Speech recognition
Speech Recognition: What It Is and How Well It Works
Speech recognition turns the sound of speech into text by converting audio into acoustic features and searching for the most probable words, and since about 2012 single neural networks trained on hundreds of thousands of hours of audio have replaced hand-built chains of models. On clean, read English the best systems now miss only a word or two in a hundred, but accuracy still depends on the audio, the speaker, the vocabulary and whether the model streams or runs on your own device, so word error rate measured on audio like yours is the number that matters.
Updated October 16, 202619 min read60 cited facts, each traced to its source
680,000hours of training audio
OpenAI's Whisper speech recognition models were trained on 680,000 hours of multilingual and multitask audio with transcripts.
Alec Radford et al., arXiv (eess.AS), 2022
2.1percent word error rate
The 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
5.9percent word error rate
On 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
9.4times
Whisper Large V2 has about 9.4 times the word error rate on CHiME6 that it has on LibriSpeech Clean.
How it is calculated
Guides in this topic
Each guide takes one part of the subject in depth.
- How Speech Recognition Works, From Sound to TextSpeech recognition finds the sequence of words most likely to have produced a sound. It turns the microphone signal into numbers, converts those numbers into a spectrogram with a new frame every 10 milliseconds, and then either a chain of acoustic, pronunciation and language models or one neural network maps the frames to words.
- What Is Whisper AI? OpenAI's Speech Model ExplainedWhisper is a free, open-source speech recognition model that OpenAI released in September 2022. It was trained on 680,000 hours of audio with transcripts, and anyone can download and run it, or call OpenAI's paid hosted API instead.
- Speech-to-Text APIs Compared: Price, Accuracy, SpeedA speech-to-text API returns a transcript for audio you send it, and bills by the minute or hour. List prices differ several-fold between providers, while independent tests show the best systems are close on clean English, so price, latency, speaker labels and data terms often decide the choice. Test two or three providers on your own audio before you commit.
- Word Error Rate (WER): How to Calculate and Read ItWord error rate (WER) is the number of substituted, deleted and inserted words divided by the number of words in a correct transcript, so a 5 percent WER means about one word in twenty is wrong. The score depends on the reference transcript, the text rules applied before counting and the audio, so compare two scores only when all three match.
- How Accurate Is Speech Recognition? The Real NumbersOn clean, read English the best recognizers miss only about one word in every hundred, but on telephone conversation the same kind of system misses about 14 in a hundred, and noise, distance, accents and rare words push the rate higher. Accuracy is a range set by the audio, so the number to trust is the error rate measured on audio like yours.
- Speech Recognition Models: How Neural ASR WorksModern speech recognition uses neural networks of three main kinds: connectionist temporal classification (CTC), transducers and attention encoder-decoders such as Whisper. Transducers are the usual choice for dictation that appears as you speak, because they stream. Whisper reads a whole clip before it writes, so it suits transcribing files better than live typing.
- Open-Source Speech Recognition: The Best ModelsOpen-source speech recognition models now match some proprietary services on English benchmarks, and many run for free on your own computer. Licenses, language coverage, streaming support and hardware needs differ widely, so choose by those needs and by word error rate on your own audio.
- How to Run Whisper Locally on Your Own ComputerYes. Whisper runs on your own computer with no API key and no usage fee, after you install an implementation and download a model. Its turbo model needs about 6 GB of video memory, and the right setup depends on whether you have an NVIDIA GPU, a Mac, or only a processor.
- Speech Recognition Benchmarks and LeaderboardsA speech recognition benchmark is a fixed set of recordings with human reference transcripts, scored by word error rate. The Open ASR Leaderboard compares 86 systems on the same audio, but the most quoted test sets are read speech, and some of their text now appears in language model training data.
- Real-Time Speech Recognition: Streaming and LatencyReal-time (streaming) speech recognition shows words while you are still talking. It uses a model that sees only a little of the audio after each word, so it gives up some accuracy for speed. Batch recognition waits for the whole recording and is usually more accurate. The delay that matters most depends on the use: the first partial words, the final text, or the wait after you stop talking.
- On-Device vs Cloud Speech Recognition: The TradeoffsOn-device recognition runs the speech model on your own phone or computer, so audio stays local and works offline. Cloud recognition can use larger models and more languages, and it bills by the minute. The right choice depends on privacy, connectivity, cost per minute and the accuracy of each engine on your own speech.
- Speech Recognition in Python: Libraries ComparedSpeechRecognition is the quickest start, but its default Google engine runs on a generic key that Google can revoke. For accurate recognition on your own computer, use faster-whisper or Hugging Face transformers, Vosk for small devices, or NVIDIA NeMo for Parakeet.
- History of Speech Recognition: From Audrey to WhisperThis history begins in 1952, when Bell Laboratories built a machine that recognized spoken digits from one speaker. Accuracy then rose in steps: statistical hidden Markov models in the 1980s, deep neural networks from 2012, end-to-end networks in the late 2010s, and Whisper, an openly licensed model trained on 680,000 hours of audio and released in 2022.
Speech recognition (also called automatic speech recognition, ASR, or speech to text) is software that turns spoken audio into written words. Modern systems do it with one neural network that listens to a recording and writes the transcript directly. On clean, read English the best research systems have reached word error rates near 2 percent 1, but the same Whisper model makes about nine times as many errors on dinner-party recordings with distant microphones as on clean read speech 2. That gap, not the headline score, is what decides whether a recognizer works for you.
This page is the overview of the whole subject. Each section gives the short answer and the evidence, then points to a guide that goes deeper. If you want to use recognition rather than understand it, start with our overview of dictation, which covers talking to your computer in practice.
Speech recognition finds the most probable words for a sound
A recognizer solves a search problem. It takes a sound wave, turns it into numbers, and looks for the word sequence that most probably produced those numbers. The classic recipe splits the work into three models: an acoustic model that maps short slices of sound to speech sounds, a pronunciation dictionary that maps sounds to words, and a language model that scores which word sequences are plausible ("recognize speech" beats "wreck a nice beach"). A decoder then searches for the best path through all three.
Before any model runs, the audio is prepared. It is sampled, cut into short overlapping frames, and converted into a mel spectrogram, a picture of how much energy the sound has at each pitch over time. The model never sees the raw waveform in most systems, only this picture. Our guide to how speech recognition works walks through each stage with an example.
The last step matters more than people expect. A raw recognizer outputs lowercase words with no punctuation, and a separate step restores capital letters, commas and numbers (for example turning "twenty twenty six" into "2026"). This is called inverse text normalization, and it is one reason the same engine can look better or worse in different products.
