Facts

Transcription: facts and numbers

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

A 2025 benchmark of speech models for African languages lists the NCHLT corpus as a source of read speech in Zulu, Xhosa and Afrikaans.arXiv, 2025
A 2025 study of non-native English speech found that speech recognition systems vary significantly in how they handle filler words, repetitions and revisions.Michael McGuire, arXiv, 2025
A 2025 study on dementia speech says that Whisper tends to omit filler words such as uh and um from its transcripts.Emmanuel Akinrintoyo et al., arXiv, 2025
A 2025 user comment on Goodnotes' feedback forum asks for audio files and MP3 files to be transcribed, not only live recordings, which suggests that imported audio is not transcribed by the app today.Goodnotes feedback forum, 2026
4-9 hours of transcribing per hour of audio or videoA Claremont Colleges oral history guide says that by some estimates one hour of audio or video can take 4 to 9 hours to transcribe, depending on the subject, the number of speakers and the audio quality.Claremont Colleges Library, Claremont Colleges Library research guides, n.d.
4-6 hours of transcribing per hour of recorded interviewA Cornell interview project guide tells interviewers to expect to spend 4 to 6 hours transcribing each hour of recorded interview by hand.Cornell University, Cornell University, Creating Profiles project guide, n.d.
A dictionary entry defines transcribing as listening to something and writing down what was said, word for word, and notes that speeches, interviews and trials are often transcribed for records.Vocabulary.com, n.d.
A federal class action filed in August 2025 alleges that Otter.ai's Otter Notebook service, by default, does not ask meeting attendees for permission to record.NPR, 2025
April 26, 2027 compliance dateA Justice Department interim final rule moved the compliance date for state and local governments with 50,000 or more people to April 26, 2027.US Department of Justice, ADA.gov, 2026
175 minutesA one-hour recording at a 150 words per minute speaking rate holds about 9,000 words, which takes about 175 minutes to type at the 51.56 words per minute online average.Calculated from other facts
A WebVTT file must be UTF-8 text labeled with the MIME type text/vtt.W3C Timed Text Working Group, W3C, 2026
95 percentAAERT lowered the CET practical exam passing score from 98 percent to 95 percent for exams taken on or after March 26, 2026.American Association of Electronic Reporters and Transcribers (AAERT), 2026
Adobe's Speech to Text page for Premiere Pro says the feature is powered for English, Spanish, Japanese, Korean, French, German, Chinese and Hindi, and more.Adobe, 2026
80.4 percent word error rate (isiXhosa)After fine-tuning Whisper medium on the same amount of transcribed child speech, the word error rate was 80.4 percent for isiXhosa, against 47.4 percent for Afrikaans.Christiaan Jacobs et al., arXiv (Stellenbosch University and University of Edinburgh), 2025
25 seconds per hour of audioAn inverse real-time factor of 146 means the leaderboard engine processes about 146 seconds of audio for each second of compute, so one hour of audio takes about 25 seconds of compute time.Calculated from other facts
5 minutes between timestampsAn oral history guide from Oregon State University Libraries tells transcribers to add timestamps at five-minute intervals.Oregon State University Libraries and Press, Special Collections and Archives Research Center, Oregon State University LibGuides, n.d.
Apple Podcasts automatically generates a transcript after a new episode is published, and the app labels it as automatically generated.Apple Podcasters Support, n.d.
Apple Podcasts can show a name for each speaker in a transcript when the podcaster provides a VTT file.Apple Podcasters Support, n.d.
Apple Podcasts shows transcripts on iOS 17.4 or later for podcasts in English, Danish, Dutch, Finnish, French, German, Italian, Norwegian, Portuguese, Spanish and Swedish.Apple Podcasters Support, n.d.
