 0.6.OpenAI, GitHub (openai/whisper), 2026
75 percentIn tests of word boundary marking by native Mandarin speakers, their agreement on where the word boundaries fall was about 75 percent, so the authors use character error rate as the usual metric for Mandarin speech recognition.Xunying Liu et al., Journal of the Acoustical Society of America, 2013
2048 units per hidden layerIn the 2012 Switchboard experiment, a seven-hidden-layer deep network with 2,048 units in each layer replaced the Gaussian mixture model in the acoustic model.Geoffrey Hinton et al., IEEE Signal Processing Magazine, 2012
In the 2019 paper, a word counts toward word diarization error rate when it is a correctly recognized word with the wrong speaker label, or a substituted word with a wrong speaker label.Laurent El Shafey et al., Interspeech 2019, 2019
In the 2024 Mandarin stuttering challenge, character error rate counts the minimum number of insertions, substitutions and deletions needed to turn the recognizer output into the reference.Hongfei Xue et al., arXiv (eess.AS), 2024
3.9 timesIn the 2024 study of university lecture recordings, IBM's average English word error rate was about 3.9 times that of Whisper large-v2.Calculated from other facts
In the 2024 study, Google had the highest mean word error rate of the eleven services, with the largest spread between recordings.Korbinian Kuhn et al., arXiv, 2024
In the 2025 study, LLM-based recognizers from contaminated language models showed only subtle error rate differences, but assigned much higher probability to transcriptions they had seen during training.Yuan Tseng et al., arXiv (Audio and Speech Processing), 2025
2.3 percent AA-WERIn the Artificial Analysis AA-WER v2.0 benchmark, ElevenLabs Scribe v2 had the lowest error rate at 2.3 percent, ahead of Google's Gemini 3 Pro at 2.9 percent.Artificial Analysis, 2026
2.3 percent AA-WERIn the Artificial Analysis AA-WER v2.0 key results, ElevenLabs Scribe v2 leads at 2.3 percent, followed by Gemini 3 Pro at 2.9, Voxtral Small at 3.0, Gemini 2.5 Pro at 3.1 and Gemini 3 Flash at 3.2 percent.Artificial Analysis, 2026
418 secondsIn the faster-whisper CPU benchmark, openai/whisper took 6 minutes 58 seconds (418 seconds) with the small model in fp32 precision.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
143 secondsIn the faster-whisper GPU benchmark, openai/whisper took 2 minutes 23 seconds (143 seconds) with the large-v2 model in fp16 precision.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
2926 MB of VRAMIn the faster-whisper README's GPU benchmark, faster-whisper in int8 precision used 2926 MB of GPU memory with the large-v2 model.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
5.1 percent word error rateIn the IBM study, the best of three human transcribers reached a word error rate of 5.1 percent on the Switchboard telephone test after quality checks.George Saon et al., arXiv (cs.CL), IBM Research, 2017
-100 percentIn the Morris and colleagues example of a system that outputs two wrong words for each input word, word accuracy computed as 1 minus WER is -100 percent.Calculated from other facts
480000 samples per 30-second chunkIn the open-source Whisper code, a chunk of 30 seconds at a sample rate of 16,000 per second is 480,000 samples.OpenAI, GitHub, openai/whisper repository (main branch), 2022
4 percent word error rateIn the same 1997 review, human transcribers had an error rate of 4 percent on spontaneous telephone conversation, more than ten times lower than the 43 percent of the best machine recognizer.Richard P. Lippmann, Speech Communication, vol. 22, pp. 1-15, 1997
6.9 times the word error rateIn the same 2024 study of university lecture recordings, Google's average English word error rate was about 6.9 times that of Whisper large-v2.Calculated from other facts
2.9 percent average word error rateIn the same 2024 study, Whisper large-v2 had an average word error rate of 2.9 percent on the English lecture recordings.Korbinian Kuhn et al., arXiv (preprint), 2024
12.3 percent word error rateIn the same 2025 9to5Mac test of one 7-minute-31-second episode, NVIDIA's Parakeet v2 had a word error rate of 12.3 percent, higher than Apple's 10.3 percent.Marcus Mendes, 9to5Mac, 2025
0.47 seconds (time to first partial after end of speech)In the same Artificial Analysis benchmark, AssemblyAI U3 Realtime Pro had a first partial transcript 0.47 seconds after end of speech, at 4.46 percent word error rate.Artificial Analysis, Artificial Analysis, 2026
In the same Artificial Analysis benchmark, ElevenLabs Scribe v2 leads on two of three datasets, and Google's Gemini 3 Pro leads on the parliamentary VoxPopuli dataset.Artificial Analysis, 2026
0.13 seconds (time to first partial after end of speech)In the same Artificial Analysis benchmark, ElevenLabs Scribe v2 Realtime had the highest accuracy on the first partial after end of speech, at 3.65 percent word error rate and 0.13 seconds of latency.Artificial Analysis, Artificial Analysis, 2026
102 secondsIn the same CPU benchmark, faster-whisper with int8 precision took 1 minute 42 seconds (102 seconds) with the small model.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
125 secondsIn the same CPU benchmark, whisper.cpp took 2 minutes 5 seconds (125 seconds) with the small model in fp32 precision.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
6.8 percent word error rateIn the same IBM study, the best human transcriber reached a word error rate of 6.8 percent on the CallHome telephone test.George Saon et al., arXiv (cs.CL), IBM Research, 2017
In the same June 2025 report, Wispr Flow told 9to5Mac that it may bring in other model providers as it expands its features.9to5Mac, 9to5Mac, 2025
In the whisper.cpp streaming example, setting the step argument to 0 switches the program to a sliding-window mode that transcribes after voice activity is detected.ggml-org, GitHub (ggml-org/whisper.cpp), n.d.