Common questions about the basics have short answers of their own: is speech recognition AI, is it NLP, how it differs from voice recognition (which identifies who is speaking, not what they say), and what it is used for.
The field moved from three hand-built models to one network
For forty years, speech recognition was an engineering project in assembling parts. In 1952, Bell Laboratories built a system that recognized isolated spoken digits from a single speaker 3. A machine called AUDREY recognized the digits zero to nine with 90 percent accuracy, but only when its inventor spoke 4. For decades, statistical methods built on hidden Markov models were the standard 5, and in April 1997 Dragon Systems unveiled NaturallySpeaking, the first general-purpose, large-vocabulary continuous speech recognition product 6.
The next break came from deep learning. A 2012 paper from four research groups reported that deep neural networks outperformed the Gaussian mixture models on a variety of speech recognition benchmarks 7, and they took over the acoustic model inside the old pipeline. The bigger change followed: networks that learn the whole job end to end. A key obstacle was alignment. Alex Graves wrote in 2012 that recurrent networks traditionally required a pre-defined alignment between input and output sequences, and proposed a transduction system that learns the alignment itself 8. That idea, along with connectionist temporal classification (CTC) and attention-based encoder-decoders, is the basis of everything in the next section.
A milestone showed how far this had gone. In 2017, Microsoft reported that its system reached a word error rate of 5.8 percent on the Switchboard telephone conversations 9, slightly below the 5.9 percent of professional human transcribers on the same test 10. The claim of "human parity" was measured on one benchmark of two-person phone calls, and later work showed that it did not carry over to every kind of audio. The full story, from Audrey to Whisper, is in our history of speech recognition, and the questions who invented speech recognition and which papers are the important ones have short answers.
Three kinds of neural network now do the work
Today's recognizers are neural networks with an encoder that turns the spectrogram into a sequence of internal representations and a decoder that turns those into text. The encoder is often a Conformer, a design from Google that combines convolution, which captures local sound detail, with transformer self-attention, which captures long-range context. On the LibriSpeech benchmark, the Conformer reached word error rates of 2.1 percent on the clean test set and 4.3 percent on the harder one without a language model 1.
The three families differ in how the decoder works:
| Family | How it writes text | Strength | Weakness | Typical use |
|---|---|---|---|---|
| CTC ( connectionist temporal classification) | Predicts a letter or word piece for every audio frame, then collapses repeats and blanks | Very fast, simple, easy to run in batches | Each output is chosen without looking at the previous ones, so it leans on an external language model | Fast offline transcription, keyword spotting |
| Transducer ( RNN-T and TDT) | Adds a small prediction network that remembers what was already written, and decides frame by frame whether to emit a token | Streams naturally, accurate | More complex to train | Live captions, phone dictation |
| Attention encoder-decoder (Whisper style) | Reads the whole encoded audio and writes text one token at a time, like a language model | Most accurate on average, handles many languages and tasks | Slower, needs a chunk of audio before it starts, can invent text | Batch transcription, translation |
The Open ASR Leaderboard, which compared 86 open-source and proprietary systems across 12 datasets 11, states the tradeoff directly. Conformer-based encoders paired with transformer-based decoders achieve the best average word error rate 12, while token-and-duration transducer (TDT) decoders offer superior speed, which makes them better suited for long-form and batched processing 13.
Training is mostly data. Supervised models learn from audio paired with human transcripts. Self-supervised models such as wav2vec 2.0 first learn the structure of speech from audio alone. With all of LibriSpeech's labeled data, wav2vec 2.0 reached 1.8 percent word error rate on the clean test set and 3.3 percent on the other set 14, and with only ten minutes of labeled data, after pre-training on 53,000 hours of unlabeled audio, it still reached 4.8 and 8.2 percent 15. That result points toward recognizers for languages with few transcribed recordings.
Our guide to speech recognition models covers architectures and training in detail. Related questions: how models are trained, the best neural network for speech recognition, and whether large language models can do speech recognition. Some top-ranked systems, such as NVIDIA Canary Qwen 2.5B, pair a speech encoder with a language-model decoder 16, a design the models guide explains.
Whisper made accurate recognition free and open
Whisper is the model most people have heard of, and the reason is scale and licensing. OpenAI trained it on 680,000 hours of multilingual and multitask audio with transcripts 17, far more than the 1,000 hours of read English in LibriSpeech 18. The authors report that compared with humans, the models approach human accuracy and robustness 19. The code and model weights use the MIT License 20, so anyone can run, modify and ship them.
Whisper comes in sizes. The smallest, tiny, has 39 million parameters and needs about 1 GB of memory 21. The largest has 1,550 million parameters and needs about 10 GB 22. That range is why Whisper runs on a phone at one end and a data-center GPU at the other, and why "Whisper" alone does not tell you how accurate a given app is.
Whisper has well-known limits. It processes audio in fixed-length windows, so it is not built for live streaming without extra engineering, and it can hallucinate (see the accuracy section). Our guides explain what Whisper is and how to run Whisper on your own computer. Quick answers: is Whisper free, who owns it, can it identify speakers (not by itself; that is speaker diarization), which model size is best, and Whisper versus Google speech to text.
Word error rate counts wrong words, and the denominator is easy to misread
The standard accuracy number is word error rate (WER): the number of substituted, deleted and inserted words divided by the number of words actually spoken. Researchers describe it as a standard measure of the discrepancy between machine and human transcriptions 23, and the formula puts the total number of words in the correct transcript in the denominator 24. A WER of 5 percent means about 5 wrong words in every 100 spoken. Because insertions count, WER can exceed 100 percent for a badly broken transcript 25.
Three habits make the number honest:
- Check what was normalized. The Open ASR Leaderboard, for example, removes punctuation and casing, and drops fillers such as "um," before scoring 26. A transcript that is perfect in words but has no punctuation has no word errors on such a test, yet is harder to read.
- Check the audio. Whisper Large V2 scored 2.7 percent on LibriSpeech Clean 27, 13.8 percent on the Switchboard phone conversations 28, and 25.5 percent on CHiME6, recordings of dinner-party conversations 29. That is about 9.4 times the error rate on CHiME6 as on clean audio 2.
- Check the sample size. A single recording tells you little. In a 2025 test of one 7-minute-31-second episode, Apple's speech API had a WER of 10.3 percent against 1.5 percent for Whisper Large V3 Turbo 30, a useful illustration but not a ranking of the products.