Apple says a screen recording saved from QuickTime Player on a Mac is an H.264 format QuickTime movie, with an AAC audio track when audio was selected.Apple, 2026
Apple says that Maximum-quality audio recordings in QuickTime Player are uncompressed, and that they can use large amounts of storage space.Apple, 2026
Apple says that, in macOS 27 Golden Gate or later, a screen recording can include system audio through an Include System Audio option.Apple Support, 2026
AssemblyAI's guide to pre-recorded audio says that to transcribe a local file, you must upload it first, so the service can then transcribe it.AssemblyAI, n.d.
0.02 US dollars per hour of audioAssemblyAI's pricing page lists speaker diarization, which splits a transcript by speaker, at $0.02 per hour of audio.AssemblyAI, 2026
442 timesAt $0.0045 per minute an hour of audio costs about $0.27 through OpenAI's API, against about $119.40 at Rev's $1.99 per minute for human transcription, roughly 440 times more.Calculated from other facts
5 one-hour meetings per monthAt 300 transcription minutes a month, the NotePin Starter allowance covers five one-hour meetings.Calculated from other facts
0.96 US dollars per hour of audioAt Google's V2 standard list price, one hour of audio costs about $0.96 before any volume discount.Calculated from other facts
12 timestamps per hourAt one timestamp every five minutes, a one-hour recording gets 12 timestamps.Calculated from other facts
119.40 US dollars per hour of audioAt Rev's list price for human transcription, one hour of audio costs about 119 US dollars.Calculated from other facts
$33 to $120 per audio hour (floor)At Rev's published legal transcriptionist rates, one hour of audio pays from about $33 up to $120 or more.Calculated from other facts
$24 to $66 per audio hourAt Rev's published transcriptionist rates, one hour of audio pays between $24 and $66 before any hours are counted.Calculated from other facts
$8.23 per hour of typingAt the bottom of Rev's transcriptionist range, an audio hour pays about $8.23 for each hour of typing at average speed, before any listening, replay or checking.Calculated from other facts
1.09 million audio tokensAt the stated audio rate, a 9.5-hour recording in Gemini is about 1.09 million audio tokens.Calculated from other facts
$22.63 per hour of typingAt the top of Rev's transcriptionist range, an audio hour pays about $22.63 for each hour of typing at average speed, before any listening, replay or checking.Calculated from other facts
Blackmagic Design's DaVinci Resolve tech specs page, read on 2026-10-17, lists DaVinci Resolve Studio 21 as its current Studio release.Blackmagic Design, n.d.
Blackmagic's DaVinci Resolve 20 New Features Guide labels its improved transcription and extra language support as Studio Version Only.Blackmagic Design, 2025
16 milliseconds (model output resolution)Bredin and Laurent's 2021 segmentation model works on 5-second audio chunks at a temporal resolution of every 16 milliseconds.Hervé Bredin and Antoine Laurent, Interspeech 2021, 2021
CapCut's help article says that auto captions depend on the selected language, and that choosing the wrong language can result in no captions or incorrect transcription.CapCut Help Center, 2026
CapCut's help article says that bugs in older versions can lead to auto captions not working.CapCut Help Center, 2026
CapCut's help article says that caption generation often uses cloud processing, so a slow or unstable connection may interrupt it.CapCut Help Center, 2026
CapCut's help article says that if a video has no audio, or the volume is too low, CapCut cannot generate captions.CapCut Help Center, 2026
CapCut's help article says that if the audio is unclear, distorted or full of background noise, CapCut may fail to detect speech accurately.CapCut Help Center, 2026
Chico State's Zoom guide says the audio transcript option is set under Advanced Cloud Recording settings, in the Zoom settings page for the account.California State University, Chico, IT Services Knowledge Base, undated
Closed captions are not visible until the viewer activates them, usually with the remote control or a menu option.Wikipedia, n.d.