Inverse text normalization converts the spoken-domain output of a speech recognizer into written-domain text, to make the recognizer's output easier to read.Yang Zhang et al., arXiv, 2021
150 milliseconds (mouth-to-ear delay)ITU-T Recommendation G.114 says that if mouth-to-ear speech delay can be kept below 150 ms, most speech and non-speech applications experience essentially transparent interactivity.International Telecommunication Union, ITU-T, 2003
4 times smaller memory footprintJacob and colleagues report that storing both activations and weights as 8-bit integers gives a close to 4x memory footprint reduction compared with 32-bit floating-point numbers.Benoit Jacob et al., IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2018, 2018
Juang and Rabiner write in 2005 that most practical speech recognition systems were based on statistical methods developed in the 1980s.B. H. Juang and Lawrence R. Rabiner, Elsevier Encyclopedia of Language and Linguistics (author copy, UC Santa Barbara), 2005
Juang and Rabiner write that Hearsay-II and BBN's HWIM did not meet the DARPA program's performance goal when it ended in 1976.B. H. Juang and Lawrence R. Rabiner, Elsevier Encyclopedia of Language and Linguistics (author copy, UC Santa Barbara), 2005
Juang and Rabiner write that IBM and AT&T Bell Laboratories were two different schools of thought on commercial speech recognition in the 1970s, and that IBM's effort was led by Fred Jelinek.B. H. Juang and Lawrence R. Rabiner, Elsevier Encyclopedia of Language and Linguistics (author copy, UC Santa Barbara), 2005
Jurafsky and Martin's draft says encoder-decoder speech models implicitly learn a language model from their training data.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
Kacprzak and Fraś conclude that without number normalization, a comparison between two recognizers can be misleading.Stanisław Kacprzak and Mieszko Fraś, arXiv (eess.AS), AGH University of Krakow, 2026
2 percentage pointsKacprzak and Fraś found that the lack of number normalization can change WER by more than 2 percentage points on some datasets.Stanisław Kacprzak and Mieszko Fraś, arXiv (eess.AS), AGH University of Krakow, 2026
Kaldi, the speech recognition toolkit written in C++, is licensed under the Apache License v2.0.Kaldi project, kaldi-asr.org, n.d.
Kim and colleagues write that WER takes into account only literal correctness, not semantic correctness, which matters for downstream language understanding.Suyoun Kim et al., Proceedings of Interspeech 2021, ISCA, 2021
47 percent less energy useLi and colleagues report that their component-level design changes cut the energy use of an on-device streaming speech recognition model by up to 47 percent while keeping accuracy comparable.Yang Li et al., arXiv (industry track of NAACL 2025, per the arXiv comments), 2024
1000 hoursLibriSpeech is a corpus of about 1,000 hours of 16 kHz read English speech, and its data is derived from read audiobooks from the LibriVox project.Vassil Panayotov et al., Open Speech and Language Resources (OpenSLR), 2015
1000 hoursLibriSpeech is a large-scale corpus of 1,000 hours of read English speech, distributed with clean and harder other test sets.Vassil Panayotov et al., Open Speech and Language Resources (OpenSLR), 2015
Lippmann's 1997 paper reviews past work comparing modern speech recognition systems with humans, to see how far the technology has moved toward human-like performance.Richard P. Lippmann, Speech Communication, vol. 22, pp. 1-15, 1997
Many state-of-the-art inverse text normalization systems use hand-written weighted finite-state transducer grammars, because the task has extremely low tolerance to unrecoverable errors.Yang Zhang et al., arXiv, 2021
1600+ languagesMeta's Omnilingual ASR is an open-source speech recognition model family for more than 1,600 languages.Meta FAIR, Hugging Face, 2025
Meta's Omnilingual ASR model and its code are licensed under Apache 2.0, while the training corpus is licensed separately under CC-BY-4.0.Meta FAIR, Hugging Face, 2025
Microsoft defines the real-time factor (RTF) of a device as the ratio of its processing time to the audio length.Microsoft Learn, n.d.
1 real-time factorMicrosoft says a device should have a real-time factor of 1 or lower to support real-time speech experiences.Microsoft Learn, n.d.
Microsoft says Azure real-time speech to text gives instant transcription with intermediate results for live audio.Microsoft, Microsoft Learn, n.d.
24 hours (maximum completion time at peak hours)Microsoft says that at peak hours a batch transcription job might take up to 30 minutes to start and up to 24 hours to complete.Microsoft, Microsoft Learn, n.d.
Microsoft says that submitting batch jobs at a higher rate does not speed up processing.Microsoft, Microsoft Learn, n.d.
5.8 percent word error rateMicrosoft's automated system reached word error rates of 5.8 percent on Switchboard and 11.0 percent on CallHome, slightly below the professional transcribers.W. Xiong et al., arXiv (Microsoft Research), 2017
Microsoft's batch transcription request must set either a contentContainerUrl or a contentUrls property to point at the audio.Microsoft Learn, 2026
Microsoft's Python quickstart installs the Azure Speech SDK with one pip command, pip install azure-cognitiveservices-speech.Microsoft Learn, 2026
MLX is an array framework for machine learning on Apple silicon, from Apple's machine learning research team.Apple machine learning research, GitHub (ml-explore/mlx), 2026
MLX, Apple's array framework, keeps its arrays in shared memory, so an operation can run on the CPU or the GPU without first copying the data between them.Apple machine learning research, GitHub (ml-explore/mlx), 2026
200 percentMorris et al. give an example in which a system that outputs two wrong words for each input word gets a WER of 200 percent, although both systems in their example communicate exactly zero information.Andrew Cameron Morris et al., Proceedings of Interspeech 2004, ISCA, 2004
Morris et al. state that the usual alignment, with zero weight for hits and equal weight for substitutions, deletions and insertions, is optimal.Andrew Cameron Morris et al., Proceedings of Interspeech 2004, ISCA, 2004
Morris, Maier and Green describe word error rate as the de facto standard index for speech recognition assessment, defined as the proportion of word errors to words processed.Andrew Cameron Morris et al., Proceedings of Interspeech 2004, ISCA, 2004
Morris, Maier and Green propose match error rate (MER), the proportion of input and output word matches that are errors, as an alternative absolute measure.Andrew Cameron Morris et al., Proceedings of Interspeech 2004, ISCA, 2004
44.5 thousand hours of EnglishMultilingual LibriSpeech has about 44.5 thousand hours of English audio, drawn from read audiobooks, and about 6 thousand hours across its other languages.Vineel Pratap et al., arXiv (Computation and Language), 2020
NVIDIA describes NeMo Speech as a toolkit for PyTorch developers working on automatic speech recognition, text to speech and speech language models.NVIDIA, GitHub, NVIDIA, n.d.