Our guide to word error rate explains the calculation step by step, including a Python version. Short answers: what a good word error rate is, how to calculate it, how to calculate it in Python, whether it can be over 100 percent, how it differs from character error rate, and WER versus accuracy.
Accuracy is excellent on clean speech and falls with noise, distance and difference
The honest summary of accuracy is a range, not a number. Best-case figures on benchmarks of read speech are very low: 1.8 to 4.3 percent on the LibriSpeech test sets 14 1. Real use is harder. Our guide to how accurate speech recognition is collects the numbers; the main factors are these.
Distance and noise
A microphone far from the speaker hurts more than most people expect. In a 2017 study, decoding close-talking sentences recorded in a studio gave a WER of 3.7 percent 31, while a single distant ceiling microphone in an apartment gave 12.0 percent 32. That is about 3.2 times the close-talk rate 33. The authors note that distant speech recognition still lacks robustness when noise and reverberation are present 34. Narrower telephone-style bandwidth costs accuracy too, though less on clean speech: in a preliminary Wall Street Journal test, the same recognizer gave 8.12 percent WER on wideband 16 kHz speech and 8.67 percent on narrowband 8 kHz speech 35 36. Practical advice on microphones is in our guides to dictation microphones and dictation in background noise.
Accents, age and speech differences
Accuracy is not equal across speakers. A study of five commercial systems found an average word error rate of 0.35 for Black speakers and 0.19 for white speakers 37, and traced the gap to the acoustic models 38. Children are a hard case too: fine-tuning Whisper-Small on children's speech cut its error rate on the MyST test set to 9.11 percent 39. Whispered speech is harder again: a 2024 baseline test gave Whisper 18.8 percent WER on whispered speech 40. And for people who stutter, technical changes to a consumer recognizer improved WER from 25.4 percent to 9.9 percent 41, which shows that much of the gap is fixable once someone tests for it. See the questions is speech recognition biased and is speech recognition more accurate than humans, and our overview of accessibility for how speech tools serve people with different speech.
Hallucination and names
Newer end-to-end models fail differently from older ones. They do not merely garble a word; sometimes they write fluent text that was never said. A 2024 study found that roughly 1 percent of Whisper's audio transcriptions contained entire hallucinated phrases or sentences 42, and 38 percent of those hallucinations included explicit harms such as violence or false authority 43. Silence handling is one defense (see the voice activity section). Rare names, drug names and brand names are the other recurring failure, because the model has seen them little. Read why Whisper hallucinates and why recognizers get names wrong.
Finally, independent measurements show the services are not interchangeable. A study of eleven common ASR services on university lecture recordings 44 concluded that, despite recent improvements, common ASR services lack reliability in accuracy 45.
Benchmarks are useful only if you know what they measure
A benchmark is a fixed set of recordings with human reference transcripts, scored by WER. The comparisons in this field rest on a handful of them, and each has a bias.
| Benchmark | What it contains | Best for | Watch out for |
|---|---|---|---|
| LibriSpeech | 1,000 hours of read English from audiobooks, with clean and harder "other" test sets 18 | Comparing research models | Clean read speech, now nearly saturated |
| Common Voice | Crowdsourced read sentences; 2,500 hours from over 50,000 people by 2019 46 | Many languages and accents | Microphone quality varies, read not spontaneous |
| Switchboard | Two-person telephone conversations | Conversational speech | Old, telephone bandwidth, American English |
| CHiME6 | Dinner-party conversations recorded with distant microphones | Hard, noisy far-field audio | Very difficult, rarely reported by vendors |
| Open ASR Leaderboard | 86 systems on 12 datasets, scoring WER and speed 11 | Comparing open and commercial models side by side | English short-form track dominates headlines |
Two cautions. First, models are often trained on data close to the test sets, so a public benchmark can flatter them. Second, vendors choose which benchmark to quote. A vendor claim of "99 percent accuracy" is a word error rate of 1 percent on whatever test the vendor chose; it says nothing about your meeting room. Our guides to speech recognition benchmarks cover each dataset, and the questions what the Open ASR Leaderboard is, which speech to text model is the most accurate and why accuracy claims differ go further.
Open models are competitive on accuracy, and speed varies widely
The leaderboard reports two numbers per model: average WER and inverse real-time factor (RTFx), the seconds of audio transcribed per second of compute. The real-time factor is its inverse. On the English short-form track, NVIDIA Canary Qwen 2.5B averaged 5.63 percent WER at an RTFx of 418 16. NVIDIA Parakeet TDT 0.6B v2 averaged 6.05 percent at an RTFx of 3,390 47. OpenAI Whisper Large v3 averaged 7.44 percent at an RTFx of 146 48.
Read the two charts together. Parakeet is about 23 times faster than Whisper Large v3 49 and also more accurate on English, but it was built for English at first and later extended: Parakeet TDT 0.6B v3 has 600 million parameters 50 and covers 25 European languages 51, while Whisper covers many more. NVIDIA says its models are designed or optimized to run on NVIDIA GPU-accelerated systems 52, so measure speed on your own machine before assuming the figures above apply to a laptop. For a small project, choose by language coverage first, accuracy second and speed third.
Our guide to open-source speech recognition compares the leading models, licenses and hardware needs. Quick answers: Whisper alternatives, what Vosk is, and what NVIDIA Parakeet is. If you want to write code, the guide to speech recognition in Python compares libraries, with companion questions on JavaScript, React Native and Unity and Unreal, plus the common network error in browsers.
Real-time recognition trades some accuracy for speed
Batch recognition waits for the whole recording, so it can use the entire context. Streaming recognition returns words while you are still talking, so the model may see only a little future audio. The cost is accuracy: a study of eleven ASR services found significantly lower quality for streaming speech recognition, the kind used for live events 53. A streaming result is also provisional; words near the end of the latest audio can change as more context arrives.
| Batch | Streaming | |
|---|---|---|
| Input | A complete file | A live audio stream |
| Output | Final transcript after processing | Partial words that update, then final text |
| Context | Whole recording | Past plus a short look-ahead |
| Typical models | Whisper, CTC and TDT models | Transducers, streaming Conformers |
| Best for | Interviews, podcasts, lectures | Dictation, live captions, voice assistants |
| Main metric | WER and RTFx | WER and delay from speech to text |
A real-time system also needs to know when you start and stop talking. That job belongs to a voice activity detector ( VAD). It labels each slice of audio as speech or not, so the recognizer skips silence and the app knows when to finalize a sentence. A good VAD is tiny: the open-source Silero VAD reports that one audio chunk of 30 ms or more takes less than 1 millisecond on a single CPU thread 54. VAD can also reduce hallucination, because a model that is never shown long silences has less opportunity to invent text over them.