Davidson's 2009 review of the transcription literature opens by calling transcription a practice central to qualitative research, although the literature often treats it as taken for granted.Christina Ruth Davidson, International Journal of Qualitative Methods, 2009
DaVinci Resolve 20 added transcription and automatic subtitling support for many more languages, and the Extended Transcription Language Support is a package you download in the Extras Download Manager.Blackmagic Design, 2025
Diarization systems are reported to be more accurate in balanced conversations with long turns and little turn overlap.Scott Seyfarth et al., arXiv, 2021
Discord's developer documentation says that voice data is carried over UDP in both directions, so a client that joins a voice channel must be able to receive UDP packets.Discord, Discord Developer Portal, 2026
Each entry in an SRT subtitle file is numbered sequentially, starting at 1.Wikipedia, n.d.
200 milliseconds tolerance (typical in prior work)Earlier studies of word-level timestamps usually counted a timestamp as correct if it fell within a tolerance of more than 200 milliseconds.Sung-Lin Yeh et al., arXiv (ASRU 2025), 2025
FFmpeg's subtitles filter draws subtitles onto the frames of a video.FFmpeg project, 2026
5.5 percent word error rateFine-tuning Whisper Large V3 Turbo on labeled filled pauses gave 5.5 percent word error rate, against 6.2 percent when hesitations were removed.Jhen-Ke Lin et al., arXiv (SLaTE 2025 workshop), 2025
For a 2023 UK Data Service anonymisation project of 38 interview transcripts, the recommendations say to edit a copy and keep each original until the whole project is anonymised.UK Data Service, 2023
1 to 8 channelsFor multichannel recognition, Google's Speech-to-Text API accepts from 1 to 8 channels for OGG_OPUS audio, so OGG_OPUS is one of its supported audio encodings.Google Cloud, 2026
For the Whisper paper's text scoring, the authors put a space between every letter for Chinese, Japanese, Thai, Lao and Burmese, which means they measured character error rate for those languages instead of word error rate.Alec Radford et al., OpenAI (arXiv), 2022
6 hours per hour of interviewGläser and Laudel found that verbatim transcription took an average of six hours for each hour of interview material in their own research projects, a figure Wollin-Giering and colleagues report in 2024.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
Goodnotes says transcription of its audio recordings is automatic, and that it converts recordings into searchable, editable text.Goodnotes, 2026
Google announced in September 2024 that users could add public YouTube URLs and audio files to a NotebookLM notebook.Raiza Martin, Google Blog, 2024
Google Cloud's Speech-to-Text supported-languages table lists Hindi (India), code hi-IN, for its chirp_2 model.Google Cloud, Google Cloud documentation, 2026
Google Cloud's Speech-to-Text supported-languages table lists Japanese (Japan), code ja-JP, for its chirp_2 model.Google Cloud, Google Cloud documentation, 2026
Google Cloud's Speech-to-Text supported-languages table lists Korean (South Korea), code ko-KR, for its chirp_2 model.Google Cloud, Google Cloud documentation, 2026
Google Cloud's Speech-to-Text supported-languages table lists Xhosa (South Africa), code xh-ZA, for its chirp_2 model.Google Cloud, Google Cloud documentation, 2026
Google Cloud's Speech-to-Text supported-languages table lists Zulu (South Africa), code zu-ZA, for its chirp_2 model.Google Cloud, Google Cloud documentation, 2026
Google lists Business Standard, Business Plus, Enterprise Standard, Enterprise Plus and Education Plus as the supported editions for Meet transcripts.Google Workspace Admin Help, 2026
Google says both the user and the organization must have enough free storage space for a Meet transcription to be saved to Drive.Google Workspace Admin Help, 2026
Google says Meet transcription is on by default, but someone in the meeting must start it each time.Google Workspace Admin Help, 2026
Google says Meet transcripts are saved to the meeting organizer's Drive after the meeting.Google Workspace Admin Help, 2026
Google says meeting transcription is on by default for the organization, but someone in the meeting has to start transcription for each meeting.Google Workspace Admin Help, 2026
Google says on Business Standard an admin cannot turn the Meet transcription feature off in the Admin console.Google Workspace Admin Help, 2026