NVIDIA describes Parakeet TDT 0.6B v2 as a 600-million-parameter speech recognition model designed for English transcription. Its multilingual successor is v3.NVIDIA, Hugging Face, 2025
NVIDIA says Canary-Qwen-2.5B is ready for commercial use and is released under the CC-BY-4.0 license.NVIDIA, Hugging Face, n.d.
24 minutesNVIDIA says Parakeet TDT 0.6B v3 handles audio up to 24 minutes long with full attention, or up to 3 hours with local attention.NVIDIA, Hugging Face, 2025
25 languagesNVIDIA's Canary 1B v2 supports speech transcription for 25 languages, and speech translation from English into 24 languages.NVIDIA, Hugging Face, 2025
NVIDIA's model card for Parakeet TDT 0.6B v3 says its transcripts may be less than 100 percent accurate, and that accuracy depends on the language and the characteristics of the input audio.NVIDIA, Hugging Face, 2025
NVIDIA's model card for Parakeet TDT 0.6B v3 says that to run Python inference with the model, you install NVIDIA NeMo after installing a recent PyTorch version.NVIDIA, Hugging Face, 2025
NVIDIA's Nemotron Speech Streaming EN 0.6B model uses a cache-aware FastConformer encoder with 24 layers and an RNN-T decoder, which lets it stream audio as it arrives.NVIDIA, Hugging Face, n.d.
On a local computer, Google's quickstart says to create local authentication credentials for your user account with gcloud auth application-default login.Google Cloud, 2026
14.1 percent word error rateOn a subset of Google's voice search task, the LAS model reached a word error rate of 14.1 percent without a dictionary or a language model.William Chan et al., arXiv, 2015
3.5 percentage pointsOn CallHome, the IBM recognizer's word error rate was 3.5 percentage points higher than the best human transcriber's.Calculated from other facts
17.6 percent word error rateOn the CallHome test set of conversations between friends and family, Whisper Large V2 had a word error rate of 17.6 percent.Alec Radford et al., OpenAI (arXiv), 2022
4.1 times fasterOn the faster-whisper project's CPU benchmark with the small model, faster-whisper with int8 precision took about 4.1 times less time than openai/whisper in fp32 precision.Calculated from other facts
5.9 percent word error rateOn the NIST 2000 Switchboard test of telephone conversations, professional human transcribers had a word error rate of 5.9 percent, and 11.3 percent on the harder CallHome conversations.W. Xiong et al., arXiv (Microsoft Research), 2017
23 timesOn the Open ASR Leaderboard, Parakeet TDT 0.6B v2 processes audio about 23 times faster than Whisper Large v3 (RTFx 3,390 against 146).Calculated from other facts
5.42 percent average WEROn the Open ASR Leaderboard's English short-form tests, Cohere Labs Transcribe had an average word error rate of 5.42 percent, and the paper's table marks it as open.Hugging Face and collaborators, arXiv, 2025
7.21 percent average WEROn the Open ASR Leaderboard's English short-form tests, Distil-Whisper Large v3.5 had an average word error rate of 7.21 percent, with an RTFx of 202.Hugging Face and collaborators, arXiv, 2025
7.15 percent average WEROn the Open ASR Leaderboard's English short-form tests, NVIDIA Canary 1B v2 had an average word error rate of 7.15 percent, with an RTFx of 749.Hugging Face and collaborators, arXiv, 2025
6.32 percent average WEROn the Open ASR Leaderboard's English short-form tests, NVIDIA Parakeet TDT 0.6B v3 had an average word error rate of 6.32 percent, with an RTFx of 3,330.Hugging Face and collaborators, arXiv, 2025
7.83 percent average WEROn the Open ASR Leaderboard's English short-form tests, OpenAI Whisper Large v3 Turbo had an average word error rate of 7.83 percent, with an RTFx of 200.Hugging Face and collaborators, arXiv, 2025
6.05 percent average WEROn the Open ASR Leaderboard's English short-form track, NVIDIA Parakeet TDT 0.6B v2 had an average word error rate of 6.05 percent and an inverse real-time factor of 3,390.Hugging Face and collaborators, arXiv, 2025
0.003 US dollars per minute of audioOn the paid tier, Gemini 3.5 Transcribe lists at $2.00 per million input audio tokens, which Google estimates as $0.003 per minute of audio.Google, Google AI for Developers, 2026
5.47 percent average WEROn the same English short-form tests, Zoom Scribe v1 had an average word error rate of 5.47 percent, and the paper's table marks it as not open.Hugging Face and collaborators, arXiv, 2025
59 secondsOn the same GPU, faster-whisper with int8 precision transcribed the recording with the large-v2 model in 59 seconds.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
5.63 percent average WEROn the same leaderboard track, NVIDIA Canary Qwen 2.5B, a speech encoder with a language-model decoder, had an average word error rate of 5.63 percent and an RTFx of 418.Hugging Face and collaborators, arXiv, 2025
7.44 percent average WEROn the same leaderboard track, OpenAI Whisper Large v3 had an average word error rate of 7.44 percent and an inverse real-time factor of 146.Hugging Face and collaborators, arXiv, 2025
5.06 percentOn the same typewritten historical records, the OCR output had a word error rate of 5.06 percent.Zahra Abedi, arXiv, 2025
1.08 percentOn typewritten historical records, the OCR output had a character error rate of 1.08 percent against a labelled sample.Zahra Abedi, arXiv, 2025
September 21, 2022 dateOpenAI released Whisper as open-source software on September 21, 2022.Matthias Bastian, The Decoder, 2022
OpenAI released Whisper's code and model weights under the MIT License.OpenAI, GitHub (openai/whisper), 2026
OpenAI says data sent to its API is not used to train or improve its models unless the customer explicitly opts in.OpenAI, OpenAI Platform documentation, 2026
OpenAI says its gpt-realtime-whisper transcription model requires turn detection to be omitted or set to null, unlike models that default to server voice activity detection.OpenAI, OpenAI Developers, n.d.