Our guide to real-time speech recognition covers latency in detail. Related questions: can Whisper transcribe in real time, which speech to text API is fastest, and how recognition knows when you stop talking.
On-device recognition removes the network and keeps audio local
For years, accurate recognition meant sending audio to a server. Google's traditional server-side recognizer used a search graph of almost 2 GB, too large for a phone 55. In 2019, Google announced an end-to-end, all-neural, on-device recognizer for Gboard 56, with an 80 MB model that works offline 57, first for American English on Pixel phones, and said it offered the same accuracy as its server-based models in a much smaller model 58. Apple says that, on a Mac, voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers 59.
| On-device | Cloud | |
|---|---|---|
| Works offline | Yes | No |
| Latency | No network round trip | Depends on connection |
| Privacy | Audio stays on your device | Audio is sent to a provider |
| Model size | Limited by phone or laptop memory | Unlimited |
| Languages and vocabulary | Fewer, smaller | Broadest, updated centrally |
| Cost | Paid by your hardware and battery | Billed by the minute or hour |
The accuracy gap between the two has narrowed enough that the choice is now mostly about privacy, connection and cost. Yap, which we make, runs its dictation on the device on Mac and Windows; see the download page if you want to try that approach. For built-in options, see our guides to offline dictation on Mac and to voice privacy, which covers what providers do with audio.
Our guide to on-device versus cloud speech recognition lays out the tradeoffs. Related questions: Google's speech recognition and synthesis service, speech recognition on a Raspberry Pi, and whether on-device recognition is as accurate as cloud.
Speech-to-text APIs differ more in price and features than in accuracy
If you build a product, you will probably rent recognition by the minute. A speech-to-text API is a web service that returns a transcript for audio you send. The main providers (Google, Microsoft, Amazon, OpenAI, AssemblyAI, Deepgram and others) differ in price, supported languages, whether they stream, and extras such as speaker labels, punctuation and custom vocabulary. Google, for example, documents model adaptation, which increases the recognition model's bias toward some phrases more than others 60.
Choose with a test, not a table. Take a sample of your own audio, with a human-corrected transcript, run it through each candidate, and compare WER on the terms that matter, such as names and product words. Our guide to speech-to-text APIs compared gives the framework and the main providers. Quick answers: a free speech to text API, whether Google speech to text is free, how to use the Google API, OpenAI speech to text API pricing, and whether Gemini has a speech to text API. For file transcription without code, try our free audio to text tool in the browser, and for the wider topic see our overview of transcription.
A short guide to choosing
The right choice depends on three questions: what the audio is like, whether it must be live, and where the audio may go.
- Clean audio, a headset, English, a draft you will edit. Any current engine works. Use the free one built into your device.
- Recordings: interviews, lectures, calls. Use a batch model such as Whisper or Parakeet. Expect to proofread names and numbers.
- Live dictation or captions. Use a streaming transducer, on-device if audio must stay private.
- Accents, children, noise or medical vocabulary. Test on your own audio before choosing, and plan to add custom vocabulary or fine-tune.
- Quoted accuracy without a test set. Treat it as marketing.
For a detailed comparison of apps built on these engines, see our overview of dictation software. For the larger picture of voice interfaces and assistants, see voice and AI, and for non-English recognition see our overview of languages.
Where to go next
Everything above has a deeper page. Pick by what you need:
| If you want to know | Read |
|---|---|
| How audio becomes text, step by step | How speech recognition works |
| What the neural networks are | Speech recognition models |
| What Whisper is | What is Whisper AI |
| How to run Whisper yourself | Run Whisper locally |
| How to read an accuracy number | Word error rate |
| How accurate recognition really is | Speech recognition accuracy |
| Which test sets exist | Benchmarks and leaderboards |
| Free models you can run | Open-source speech recognition |
| Live and low-latency use | Real-time speech recognition |
| Phone and laptop versus server | On-device versus cloud |
| Paid web services | Speech-to-text APIs |
| Writing code | Speech recognition in Python |
| Where it all came from | History of speech recognition |
The one habit worth keeping from this page is to ask what audio a number was measured on. A recognizer that makes 2.7 errors per 100 words on audiobooks 27, 13.8 on phone calls 28 and 25.5 on dinner-party recordings 29 is the same recognizer, and only a test on your own audio tells you which of those you will get.
Evidence
Every number on this page, with where it comes from. Open a fact to see the exact words in its source.
- 1.
- 2.Whisper Large V2 has about 9.4 times the word error rate on CHiME6 that it has on LibriSpeech Clean.
Calculated from 2 other facts
- 3.
- 4.
- 5.
- 6.
- 7.
- 8.
- 9.
- 10.
- 11.
- 12.
- 13.
- 14.
- 15.
- 16.
- 17.
- 18.
- 19.
- 20.
- 21.
- 22.
- 23.
- 24.
- 25.
- 26.
- 27.
- 28.
- 29.
- 30.
- 31.
- 32.
- 33.
- 34.
- 35.
- 36.
- 37.
- 38.
- 39.
- 40.
- 41.In Apple research with 91 people who stutter, technical changes to a consumer speech recognizer cut utterances off 79.1 percent less often and improved word error rate from 25.4 percent to 9.9 percent.
Colin Lea et al., Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems (arXiv preprint), 2023 · Original
- 42.
- 43.
- 44.
- 45.
- 46.
- 47.
- 48.
- 49.
- 50.
- 51.
- 52.
- 53.
- 54.
- 55.
- 56.
- 57.
- 58.
- 59.
- 60.
Questions people ask
- Can an ESP32 or microcontroller do speech recognition?Yes, for a short list of voice commands and for wake words. Espressif's MultiNet command model runs on the ESP32-S3 and supports up to 300 commands, but the original ESP32 is not among its supported targets. Open dictation needs far more memory than the ESP32-S3's 512 KB of on-chip SRAM, so a microcontroller is the wrong tool for transcribing speech.
- Can large language models do speech recognition?Yes, in part: some systems pair a speech encoder with a large language model (LLM) and transcribe audio directly, and an LLM can also correct the output of an ordinary recognizer. A text-only LLM cannot hear audio, and Whisper is a speech model rather than a GPT chatbot. For live dictation, start with a streaming speech model, and check whether an LLM-based system streams before you rely on it.
- Can Whisper AI do text to speech?No. Whisper is a speech recognition model, so it turns recorded speech into text and cannot read text aloud or generate a voice. For text to speech, you need a separate model, such as OpenAI's GPT-4o mini TTS endpoint or a dedicated speech engine.