Google says that before joining a meeting, attendees see a notification informing them when transcripts are active.Google Workspace Updates, Google, 2022
Google says the accuracy of its speech recognition can be reduced when lossy codecs are used to capture or transmit audio, particularly if background noise is present.Google Cloud, 2026
Google's 2024 announcement says that when you add audio recordings to NotebookLM, it can search across the transcribed conversations to find specific information.Raiza Martin, Google Blog, 2024
Google's admin help says to check that you have a Google Workspace edition that supports transcription before you set up Meet transcripts.Google Workspace Admin Help, 2026
Google's administrator help lists Business Standard, Business Plus, Enterprise Standard and Enterprise Plus, and Education Plus as the Google Workspace editions that support meeting transcription.Google Workspace Admin Help, 2026
116 Chrome versionGoogle's developer guide says that from Chrome 116, an extension can call the tabCapture API in a service worker to obtain a stream ID after a user gesture, and pass it to an offscreen document that records the tab.Chrome for Developers, 2026
9.5 hours of audio per promptGoogle's Gemini API documentation gives the maximum audio length as 9.5 hours per prompt.Google, Google AI for Developers, 2026
Google's Gemini audio documentation lists transcription and translation of speech among the things a Gemini model can do with audio input, alongside describing, summarizing and answering questions about the audio.Google, Google AI for Developers, 2026
Google's Gemini audio overview lists describing, summarizing and answering questions about audio as use cases, alongside transcription.Google, Google AI for Developers, 2026
Google's launch post says that when a YouTube video is added to NotebookLM, it summarizes key concepts and links its citations to the video's transcript.Raiza Martin, Google Blog, 2024
GoTranscript defines clean verbatim as keeping the meaning while removing obvious filler words, stumbles and minor repetition.Daniel Chang, GoTranscript, 2026
GoTranscript says an audio minute is the length of the file, not the time it takes to transcribe it, so its pay is set by file length.GoTranscript, 2026
GoTranscript says edited transcription rewrites for clarity and flow, which helps readability but moves the text further from the spoken record.Daniel Chang, GoTranscript, 2026
GoTranscript says important background sounds should be noted, with the cues inserted in square brackets.GoTranscript, GoTranscript, n.d.
GoTranscript says its pay varies by language, audio length, difficulty, turnaround and quality, so no single rate applies to every file.GoTranscript, 2026
GoTranscript says to choose American or British English spelling and keep it uniform across the transcript.GoTranscript, GoTranscript, n.d.
GoTranscript's formatting guide describes interval timestamps, placed at regular intervals such as every one or two minutes.GoTranscript, GoTranscript, n.d.
GoTranscript's formatting guide describes speaker change timestamps: a timestamp inserted each time a new speaker talks.GoTranscript, GoTranscript, n.d.
GoTranscript's formatting guide says to always add the timestamp to inaudible and crosstalk tags, for easy reference.GoTranscript, GoTranscript, n.d.
GoTranscript's formatting guide says to format each speaker line as the speaker's name, a colon, and the text of the speech.GoTranscript, GoTranscript, n.d.
GoTranscript's jobs page lists no application fee, weekly payments for approved work, and a two-step application of guidelines and a short test.GoTranscript, 2026
GoTranscript's style advice is to mark overlapping speech that cannot be separated as [crosstalk mm:ss].GoTranscript, GoTranscript, n.d.
GoTranscript's style advice is to mark speech that cannot be understood as [inaudible mm:ss], with the timestamp.GoTranscript, GoTranscript, n.d.
If a Zoom meeting shuts down unexpectedly, or the conversion of a computer recording is interrupted, the recording files can become corrupted and cannot be recovered.University of Illinois Urbana-Champaign IT help desk, undated
If the chat bubble icon does not appear in the Notes app, the iPhone either does not meet the minimum requirements or does not detect a supported language.Bowdoin College IT knowledge base, n.d.