OpenAI says that a higher voice activity detection threshold requires louder audio to activate the detector, and so may work better in noisy environments.OpenAI, OpenAI Developers, n.d.
openai-whisper saves its model files to ~/.cache/whisper unless you set a different folder with --model_dir.OpenAI, GitHub (openai/whisper), 2026
0.003 US dollars per minuteOpenAI's API price list shows its gpt-4o-mini-transcribe model at $0.003 per minute.OpenAI, OpenAI, 2026
OpenAI's Audio API provides a speech endpoint based on its GPT-4o mini TTS model, which turns written text into spoken audio.OpenAI, 2026
OpenAI's chief executive, Sam Altman, was fired by the board and later reinstated after board members questioned his leadership.The Decoder, The Decoder, 2025
0.017 US dollars per minuteOpenAI's price list gives gpt-realtime-whisper, for live transcription, at $0.017 per minute.OpenAI, OpenAI, 2026
2.50 US dollars per 1 million input tokens (gpt-4o-transcribe)OpenAI's price table lists token prices for gpt-4o-transcribe and gpt-4o-mini-transcribe as well as per-minute estimates.OpenAI, OpenAI, 2026
OpenAI's price table puts its per-minute speech rates in a column headed Estimated cost.OpenAI, OpenAI, 2026
OpenAI's README says its English-only .en models tend to perform better for English applications, especially the tiny and base sizes.OpenAI, GitHub (openai/whisper), 2026
OpenAI's Realtime API voice activity detection, its default server-side turn detection for models that support it, uses periods of silence to split the audio into turns.OpenAI, OpenAI Developers, n.d.
OpenAI's Realtime guide says that a shorter silence duration setting for voice activity detection means a turn is detected as ended more quickly.OpenAI, OpenAI Developers, n.d.
98 languagesOpenAI's speech to text guide says Whisper supports 98 languages, and that accuracy varies by language.OpenAI, OpenAI, 2026
OpenAI's speech-to-text guide says its gpt-4o-transcribe-diarize model is for use only when you need to identify who speaks during different parts of a recording.OpenAI, OpenAI, 2026
OpenAI's speech-to-text guide says that file streaming is supported by its gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize models, and that the whisper-1 model does not support it.OpenAI, OpenAI, 2026
OpenAI's transcription API expects your secret API key in an Authorization header, written as the word Bearer followed by the key.OpenAI, OpenAI, 2026
OpenAI's Whisper Colab notebook says its first commands install the Python packages needed to use Whisper and evaluate transcription results.OpenAI (GitHub), n.d.
OpenAI's Whisper package requires the command-line tool ffmpeg to be installed on your system, in addition to the Python package.OpenAI, GitHub (OpenAI repository), 2023
33 percentOpting into data logging lowers Google's V1 standard list price by about one third, from $0.024 to $0.016 per minute.Calculated from other facts
PocketSphinx, the lightweight recognizer in the CMU Sphinx family, descends from older Carnegie Mellon University systems that were released as free software under a BSD-like license.CMU Sphinx project, GitHub (cmusphinx/pocketsphinx), n.d.
PyTorch's torch.cuda.is_available() function returns a bool that says whether CUDA is currently available to the program.PyTorch contributors, PyTorch, 2026
Qualcomm's Hugging Face page says its repository holds pre-exported Whisper Tiny model files optimized for Qualcomm devices.Qualcomm AI Hub, on Hugging Face, n.d.
30 seconds of audio per clipQualcomm's page for its Whisper Tiny build says the model transcribes audio clips up to 30 seconds long.Qualcomm AI Hub, on Hugging Face, n.d.
Quantized whisper.cpp models require less memory and disk space, and can run more efficiently on some hardware.Georgi Gerganov, GitHub, 2026
Rabiner's 1989 tutorial says HMMs must solve three fundamental problems before they apply to speech: evaluating the likelihood of an observation sequence, finding the best sequence of model states, and adjusting model parameters to fit the observed signal.Lawrence R. Rabiner, Proceedings of the IEEE, vol. 77, no. 2, pp. 257-286, 1989
Running faster-whisper on a GPU requires NVIDIA libraries to be installed first.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
8 percent relative improvement in word error rateSainath and colleagues report that their streaming on-device model, RNN-T plus LAS, obtained an 8 percent relative improvement in word error rate over a conventional model at the same latency, while being more than 400 times smaller.Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020
Sainath et al. define endpointer latency as the time it takes for the microphone to close after the user stops speaking.Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020
70 milliseconds (reduction in 90th percentile endpointer latency)Sainath et al. report that a joint RNN-T endpointer model reached a word error rate of 6.8 percent and cut the 90th-percentile endpointer latency by 70 ms.Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020
860 milliseconds (90th percentile endpointer latency)Sainath et al. report that with an acoustic endpointer alone, the 90th-percentile endpointer latency was 860 ms.Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020
450 milliseconds (median endpointer latency)Sainath et al. report that with an acoustic endpointer alone, their on-device model reached a median endpointer latency of 450 ms, and its word error rate rose from 6.2 to 7.4 percent.Tara N. Sainath et al., ICASSP 2020 (arXiv:2003.12710), 2020
Shannon and colleagues report that increasing the number of bands to beyond three or four gave no further improvement in their results.Robert V. Shannon et al., Science, 1995
Speaker diarization is not supported in Gemini live transcription sessions, and Google points to its non-streaming audio transcription endpoint for speaker labels.Google, Google AI for Developers, 2026
SpecAugment is a data augmentation method that is applied directly to a speech network's input features, the filter bank coefficients. It warps the features and masks blocks of frequency channels and blocks of time steps.Daniel S. Park et al., Proceedings of Interspeech 2019, 2019
Speech Communication describes itself as an interdisciplinary journal whose primary objective is the rapid dissemination and thorough discussion of basic and applied research results.Elsevier, with EURASIP and ISCA, n.d.