- Can Whisper identify different speakers?No. Whisper writes the words it hears but does not tag which person said each one. To get speaker labels you add a separate diarization step, either an open-source tool such as WhisperX or a hosted speech API that includes it, and then check the labels on recordings like yours.
- Can Whisper run without a GPU?Yes. Whisper runs on a processor with no GPU. In the faster-whisper README's benchmark, faster-whisper in int8 mode was the quickest CPU route, and whisper.cpp finished a 13-minute recording in about two minutes with the small model. A GPU is the better choice only when you need the large models quickly.
- Can Whisper transcribe in real time?Yes, but not out of the box. Open-source Whisper was built to transcribe 30-second chunks of audio, so real-time use means running it repeatedly on a moving window of sound, which adds delay and can revise words. OpenAI also offers a hosted streaming Whisper model, and whether either keeps up depends on your computer, your model size and how much delay you accept.
- Can word error rate be greater than 100%?Yes. Word error rate has no ceiling at 100 percent, because insertions add to the numerator while the denominator is only the reference word count. A system that writes two wrong words for every reference word scores 200 percent, so a WER above 100 percent means the output contains more errors than the reference has words.
- Can you disable or uninstall Speech Recognition and Synthesis from Google?We could not confirm from any page we opened that you can uninstall Speech Recognition & Synthesis from Google, and Google's listing does not mention removing it. Turning it off would likely stop speech-to-text and read-aloud features in apps that use it. The setting the listing describes is the preferred voice input engine, so you can test a different engine there before you disable anything.
- Can you run speech recognition on a Raspberry Pi?Yes. A Raspberry Pi can run speech recognition offline, but the result depends on the task. Home Assistant reports that Whisper takes about 8 seconds per voice command on a Raspberry Pi 4, while its closed command model transcribes in under a second. Open dictation on a Pi is slow, and an ESP32 chip handles only a fixed list of spoken commands.
- Does Gemini have a speech-to-text API?Yes. Google's Gemini Developer API includes Gemini 3.5 Transcribe, a dedicated speech-to-text model for audio files, and Gemini 3.5 Transcribe Live, a streaming model for live audio. You can also send audio to a general Gemini model in an ordinary prompt and ask for a transcript.
- Does Whisper AI have an API?Yes. OpenAI offers Whisper-family transcription through a paid speech-to-text API, and you call it with a secret API key. You do not need that API to use Whisper: the open-source model is free to download and run on your own computer.
- How are speech recognition models trained?Speech recognition models are trained on recordings paired with human-written transcripts. The model predicts text from each clip, and training adjusts its weights until those predictions match the transcript. Public datasets such as LibriSpeech and Common Voice supply much of the audio, and methods such as data augmentation and pre-training make a fixed set of recordings go further.
- How do you add speech recognition to a React or React Native app?In a React web app, use the browser's Web Speech API, directly or through a hook such as react-speech-recognition, and expect it to work mainly in Chrome. In a React Native app, use a native library such as expo-speech-recognition, and create a development build first if your Expo project has no native folders.
- How do you add speech recognition to a Unity or Unreal game?Unity has a built-in dictation class, DictationRecognizer, but it works only on Windows 10 and only when the user allows dictation. For Unreal Engine 5 and Godot, our searches found no built-in recognizer, so you add a third-party plugin and check its license and platforms before you ship.
- How do you calculate word error rate in Python?Install the jiwer package with pip install jiwer, then call jiwer.wer(reference, hypothesis). It returns a fraction, so 0.3 means 30 percent. Normalize case and punctuation first, because jiwer does not do that by default.
- How do you calculate word error rate?Word error rate (WER) is the number of substitutions (S), deletions (D) and insertions (I) needed to turn the recognizer's output into the reference transcript, divided by the number of words in the reference (N): WER = (S + D + I) / N. Align the two texts, count the three error types, add them up and divide.
- How do you get a Google Speech-to-Text API key?You create an API key on the Credentials page of a Google Cloud project that has Speech-to-Text enabled and billing turned on. Google's Speech-to-Text authentication docs describe gcloud and Application Default Credentials and do not list API keys, so send one short test request with the new key before you build on it.
- How do you install Whisper AI?Install Python 3.8 to 3.11 and ffmpeg, then run pip install -U openai-whisper in a terminal. Then run whisper audio.mp3 --model turbo to transcribe a file. The official route is this Python package, not a downloadable installer.
- How do you learn speech recognition?Learn the basics of Python and probability, read the free online draft of Jurafsky and Martin's textbook, follow a structured course such as Hugging Face's Audio Course, then practice on public data. Measure your results with word error rate, using the same text normalization as the Open ASR Leaderboard.
- How do you run Whisper in Google Colab?Open a new Colab notebook, choose a T4 GPU under Change runtime type, install Whisper with pip, and transcribe your file. Free Colab sessions are not guaranteed and can run for at most 12 hours, so download your transcript before the session ends.
- How do you use the Azure Speech-to-Text API?To use the Azure Speech to text API, create a Speech resource, send your audio to its REST endpoint with your resource key in the Ocp-Apim-Subscription-Key header and a language in the URL, and read the transcript from the JSON reply. The short-audio REST call takes up to 60 seconds of audio, so longer recordings go to batch transcription, which Microsoft says can take up to 24 hours to finish at peak times.
- How do you use the Google Speech-to-Text API?Enable the Speech-to-Text API in a Google Cloud project with billing turned on, sign in with the gcloud CLI so your code finds credentials, install the google-cloud-speech package, and call the v2 client with a short audio file. Synchronous calls take 1 minute of audio or less, so use the asynchronous method for longer recordings.
- How does speech recognition know when you stop talking?A voice activity detector labels each slice of audio as speech or silence, and an endpointer decides that a silence has lasted long enough to count as the end of your turn. A longer silence timer cuts off fewer sentences but delays the final text, so every setting is a trade-off. Whisper's no-speech threshold is a different check: it skips silent 30-second windows after decoding and does not decide when you have finished.
- How much does the ElevenLabs speech-to-text API cost, and how do you use it?The ElevenLabs speech-to-text API bills by the hour of audio you send. Its Scribe v2 model lists at $0.22 per hour, the real-time version at $0.39 per hour, and keyterm prompting and transcript editing cost extra. To use it, create an API key in the ElevenLabs dashboard and POST your audio file to the speech-to-text endpoint with the model_id set to scribe_v2.