In 2018, the authors said very limited code-switched speech resources were available, so they built a Hindi-English speech database.Ganji Sreeram et al., arXiv, 2018
In 2025 Instagram added a View transcription button under voice clips in direct messages, so the recipient can read the words of a voice message.Social Media Today, 2025
5 minutesIn 2025 Instagram raised the maximum length of voice clips in direct messages from 1 minute to 5 minutes.Social Media Today, 2025
In a 2017 comparison of machine and vendor transcripts on telephone speech, the automatic recognizer tended to confuse filled pauses such as uh with backchannels such as uhhuh, and humans tended not to make that error.Andreas Stolcke and Jasha Droppo, Proceedings of Interspeech 2017, ISCA, 2017
In a 2024 Interspeech study, adjusting the tokenizer of the Whisper model significantly improved the precision of its word-level timestamps.Laurin Wagner et al., Interspeech 2024, 2024
In a 2024 review, the editor Larry Jordan described Adobe's Speech-to-Text engine as powerful but not very accurate, and estimated that about 80 percent of its output on three of his videos was text he did not have to transcribe himself.Larry Jordan, Larry Jordan (larryjordan.com), 2024
7 times real timeIn a 2024 study of research transcripts, transcribing from scratch and correcting automatic output took equally long, about 7 times real time on average, for the data and conventions tested.Jan Gorisch and Thomas Schmidt, Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024
1 percentIn a 2024 study, roughly 1 percent of Whisper audio transcriptions contained entire hallucinated phrases or sentences that did not exist in the audio.Allison Koenecke et al., arXiv (ACM FAccT 2024), 2024
In a 2024 walkthrough, Larry Jordan made captions in Premiere Pro from his corrected transcript and exported them as a separate SRT file for YouTube.Larry Jordan, Larry Jordan (larryjordan.com), 2024
5.81 percent word error rateIn a 2025 challenge on non-native English, a fine-tuned Whisper system scored 5.81 percent word error rate when it labeled hesitations, against 6.47 percent when it removed them.Jhen-Ke Lin et al., arXiv (SLaTE 2025 workshop), 2025
6.3 times the interview durationIn a study of more than 200 measurements, students needed on average 6.3 times the interview duration to transcribe an interview manually.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
In a Zoom meeting, the host can turn captions off for everyone from the caret next to the Show Captions button, under Host Controls.Brown University IT, undated
In DaVinci Resolve 20, to transcribe clips for IntelliScript, you select them in the Media Pool, right-click, and choose AI Tools > Audio Transcription > Transcribe.Blackmagic Design, 2025
In Dragon 15 for Windows, the step to transcribe a recording asks you to select who the voice in the recording belongs to, and the choices are Me or Someone else.Nuance Communications, 2018
In Dragon 15 for Windows, you can transcribe a recording with your own user profile.Nuance Communications, 2018
In Dragon 15 on Windows, you start a recording transcription from the DragonBar by choosing Tools, then Transcribe Audio, then Transcribe Recording.Nuance Communications, 2018
In nursing research, verbatim transcription of interview data is a common data management strategy that is widely considered integral to analysing verbal data.Elizabeth Halcomb and Patricia Davidson, Applied Nursing Research, 2006
In Premiere Pro, the Automatically Transcribe Clips option is in the Transcription settings, and when it is on, transcription starts in the background as soon as you import a clip.Larry Jordan, Larry Jordan (larryjordan.com), 2024
44100 hertzIn QuickTime Player on a Mac, the High recording quality setting saves audio as 44100 Hz AAC, and the Maximum setting saves uncompressed 24-bit AIFF-C audio at 44100 Hz.Apple, 2026
4.4 percent average word error rateIn the 2024 study, Amazon had an average word error rate of 4.4 percent on the English lecture recordings.Korbinian Kuhn et al., arXiv, 2024
8.3 percent average word error rateIn the 2024 study, Deepgram had an average word error rate of 8.3 percent on the English lecture recordings.Korbinian Kuhn et al., arXiv, 2024
11.2 percent average word error rateIn the 2024 study, IBM had an average word error rate of 11.2 percent on the English lecture recordings.Korbinian Kuhn et al., arXiv, 2024
4.4 percent average word error rateIn the 2024 study, Microsoft had an average word error rate of 4.4 percent on the English lecture recordings.Korbinian Kuhn et al., arXiv, 2024
7 calendar daysIn the maximum transcript rate schedule effective October 1, 2024, an expedited transcript is one delivered within seven calendar days after the court receives the order.U.S. Bankruptcy Court, Southern District of New York, 2024
30 calendar daysIn the maximum transcript rate schedule effective October 1, 2024, an ordinary transcript is one delivered within thirty calendar days after the court receives the order.U.S. Bankruptcy Court, Southern District of New York, 2024
2 hoursIn the maximum transcript rate schedule effective October 1, 2024, the hourly category is a transcript of proceedings delivered within two hours from receipt of the order.U.S. Bankruptcy Court, Southern District of New York, 2024
In the Notes app on iPhone, a recording made inside a note shows its transcript when you tap the chat bubble icon, and spoken words in supported languages are transcribed automatically.Bowdoin College IT knowledge base, n.d.