Stanford's CS224S course, in its Spring 2025 offering, is an introduction to spoken language technology with an emphasis on dialog and conversational systems.Andrew Maas, Stanford University, 2025
The 2004 review by Morris, Maier and Green cites a 1982 workshop paper by Woodard and Nelson as the source for the definition of WER for isolated word recognition.Andrew Cameron Morris et al., Proceedings of Interspeech 2004, ISCA, 2004
The 2017 Montreal Forced Aligner paper says the aligner uses Kaldi instead of HTK, so that it can be distributed as a stand-alone package.Michael McAuliffe et al., Proceedings of Interspeech 2017, 2017
The 2019 paper says word diarization error rate must be used together with word error rate, because it does not account for deletions and insertions.Laurent El Shafey et al., Interspeech 2019, 2019
6 authorsThe 2022 paper that introduced Whisper lists six authors, among them Greg Brockman and Ilya Sutskever.Alec Radford et al., arXiv (eess.AS), 2022
The 2026 draft of Jurafsky and Martin's textbook says speech is a natural interface for communicating with appliances, digital assistants and chatbots, especially on cellphones, where keyboards are less convenient.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
The 2026 draft of Jurafsky and Martin's textbook says that automatic speech recognition has matured to the point where it is viable for many practical tasks, while automatic transcription by any speaker in any environment is still far from solved.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
The 2026 draft of Jurafsky and Martin's textbook says that understanding spoken language, or at least transcribing it into writing, is one of the earliest goals of computer language processing.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
The 2026 draft of Jurafsky and Martin's textbook says transcription is important in fields like law, where dictation plays an important role.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
50 percentThe American Speech-Language-Hearing Association defines the speech recognition threshold as the minimum hearing level for speech at which a person can recognize 50% of the speech material.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The American Speech-Language-Hearing Association recommends spondaic words as the usual test material for the speech recognition threshold, and defines them as two-syllable words with equal stress on both syllables.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The American Speech-Language-Hearing Association says that for most clinical cases the speech recognition threshold agrees well with the three-frequency pure tone average, but with a sloping loss it may agree better with the two-frequency average.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The American Speech-Language-Hearing Association says that when the speech recognition threshold and the pure tone average disagree, the results are inconsistent, which can be an early sign of pseudohypacusis, or exaggerated hearing loss.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The American Speech-Language-Hearing Association says the primary clinical purpose of a speech threshold is to serve as a validity check for the pure tone audiogram.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The American Speech-Language-Hearing Association says the speech threshold procedure may need modifying for young children and for mentally disabled, uncooperative or neurologically disabled people, and that modifications must be recorded.American Speech-Language-Hearing Association, American Speech-Language-Hearing Association, 1988
The authors of Jurafsky and Martin's textbook say readers may use the draft chapters and slides in their courses.Daniel Jurafsky and James H. Martin, Stanford University, 2026
The authors of the Open ASR Leaderboard paper say that private evaluation sets could help minimize the risk of test-set contamination.Hugging Face and collaborators, arXiv, 2025
The authors write that deep neural networks with many hidden layers, trained with new methods, have been shown to outperform Gaussian mixture models on a variety of speech recognition benchmarks.Geoffrey Hinton et al., IEEE Signal Processing Magazine, 2012
1 GB of VRAMThe base Whisper model needs about 1 GB of video memory.OpenAI, GitHub (openai/whisper), 2026
7 times the speed of largeThe base Whisper model runs at about 7 times the relative speed of the large model.OpenAI, GitHub (openai/whisper), 2026
7.4 times lower error rateThe best machine error rate on Switchboard fell from 43 percent in the 1997 review to 5.8 percent for the 2016 Microsoft system, about 7.4 times lower.Calculated from other facts
The Careless Whisper study evaluated Whisper as it stood in 2023, so its 1 percent hallucination figure describes that version of the model.Allison Koenecke et al., arXiv (ACM FAccT 2024), 2024
The Common Voice paper says the corpus is the largest audio corpus in the public domain for speech recognition, by both hours and number of languages.Rosana Ardila et al., arXiv (LREC 2020), 2020
2.1 percent word error rateThe Conformer model reached word error rates of 2.1 percent on LibriSpeech test-clean and 4.3 percent on test-other without a language model, and 1.9 and 3.9 percent with one.Anmol Gulati et al., arXiv (Interspeech 2020), 2020
The Conformer paper introduces the convolution-augmented transformer for speech recognition, named Conformer, which combines convolution neural networks with transformers.Anmol Gulati et al., arXiv (Interspeech 2020), 2020
2 secondsThe default voice activity filter in faster-whisper removes only silences longer than 2 seconds.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
6x times faster (project claim)The Distil-Whisper project says its distilled model is 6x faster and 50% smaller than Whisper, within 1% word error rate.Hugging Face, GitHub (huggingface/distil-whisper), n.d.
The docstring of OpenAI's open-source Whisper code says a segment is treated as silent when its no-speech probability is above the threshold and its average log probability is below the log probability threshold.OpenAI, GitHub (openai/whisper), n.d.
The docstring of SpeechRecognition's Google recognizer says it raises a RequestError when the recognition operation fails, when the key is not valid, or when there is no internet connection.Anthony Zhang (Uberi), GitHub, source code, n.d.
The docstring of SpeechRecognition's Google recognizer says the default key should generally be used only for personal or testing purposes, because Google may revoke it at any time.Anthony Zhang (Uberi), GitHub, source code, n.d.
The Earnings-22 paper says modern speech recognition systems have reached superhuman word error rates on many common corpora, while lacking adequate performance on speech in the wild.Miguel Del Rio et al., arXiv (Computation and Language), submitted to Interspeech 2022, 2022
5.0 GBThe ElevenLabs speech to text API reference says the file for a transcription request must be less than 5.0 GB.ElevenLabs, 2026
512 kilobytes of internal SRAMThe ESP32-S3 is a dual-core processor that runs at 240 MHz and has 512 KB of internal SRAM.Espressif Systems, 2026
The expo-speech-recognition README says that if your Expo project has no ios or android directories, you need to create a development build.jamsch, GitHub, open-source project README, 2026
The expo-speech-recognition README says the requiresOnDeviceRecognition option, which defaults to false, stops the device from sending audio over the network, and that it is enabled only where the device supports it.jamsch, GitHub, open-source project README, 2026
The expo-speech-recognition README says you can use its module to call the native speech APIs directly, and that its listener events are similar to the Web Speech API.jamsch, GitHub, open-source project README, 2026
13 minutesThe faster-whisper benchmark measured how long each implementation took to transcribe a 13-minute recording.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
4 times faster (project claim)The faster-whisper project says its implementation is up to 4 times faster than the original Whisper code for the same accuracy, while using less memory.SYSTRAN, GitHub (SYSTRAN/faster-whisper), n.d.