- How much does the OpenAI speech-to-text API cost?OpenAI's list prices for speech-to-text run from $0.003 per minute for gpt-4o-mini-transcribe to $0.006 per minute for gpt-4o-transcribe. Live transcription with gpt-realtime-whisper is $0.017 per minute. We found no whisper-1 price on the price list we read on 2026-10-17, so check that row before you budget for it.
- Is ElevenLabs speech to text better than Whisper?On the independent AA-WER v2.0 benchmark, ElevenLabs Scribe v2 ranked first at 2.3 percent, and the article's key results do not include Whisper. No published test we found runs both on the same audio, though, and Whisper is free and runs on your own computer while Scribe v2 is a paid cloud API. Choose Scribe v2 when accuracy and speaker features matter most, and Whisper when cost, privacy or offline use matters more.
- Is Google Speech-to-Text free?No, not for regular use. Each account gets 60 minutes a month at no charge on Google's V1 price table, and after that standard recognition costs $0.016 per minute with data logging or $0.024 without it. A $300 credit under Google Cloud's free program can pay for a trial.
- Is on-device speech recognition as accurate as cloud?For short dictation in major languages, on-device recognition is now close to cloud, and some studies and vendor reports put it level. Accuracy depends more on the engine, the language and the audio than on where the model runs. Cloud services can still run larger models than a phone, and some less common languages may only work well in the cloud, so test both on your own recordings before you choose.
- Is speech recognition AI?Yes. Speech recognition is a field within artificial intelligence, and modern systems learn to map sound to words from large collections of recorded speech, and many of them use neural networks. Early systems from the 1950s and 1960s recognized only a handful of words.
- Is speech recognition less accurate for some people?Yes. A 2020 study of five commercial systems found an average word error rate of 0.35 for Black speakers and 0.19 for white speakers. People with speech impairments, young children and some accents also get higher error rates, mostly because the training data contains less of their speech, and the gaps can be narrowed.
- Is speech recognition more accurate than humans?Not in general. On one telephone benchmark in 2016, a Microsoft system matched professional transcribers, at 5.8 percent word error against their 5.9 percent. Humans remain far more accurate in noise and on speech the system has not learned, so the answer depends on the audio and on which human you compare with.
- Is speech recognition part of NLP?Partly. Speech recognition turns sound into words, while natural language processing (NLP) works on the meaning of language. A widely used textbook treats speech recognition as one of the topics of language processing, so it counts as part of NLP in the broad sense. A tool that only writes what you say uses less NLP than an assistant that acts on your words.
- Is there a free speech-to-text API?Yes, but only as an allowance. Google includes 60 free minutes a month, Azure 5 audio hours a month of real-time transcription, AssemblyAI up to 185 hours of recorded audio, and Deepgram a $200 credit, after which each bills. Among these, open-source Whisper is the only option with no per-minute bill. It is free to license but needs your own hardware.
- Is there an official Whisper AI app?We found no official Whisper app from OpenAI. OpenAI publishes Whisper as open-source software and offers hosted transcription through its API, while dictation in ChatGPT is a feature of that product. The apps you can install and use with Whisper come from other developers, so check who makes each one.
- Is Whisper AI free?Yes. OpenAI releases Whisper's code and model weights under the MIT License, which charges no fee to download or run the model. Costs come from OpenAI's hosted API, from paid apps built on Whisper, and from the computer you need to run it yourself.
- Is Whisper AI the same as Wispr Flow?No. Whisper is an open-source speech model from OpenAI, while Wispr Flow is a dictation app made by a separate company. A June 2025 report said Wispr Flow's intelligence mixes Meta's Llama with Whisper, and Wispr Flow sends dictated audio to its servers, so it is a different product with a different privacy model.
- Is Whisper more accurate than Google Speech-to-Text?Not on every test. In one 2024 study of English lecture recordings, Whisper large-v2 made far fewer word errors than Google's service, but a 2024 classroom study found Google more robust to noise, so the answer depends on your audio, the model you pick and the date of the test.
- Is word error rate the same as accuracy?No. Word error rate is the share of errors, and word accuracy is 1 minus WER, so the two carry the same information. The two can differ because some tools report percent correct, which leaves out insertions, and because accuracy can fall below zero when a system writes many extra words.
- What are the best alternatives to Whisper?For English, a strong open alternative to Whisper is NVIDIA's Parakeet TDT 0.6B v2, which beat Whisper Large v3 on the leaderboard and ran much faster. For the longest list of languages, test Meta's Omnilingual ASR, and for devices too small for Whisper choose Vosk, but test each one on your own recordings before you switch.
- What are the best datasets for speech recognition?There is no single best dataset. For English, LibriSpeech is the most quoted free test set, with 1,000 hours of read speech under CC BY 4.0, and GigaSpeech has 10,000 hours, but its audio is offered only for non-commercial research and education. Choose by the speech and language you need, then check the license.
- What are the best Whisper settings for accuracy?Start with the turbo model, or large if your hardware has the memory for it. Set the language yourself, add a short prompt of names and jargon, and keep the default temperature fallback. Leave the no-speech threshold at its default of 0.6 unless a test shows dropped speech or invented text.
- What are the most important speech recognition research papers?The most important papers changed either the method or the way results are measured. Start with Rabiner's 1989 tutorial on hidden Markov models, then read Shannon and colleagues' 1995 study of temporal cues, the 2012 deep neural network review, and the 2022 Whisper paper. For new work, follow IEEE Transactions on Audio, Speech and Language Processing and Speech Communication.
- What is a good word error rate?There is no single good word error rate, because the target depends on the audio and the speaker. On typical speech a WER below 5 percent is the benchmark, while speech from people with dysarthria or Parkinson's disease can score several times higher. Judge a score by whether it meets the accuracy you need on audio like yours.
- What is a speech recognition threshold in a hearing test?A speech recognition threshold is the softest level at which a person can repeat or identify half of a list of spoken words in a hearing test, usually two-syllable spondee words played through headphones. It checks the pure-tone hearing test, and it measures a person's hearing, not how accurate speech recognition software is.
- What is IBM Watson Speech to Text?IBM Watson Speech to Text is IBM's cloud service that turns audio into text through an API, with pre-trained speech models and the option to tune them to improve accuracy. Its free Lite plan includes 500 minutes a month. In one 2024 test on English lecture recordings it made more errors than Whisper large-v2, so test it on your own audio before you choose it.
- What is Kaldi speech recognition?Kaldi is an open-source speech recognition toolkit, written in C++ and licensed under Apache 2.0, for building a recognizer rather than dictating with a finished model. Its project says it is intended for speech recognition researchers, so most people who want to dictate will get more from a ready-made model or app.