70 minutes saved per hour of interviewIn the same F4x study, students needed on average 5.1 times the interview duration to correct an automatic transcript, which the study said saved 70 minutes of correction time per hour of interview.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
25.25 percent word error rateIn the same Voxtral table, Gemini 2.5 Flash had a word error rate of 25.25 percent in Arabic, the highest of the three systems.Mistral AI, arXiv (Audio and Speech Processing), 2025
6.76 percent word error rateIn the same Voxtral table, Gemini 2.5 Flash had a word error rate of 6.76 percent in Hindi.Mistral AI, arXiv (Audio and Speech Processing), 2025
12.36 percent word error rateIn the same Voxtral table, GPT-4o mini Transcribe had a word error rate of 12.36 percent in Hindi.Mistral AI, arXiv (Audio and Speech Processing), 2025
14.02 percent word error rateIn the same Voxtral table, GPT-4o mini Transcribe had a word error rate of 14.02 percent in Arabic.Mistral AI, arXiv (Audio and Speech Processing), 2025
27.90 percent DER, Track 2In the second DIHARD challenge, the DKU-LENOVO team reached a diarization error rate of 27.90 percent on Track 2, from the same system that reached 18.84 percent on Track 1.Qingjian Lin et al., arXiv, 2020
18.84 percent DER, Track 1In the second DIHARD challenge, the DKU-LENOVO team's diarization system reached a diarization error rate of 18.84 percent on Track 1.Qingjian Lin et al., arXiv, 2020
39 benchmarks out of 59In the Vistaar benchmark, IndicWhisper, a Whisper model fine-tuned on Indian-language data, had the lowest word error rate in 39 of the 59 benchmarks.Kaushal Santosh Bhogale et al., Interspeech 2023 (arXiv:2305.15386), 2023
15.44 percent word error rateIn the Voxtral paper's FLEURS table, Whisper large-v3 had a word error rate of 15.44 percent in Arabic.Mistral AI, arXiv (Audio and Speech Processing), 2025
28.87 percent word error rateIn the Voxtral paper's FLEURS table, Whisper large-v3 had a word error rate of 28.87 percent in Hindi.Mistral AI, arXiv (Audio and Speech Processing), 2025
In US and Canadian usage, subtitles assume the viewer can hear the audio but cannot understand the language or accent.Wikipedia, n.d.