MIT licenseThe faster-whisper repository is under the MIT License, which grants permission to use and copy the software free of charge.SYSTRAN, GitHub, SYSTRAN, 2023
The first run of whisper.cpp on a device with Core ML is slow, because the Neural Engine service compiles the model to a device-specific format.Georgi Gerganov, GitHub, 2026
102 languagesThe FLEURS benchmark is a parallel speech dataset in 102 languages, with about 12 hours of speech supervision per language.Alexis Conneau et al., arXiv (Computation and Language), 2022
The GigaSpeech dataset card says its creators do not own the copyright of the audio, and that access is offered for non-commercial research and educational purposes.SpeechColab, Hugging Face, n.d.
The Godot Asset Library listing for Godot Whisper describes it as an addon that enables realtime audio transcription, supports OpenCL for most platforms and Metal for Apple devices, and runs on a separate thread.Godot Engine Asset Library, 2026
The Hello Edge authors say keyword spotting, which must always listen, typically runs on tiny microcontrollers with limited memory and compute capability, so the network design has to fit those limits.Yundong Zhang et al., arXiv, 2017
The HMM, with Gaussian mixture models as the phonetic component, spread through the speech community and became the dominant paradigm by the 1990s.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
The IEEE Transactions on Audio, Speech and Language Processing is dedicated to innovative theory and methods for processing signals that represent audio, speech and language, and their applications.IEEE Signal Processing Society, n.d.
The jiwer Python package computes its speech measures, including WER, from the minimum edit distance between the reference and the hypothesis sentences.Jitsi open source project, GitHub, jitsi/jiwer, n.d.
The jiwer README offers pip as the alternative to uv for installing the package.Jitsi open source project, GitHub, jitsi/jiwer, n.d.
The jiwer README says that applying an empty reference and hypothesis pair to the WER formula is undefined behaviour, because it divides by zero.Jitsi open source project, GitHub, jitsi/jiwer, n.d.
The jiwer README says that when the hypothesis is non-empty and the reference is empty, every word or character in the hypothesis counts as an insertion.Jitsi open source project, GitHub, jitsi/jiwer, n.d.
The Kaldi project says it is not a speech recognition toolkit for beginners, because it lets you do many operations that do not make sense.Kaldi project, kaldi-asr.org, n.d.
2011 year of first public releaseThe Kaldi project says its code was released on 14 May 2011 and presented publicly at ICASSP 2011 in Prague.Kaldi project, kaldi-asr.org, n.d.
The Kaldi project says its goal is to provide complete recipes for building speech recognition systems from widely available databases, such as those from the Linguistic Data Consortium.Kaldi project, kaldi-asr.org, n.d.
The Kaldi project says its toolkit is intended for use by speech recognition researchers.Kaldi project, kaldi-asr.org, n.d.
The Kaldi project's home page says the model builds it offers for download are few in number.Kaldi project, kaldi-asr.org, n.d.
3.9 GB of memoryThe large model in whisper.cpp needs about 3.9 GB of memory and takes 2.9 GiB on disk.Georgi Gerganov, GitHub, 2026
The LAS paper says its network produces character sequences without any independence assumptions between the characters, which it names as the key improvement over previous end-to-end CTC models.William Chan et al., arXiv, 2015
The license file of Hugging Face's transformers library is the text of the Apache License, Version 2.0.Hugging Face, GitHub, huggingface/transformers, n.d.
The license file of the Vosk API repository is the text of the Apache License, Version 2.0.Alpha Cephei, GitHub (alphacep/vosk-api), n.d.
The Linguistic Data Consortium distributes Switchboard-1 Release 2 under its User Agreement for Non-Members, and the licensing instructions depend on membership.John J. Godfrey and Edward Holliman, Linguistic Data Consortium, University of Pennsylvania, 1997
80 Mel-frequency filtersThe log-Mel spectrogram function in the Whisper code uses 80 Mel-frequency filters by default, and it supports only 80 or 128 filters.OpenAI, GitHub, openai/whisper repository (main branch), 2022
2 times the speed of largeThe medium Whisper model runs at about 2 times the relative speed of the large model.OpenAI, GitHub (openai/whisper), 2026
19,500 times the on-chip SRAMThe memory a Whisper large model needs, about 10 GB, is roughly 19,500 times the 512 KB of internal SRAM on an ESP32-S3.Calculated from other facts
The Moonshine project says its models are optimized for live streaming, with low latency because they do work while the user is still talking.Moonshine AI (formerly Useful Sensors), GitHub (moonshine-ai/moonshine), n.d.
The Moonshine speech recognition models are MIT licensed by default, in every language and at every size.Moonshine AI (formerly Useful Sensors), GitHub (moonshine-ai/moonshine), n.d.
2.0 license versionThe NeMo repository's LICENSE file is the Apache License, Version 2.0, the same license as the NeMo code.NVIDIA, GitHub, NVIDIA, n.d.
The NIST sclite scorer reports the percent of correct words as the number of correct words divided by the number of reference words, and reports inserted words as a separate percentage.NIST Speech Group, NIST Speech Recognition Scoring Toolkit (SCTK), documentation hosted at MIT, n.d.
The NVIDIA Canary 1B model card lists its license as CC-BY-NC-4.0, a Creative Commons license that does not allow commercial use.NVIDIA, Hugging Face, n.d.