- What is NVIDIA Parakeet, and is it better than Whisper?NVIDIA Parakeet is a family of open speech recognition models. Parakeet TDT 0.6B v2 beats Whisper Large v3 on English benchmarks and runs much faster, but it covers only English, and in one 2025 test of a single episode it made more errors than Whisper Large V3 Turbo, so it is not better in every case.
- What is Speech Recognition and Synthesis from Google?Speech Recognition & Synthesis from Google is the speech engine on Android that turns your speech into text for Google and other apps, and reads on-screen text aloud. You can select it as your voice input engine in your phone's settings. Its Google Play listing does not say whether recognition runs on your phone or on Google's servers.
- What is speech recognition used for?Speech recognition turns spoken words into text for dictation, live captions, voice assistants, relay calls for deaf and hard of hearing people, and medical and legal records. It is least reliable when the audio differs from the speech it was trained on, so test it on your own recordings before you rely on it.
- What is the best GPU for Whisper?There is no single best GPU for Whisper. Choose a card with enough video memory for the model you run: about 6 GB for turbo and about 10 GB for large. On a Mac, use whisper.cpp or MLX, which are built for Apple silicon.
- What is the best Whisper GUI or desktop app?No single app is the best for everyone. For a free, open-source desktop app that runs Whisper on your own computer and transcribes files, Buzz is the closest match. For dictating into any app with a keyboard shortcut, Handy and Superwhisper are built for that job, and MacWhisper suits Mac users who transcribe files.
- What is the cheapest speech-to-text API?AssemblyAI's Universal-2 model has the lowest published list rate we checked, at $0.15 per hour of audio, with OpenAI's gpt-4o-mini-transcribe next at $0.18. For small volumes, free allowances can cost less still, and the cheapest bill only matters if the model also meets your accuracy and data needs.
- What is the difference between speech recognition and voice recognition?Speech recognition converts spoken words into text and answers what was said. Voice recognition is an ambiguous term: it can mean speech recognition, or speaker recognition, which identifies or verifies who is speaking. Speaker recognition can identify a person from their voice, but a voice match is not proof of identity.
- What is the difference between WER and character error rate?Word error rate (WER) and character error rate (CER) use the same count of substitutions, deletions and insertions, but WER counts words and CER counts characters. Use WER for languages written with spaces between words, use CER for Mandarin and other scripts without word spaces, and report both for OCR, where one wrong letter can make a whole word wrong.
- What is the difference between whisper.cpp, faster-whisper and WhisperX?Pick whisper.cpp for a native build that runs without Python, especially on a Mac or a CPU. Pick faster-whisper for Python work, including transcription on a CPU in int8 mode. Pick WhisperX when you need word-level timestamps or speaker labels, since it adds those to faster-whisper's output.
- What is the fastest speech-to-text API?In the 2026 streaming benchmark by Artificial Analysis, Deepgram Flux was the fastest service, returning its first partial transcript 0.019 seconds after speech ended. It was not the most accurate, and the answer changes with the measure: first words, final text or throughput on long files, so time the service on your own audio before you choose.
- What is the most accurate speech-to-text model?No single speech-to-text model is the most accurate for every speaker and setting. On the Open ASR Leaderboard's English tests, the best average word error rate in 2026 was about 5.3 to 5.4 percent, shared by several open and proprietary systems, and the leader changes with each update. Test the top few on your own recordings before you choose.
- What is the Open ASR Leaderboard?The Open ASR Leaderboard is a public scoreboard that runs open-source and proprietary speech recognition systems on the same test audio, with the same scoring code. It reports word error rate for accuracy and inverse real-time factor for speed, so you can compare systems on equal terms, but only on the test sets it uses.
- What is Vosk speech recognition?Vosk is an open-source speech recognition toolkit from Alpha Cephei that runs offline, even on small devices such as the Raspberry Pi. It supports more than 20 languages, and its per-language models are about 50 MB. It is best suited to small devices and fixed commands, and for long, noisy dictation you should test it against a neural model first.
- What languages does Whisper support?Whisper accepts 98 languages, according to OpenAI, and its accuracy varies widely from one language to another. Hebrew, Persian, Arabic and Indonesian have codes in its language table, but the results are uneven, and Kurdish has no code in the table at all.
- What role do hidden Markov models play in speech recognition?Hidden Markov models (HMMs) handled the timing of speech in most recognizers from the early 1970s until about 2012. An HMM models a word as a sequence of hidden sound states, and a separate model scored how well each state fit the audio. Neural networks took over that scoring from about 2012, and end-to-end networks now often replace the HMM entirely.
- Which audio formats and file sizes does Whisper accept?Yes, Whisper accepts m4a. OpenAI's transcription API takes mp3, mp4, mpeg, mpga, m4a, wav and webm files up to 25 MB each, and the open-source package reads any format that ffmpeg can decode, while each app sets its own limits.
- Which neural network is best for speech recognition?No single network is best for every job. Transducers are the usual choice for dictation that appears as you speak, because they stream text while the audio arrives. For transcribing recordings, Conformer-based encoders and Whisper's encoder-decoder are strong, so test both on your own audio before you choose.
- Which speech-to-text API is the best?No single speech-to-text API is best for every job. On clean English the strongest systems are close on independent benchmarks, so the best choice is the one with the fewest errors on your own audio that also meets your needs for speed, speaker labels and data terms.
- Which speech-to-text APIs can identify speakers?Several hosted speech-to-text APIs label speakers, including Google Cloud, Amazon Transcribe, AssemblyAI, Deepgram, ElevenLabs and OpenAI's gpt-4o-transcribe-diarize. Whisper on its own does not identify speakers, but the open-source WhisperX project adds speaker diarization to Whisper's output, so the choice depends on language, price and how you run the model.
- Which Whisper model should I use?For most English recordings, use the turbo model. It needs about 6 GB of video memory and scores close to the largest model on English accuracy tests. Pick medium or large for translation, and a smaller model only when your computer cannot run turbo.
- Who invented speech recognition?No single person invented speech recognition. Bell Laboratories built a digit recognizer in 1952, IBM's Shoebox is usually credited with 16 spoken words in 1961, and later research teams built the statistical and neural methods that today's systems use.
- Who owns Whisper AI?OpenAI publishes Whisper, and its nonprofit OpenAI Foundation controls the for-profit company, OpenAI Group PBC. Whisper is not a separate company. Its code and weights are released under the MIT License, so anyone can use them.