The NVIDIA Canary 1B v2 model card lists its license as CC-BY-4.0.NVIDIA, Hugging Face, 2025
The NVIDIA Parakeet TDT 0.6B v2 model card lists its license as CC-BY-4.0.NVIDIA, Hugging Face, 2025
The Open ASR Leaderboard paper observes that Conformer-based encoders paired with transformer-based decoders achieve the best average word error rate.Hugging Face and collaborators, arXiv, 2025
The Open ASR Leaderboard paper says that the full and latest results table can be found on Hugging Face.Hugging Face and collaborators, arXiv, 2025
The Open ASR Leaderboard reports the inverse real-time factor (RTFx) rather than the real-time factor.Hugging Face and collaborators, arXiv, 2025
The Open ASR Leaderboard reports two metrics: word error rate for transcription quality and inverse real-time factor for inference speed.Hugging Face and collaborators, arXiv, 2025
130 billion US dollarsThe OpenAI Foundation holds an equity stake in OpenAI Group PBC that was valued at about $130 billion at the time of the October 2025 restructuring.Converge Digest, Converge Digest, 2025
The openai-whisper command line runs on an NVIDIA GPU (CUDA) only when PyTorch can detect one, and otherwise falls back to the CPU.OpenAI, GitHub (openai/whisper), 2026
The openai-whisper command line turns its starting temperature and its fallback increment into a list of temperatures that runs up to 1.0, which the decoder tries in turn when a decode fails.OpenAI, GitHub (openai/whisper), 2026
The OpenSLR page for LibriSpeech lists its license as CC BY 4.0, and describes it as a large-scale corpus of read English speech.Vassil Panayotov et al., Open Speech and Language Resources (OpenSLR), 2015
The paper says most speech recognition systems used hidden Markov models to handle the timing of speech, and Gaussian mixture models to score how well each HMM state fits a short window of acoustic input.Geoffrey Hinton et al., IEEE Signal Processing Magazine, 2012
The python-sounddevice package records audio from the sound device into a NumPy array with its rec() function.Matthias Geier, python-sounddevice documentation, n.d.
The PyTorch guide says to change the Colab runtime through the Runtime menu and select a T4 GPU as the hardware accelerator.PyTorch documentation, n.d.
The react-speech-recognition README says browser support for the Web Speech API it uses is currently limited, with Chrome giving the best experience.James Brill, GitHub, open-source project README, 2026
The react-speech-recognition README says that without a polyfill the Web Speech API is largely supported only by Google browsers, and gives a list of browsers dated May 2021.James Brill, GitHub, open-source project README, 2026
The SALMONN authors say their model lets a large language model process and understand general audio inputs directly, with competitive performance on speech tasks such as automatic speech recognition.Changli Tang et al., arXiv (ICLR 2024), 2023
0, 3, 3, 4 alignment cost weightsThe sclite scoring program aligns words with a cost of 0 for a correct word, 3 for an insertion, 3 for a deletion and 4 for a substitution.NIST Speech Group, NIST Speech Recognition Scoring Toolkit (SCTK), documentation hosted at MIT, n.d.
1 millisecondThe Silero VAD project says one audio chunk of 30 ms or more takes less than 1 millisecond to process on a single CPU thread.Silero Team, GitHub (snakers4/silero-vad), 2026
4 times the speed of largeThe small Whisper model runs at about 4 times the relative speed of the large model.OpenAI, GitHub (openai/whisper), 2026
The SpeechRecognition package lists Vosk and OpenAI's Whisper among the engines it can call, both marked as working offline, alongside the Google Speech Recognition endpoint.Anthony Zhang (Uberi), GitHub, project README, n.d.
The SpeechRecognition package needs the separate PyAudio package only if you want to use microphone input; the rest of the library works without it.Anthony Zhang (Uberi), GitHub, project README, n.d.
The SpeechRecognition package's Google recognizer uses a generic API key that works without setup, and its docstring says that key should generally be used only for personal or testing purposes.Anthony Zhang (Uberi), GitHub, source code, n.d.
The SpeechRecognition package's license file is a BSD-style license with three numbered conditions, which allows redistribution in source and binary form with or without modification.Anthony Zhang (Uberi), GitHub, Uberi/speech_recognition, n.d.
The third edition of Jurafsky and Martin's textbook is titled as an introduction to natural language processing, computational linguistics and speech recognition, with language models.Daniel Jurafsky and James H. Martin, Stanford University, 2026
The third-edition draft of Jurafsky and Martin's textbook says the input to a speech recognizer is a series of acoustic waves.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026
809 million parametersThe turbo Whisper model has 809 million parameters.OpenAI, GitHub (openai/whisper), 2026
The turbo Whisper model is not trained for translation tasks, so translating non-English speech into English needs a multilingual model such as medium or large.OpenAI, GitHub (openai/whisper), 2026
6 GB of VRAMThe turbo Whisper model needs about 6 GB of video memory, less than the 10 GB the large model needs.OpenAI, GitHub (openai/whisper), 2026
8 times the speed of largeThe turbo Whisper model runs at about 8 times the relative speed of the large model.OpenAI, GitHub (openai/whisper), 2026
The Vosk project says it allows quick reconfiguration of the vocabulary for best accuracy.Alpha Cephei, 2026
The Vosk project says it provides a streaming API for the best user experience.Alpha Cephei, 2026
20+ languages and dialectsThe Vosk project says it supports more than 20 languages and dialects, with English, Indian English, German and French among them.Alpha Cephei, 2026
The Vosk project says it works offline even on lightweight devices such as the Raspberry Pi, Android and iOS.Alpha Cephei, 2026
The Vosk project's accuracy guide says modern systems, Vosk included, can be very accurate at one time and poor at another, because audio quality, vocabulary and accent all matter.Alpha Cephei, 2026
18 languagesThe VoxPopuli dataset card says it contains transcribed speech for 18 languages, and 29 hours of transcribed non-native English for research on accented speech.Meta AI, Hugging Face, n.d.
The Whisper authors chose an encoder-decoder Transformer as the architecture for their speech recognition models because they judged it well validated to scale reliably.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper authors list complete hallucination among known failure modes: the model outputs a transcript entirely unrelated to the actual audio, along with repeat loops and missing words at segment edges.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper paper concludes that its English speech recognition is not perfect but very close to human-level accuracy.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper paper says a complete speech system has several additional components, which are often handled separately, making the system relatively complex.Alec Radford et al., OpenAI (arXiv), 2022
30 secondsThe Whisper paper says its models are trained on 30-second audio chunks and cannot take longer audio as a single input.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper paper says many different tasks can be performed on the same input audio, and lists transcription, translation, voice activity detection and alignment among them.Alec Radford et al., OpenAI (arXiv), 2022
The Whisper README gives a pip command with --force-reinstall to update the package to the latest version of its repository.OpenAI, GitHub (openai/whisper), 2026
The Whisper repository says its turbo model is an optimized version of large-v3 that transcribes faster, with minimal degradation in accuracy.OpenAI, GitHub (OpenAI repository), 2023
The whisper.cpp command-line example runs only with 16-bit WAV files, so audio in other formats must be converted before transcription.Georgi Gerganov, GitHub, 2026
The whisper.cpp project says its quantized Whisper models require less memory and disk space, and that depending on the hardware they can be processed more efficiently.Georgi Gerganov, GitHub, 2026
The whisper.cpp README describes the project as high-performance inference of OpenAI's Whisper automatic speech recognition model.Georgi Gerganov, GitHub, 2026
0.5 seconds (sampling interval)The whisper.cpp stream example samples the microphone audio every half a second and runs the transcription continuously.ggml-org, GitHub (ggml-org/whisper.cpp), n.d.
The WhisperX README lists the faster-whisper project as the backend it uses for batched inference.m-bain/whisperX project, GitHub (m-bain/whisperX), 2026
The XXL README says its standalone executables of Whisper and Faster-Whisper are for people who do not want to bother with Python.GitHub (Purfview), n.d.
Turning off condition_on_previous_text can make the text inconsistent across windows, but makes the model less likely to get stuck in a failure loop.OpenAI, GitHub (openai/whisper), 2026
Unity's scripting reference says that if dictation is not enabled in the Speech privacy settings, the DictationRecognizer fails when you call Start.Unity Technologies, Unity Scripting API documentation, 2026
10 Windows versionUnity's scripting reference says the DictationRecognizer is currently functional only on Windows 10, and that dictation must be permitted in the user's Speech privacy policy.Unity Technologies, Unity Scripting API documentation, 2026
Unity's scripting reference says the DictationRecognizer must be disposed with Dispose() once you are done with it, to release the resources it uses.Unity Technologies, Unity Scripting API documentation, 2026
Unlike openai-whisper, faster-whisper does not need FFmpeg installed on the system, because it decodes audio with PyAV, which bundles the FFmpeg libraries.SYSTRAN, GitHub (SYSTRAN/faster-whisper), 2026
105,829 audio filesVersion 0.02 of Google's Speech Commands dataset contains 105,829 audio files, each a one-second recording of a single spoken English word or background noise.Hugging Face, dataset card for Google's Speech Commands dataset, 2018
300 MB of memory at runtimeVosk says a small model is around 50 MB on disk and needs about 300 MB of memory while it runs.Alpha Cephei, 2026
16 GB of memoryVosk says its big models can require up to 16 GB of memory, because they use more advanced algorithms.Alpha Cephei, 2026
Vosk says its big models have a fixed vocabulary that cannot be changed while the model runs.Alpha Cephei, 2026
5.69 percent word error rateVosk's 1.8 GB English model vosk-model-en-us-0.22 had a word error rate of 5.69 percent on LibriSpeech test-clean, and it is listed under the Apache 2.0 license.Alpha Cephei, 2026
9.85 percent word error rateVosk's small English model vosk-model-small-en-us-0.15, at 40 MB, had a word error rate of 9.85 percent on LibriSpeech test-clean, and it is listed under the Apache 2.0 license.Alpha Cephei, 2026
When openai-whisper loads its model on the CPU while CUDA is available, it prints a warning that says inference is running on the CPU.OpenAI, GitHub (openai/whisper), 2026
When openai-whisper runs on a CPU, it warns that FP16 is not supported there and uses FP32 instead.OpenAI, GitHub (openai/whisper), 2026
Whisper can retry a segment with a sequence of higher sampling temperatures when a decode fails its quality checks.OpenAI, GitHub (openai/whisper), 2026
25 milliseconds per windowWhisper re-samples all audio to 16,000 Hz and computes its 80-channel log-magnitude Mel spectrogram on 25-millisecond windows with a stride of 10 milliseconds.Alec Radford et al., OpenAI (arXiv), 2022
Whisper's initial prompt can give the model custom vocabulary or proper nouns, which makes those words more likely to be transcribed correctly.OpenAI, GitHub (openai/whisper), 2026
Whisper's own timestamps are per utterance rather than per word, and can be inaccurate by several seconds.m-bain/whisperX project, GitHub (m-bain/whisperX), 2026
Whisper's transcribe code treats a segment as silence, and skips it, when its no-speech probability is above one threshold and its average log probability is below another.OpenAI, GitHub (openai/whisper), 2026
70 times real timeWhisperX reports 70 times real-time transcription with the large-v2 model, using batched inference.m-bain/whisperX project, GitHub (m-bain/whisperX), 2026
WhisperX, an open-source project that runs Whisper-based recognition, adds word-level timestamps and speaker diarization to its output.m-bain/whisperX project, GitHub (m-bain/whisperX), 2026
100 frames per secondWith a 10-millisecond stride, Whisper's spectrogram has 100 frames for every second of audio.Calculated from other facts
1.8 percent word error rateWith all of LibriSpeech's labeled training data, wav2vec 2.0 reached word error rates of 1.8 percent on the clean test set and 3.3 percent on the harder other test set.Alexei Baevski et al., arXiv (NeurIPS 2020), 2020
4.8 percent word error rateWith just ten minutes of labeled data and pre-training on 53,000 hours of unlabeled audio, wav2vec 2.0 still reached 4.8 percent word error rate on LibriSpeech clean and 8.2 percent on other.Alexei Baevski et al., arXiv (NeurIPS 2020), 2020
100 percentWord error rate can be greater than 100 percent, because its equation counts insertions as errors.Daniel Jurafsky and James H. Martin, Stanford University, online manuscript of the third edition, 2026