- Why do speech-to-text accuracy claims disagree?Speech-to-text accuracy claims disagree because each one uses different audio, a different reference transcript, different scoring rules and sometimes a test set the model has already seen. The same model can score 6.05 percent on one public benchmark and 12.3 percent in a single 2025 test. Compare only numbers that share a method, and test on your own recordings.
- Why does speech recognition get names and jargon wrong?Names and jargon are rare in the speech a recognizer learns from, so the model often picks a more common word that sounds similar. Measured error rates for medical dictation were 14 to 15 percent in 2000, and a 2024 study still found significant errors on rare entity names, so check and add your own terms before you rely on the text.
- Why does speech recognition show a network or not-allowed error?A **network** error means the network communication that speech recognition needs failed, so the request never finished. A **not-allowed** error means the browser blocked speech input for security, privacy or user preference reasons, and a refused microphone permission is one common cause. MDN lists Firefox's support for speech recognition only as preview, so a page there cannot rely on the feature.
- Why does Whisper make up words nobody said?Whisper writes the text its training makes most likely, so on silence, background noise or unclear audio it can produce fluent words that nobody said. This is called hallucination. A 2024 study found that about 1 percent of its transcriptions contained whole invented phrases, so proofread text that follows a long pause and compare it with what you said.
- Why is Whisper not using my GPU?Whisper uses your GPU only when PyTorch can detect an NVIDIA card with CUDA. If it cannot, openai-whisper quietly runs on the processor. Run one check in Python: if it prints False, install a CUDA build of PyTorch, then run Whisper again.
Terms in this topic
- Acoustic modelAn acoustic model is the part of a speech recognizer that scores how well a short window of audio matches the sounds of speech. Since about 2012 most acoustic models are deep neural networks, and their training data strongly affects accuracy.
- Automatic speech recognition (ASR)Automatic speech recognition (ASR) is software that converts spoken audio into written words. Modern systems are neural networks trained on large amounts of transcribed speech, and they run in the cloud or on the device.
- ConformerA Conformer is a neural encoder for speech recognition that combines self-attention, which sees the whole input, with convolution, which captures local sound patterns. FastConformer is NVIDIA's faster variant of the same design, used in streaming models.
- Connectionist temporal classification (CTC)Connectionist temporal classification (CTC) is a way to train a neural network to turn audio into text without knowing in advance which moment of sound matches which letter. The network labels every short slice of audio with a character or a blank, and the blanks and repeated labels are then removed to leave the transcript.
- Hidden Markov model (HMM)A hidden Markov model (HMM) is a statistical model in which a sequence of hidden states produces the observations you can measure, and each state changes to the next with set probabilities. In speech recognition, the hidden states are units of sound and the observations are short frames of audio.
- Inverse text normalization (ITN)Inverse text normalization (ITN) is the step that converts a speech recognizer's spoken-form output into written form, so that 'two hundred' becomes 200. It changes how the text reads, not which words the recognizer heard.
- Language model (in speech recognition)In speech recognition, a language model (LM) scores how probable a word sequence is in a language. The recognizer combines that score with the acoustic evidence to choose between words that sound alike, such as homophones.
- LibriSpeechLibriSpeech is a free corpus of about 1,000 hours of read English speech, taken from LibriVox audiobooks, and it is one of the most widely used benchmarks for English speech recognition. Its scores describe clean, read speech, which differs from everyday dictation.
- Mel spectrogramA Mel spectrogram is a grid of sound energy over time, with frequency bands spaced to match human hearing. Many speech recognizers compute its logarithm, the log-Mel spectrogram, and feed it to a neural network instead of the raw waveform.
- Real-time factor (RTF)The real-time factor (RTF) is the time a system takes to process audio divided by the length of that audio. An RTF of 1 or lower means the system keeps up with real-time speech. Some leaderboards report the inverse, RTFx, where higher is faster.
- RNN transducer (RNN-T)An RNN transducer (RNN-T) is a neural speech model that reads audio in order and emits words as it goes, so text can appear while the speaker is still talking. It has an encoder for the sound, a prediction network for the words already written, and a joint network that chooses the next output.
- Speaker diarizationSpeaker diarization is the task of working out who spoke when in a recording, by splitting the audio into turns and grouping turns that come from the same voice. It is separate from transcribing the words, and a transcript gains speaker labels only when diarization is run alongside recognition.
- Voice activity detection (VAD)Voice activity detection (VAD) is a component that decides whether a stretch of audio contains speech. Speech recognizers use it to split audio into turns, skip silence and decide when a speaker has finished.
- wav2vec 2.0wav2vec 2.0 is a speech model that first learns the structure of speech from large amounts of unlabeled audio, then is fine-tuned on a much smaller set of transcribed recordings. Because transcribed speech is expensive to make, this approach helps when labeled data is scarce.
- Word error rate (WER)Word error rate (WER) is the number of word substitutions, deletions and insertions in a transcript, divided by the number of words in the correct transcript. Lower is better, and a score above 100 percent is possible.
Referenced by
- Dictation Software: How to Choose the Right AppTopic
- Typing Speed: Average WPM and How to Type FasterTopic
- Mac Dictation: How to Dictate and Control a Mac by VoiceTopic
- Dictation on Windows: Voice Typing, Voice Access, AppsTopic
- Speech to Text for Accessibility, RSI and DisabilitiesTopic
- Dictation for Work: Doctors, Lawyers, Writers, StudentsTopic
- Voice Typing in Google Docs, Word and Every Other AppTopic
- Dictation and Voice to Text on iPhone and AndroidTopic
- Transcription: How to Turn Audio and Video Into TextTopic
- Speech to Text Privacy: Where Your Voice GoesTopic
- Dictation Languages, Accents and Voice TranslationTopic
- Speech to Text AI: Dictation, Assistants and Voice AgentsTopic
- How Speech Recognition Works, From Sound to TextGuide
- What Is Whisper AI? OpenAI's Speech Model ExplainedGuide
- Speech-to-Text APIs Compared: Price, Accuracy, SpeedGuide
- Word Error Rate (WER): How to Calculate and Read ItGuide
- Speech to Text and LLMs: What Language Models ChangeGuide
- Best Dictation Apps for iPhone and AndroidGuide
- Speech Recognition Models: How Neural ASR WorksGuide
- Open-Source Speech Recognition: The Best ModelsGuide
- Dictation vs Typing: When Speaking Beats the KeyboardGuide
- How to Run Whisper Locally on Your Own ComputerGuide
- Speech Recognition Benchmarks and LeaderboardsGuide
- Real-Time Speech Recognition: Streaming and LatencyGuide