30 daysIn Voice Memos on iPhone, deleted recordings move to the Recently Deleted folder, where they are kept for 30 days by default.Apple, 2026
In WebVTT, the thousandths of a second in a timestamp follow a full stop character.W3C Timed Text Working Group, W3C, 2026
In Whisper's transcribe function, a segment counts as silent when its no-speech probability is above the no_speech_threshold setting and its average log probability is below logprob_threshold.OpenAI, GitHub (openai/whisper), 2026
In Zoom's automated captions, you open the full transcript from the menu next to the captions icon, under Captions and translation.Zoom Help Center, 2026
5 hours per hour of interviewKvale and Brinkmann estimated that manual transcription takes five hours for each hour of interview, a figure Wollin-Giering and colleagues cite in 2024.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
Limitless says it stopped selling its Pendant recorder to new customers as of December 5, 2025.Limitless, 2025
Limitless says it will support existing Pendant customers throughout 2026, and it does not promise support beyond that year.Limitless, 2025
Many ASR systems discard or generalize hesitations, which loses acoustic detail that verbatim transcription depends on.Jhen-Ke Lin et al., arXiv (SLaTE 2025 workshop), 2025
Meetily's README lists Ollama, which runs locally, as one choice for summaries, alongside cloud providers such as Claude, Groq and OpenRouter, or any OpenAI-compatible endpoint.Zackriya Solutions, GitHub (Zackriya-Solutions/meetily), 2026
Meetily's README says it captures microphone and system audio at the same time, so a meeting played on your computer can be recorded without a bot joining the call.Zackriya Solutions, GitHub (Zackriya-Solutions/meetily), 2026
78 percent of languagesMeta says its 7B-LLM-ASR system reaches character error rates (CER) below 10 for 78 percent of the more than 1,600 languages it covers.Meta AI, Meta AI, 2025
Microsoft says a Teams meeting includes transcripts only if the organizer has the transcription setting turned on, and the user who starts a transcript must also have it on.Microsoft Learn, 2026
Microsoft says live transcription in Microsoft Teams is available only on the Teams desktop app.Microsoft, Microsoft Support, 2026
Microsoft says standard Teams transcription is included in Microsoft 365 E3 and E5, and that the Transcribe only option needs no additional license.Microsoft 365 Message Center, mirrored at mc.merill.net, 2026
Microsoft says Teams live transcription appears in real time and includes each speaker's name and time stamp.Microsoft, Microsoft Support, 2026
Microsoft says Teams live transcripts for meetings with recording or transcription are stored in the meeting organizer's OneDrive for Business.Microsoft, Microsoft Support, 2026
Microsoft says that by default meeting organizers and co-organizers can download a Teams transcript as a .docx or .vtt file.Microsoft, Microsoft Support, 2026
Microsoft says that when live transcription starts in Teams, all participants see a notification that the meeting is being transcribed.Microsoft, Microsoft Support, 2026
Microsoft says the captions generated by Windows 11 live captions are not stored anywhere on the device or in the cloud.Microsoft Support, 2026
Microsoft says the meeting organizer must have a Teams Premium license for attendees to use live translated transcription.Microsoft Learn, 2026
Microsoft says the Teams mobile app supports viewing a transcript after a meeting ends, and not live transcription.Microsoft, Microsoft Support, 2026
Microsoft says the Transcribe only option creates a meeting transcript without storing any recording.Microsoft 365 Message Center, mirrored at mc.merill.net, 2026
Microsoft says the Transcription toggle in a new Teams meeting policy is On by default.Microsoft Learn, 2026
Microsoft says transcription in Teams is both a per-organizer and a per-user policy setting.Microsoft Learn, 2026
250 milliseconds excluded around speaker boundariesMost earlier diarization papers scored a modified DER that leaves out speech within 250 milliseconds of a speaker boundary and all overlapped speech.Neville Ryant et al., arXiv (Interspeech 2019 challenge paper), 2019
Most subtitles distributed on the internet are in the SRT format.Wikipedia, n.d.
7 secondsNetflix's general timed text requirements set a maximum duration of 7 seconds for each subtitle event.Netflix Backlot Help Center, n.d.
5/6 of a secondNetflix's general timed text requirements set a minimum duration of 5/6 of a second for each subtitle event.Netflix Backlot Help Center, n.d.
42 characters per lineNetflix's subtitle template guide sets a limit of 42 characters per line for most Latin-alphabet languages.Netflix Backlot Help Center, n.d.
Northwestern University's guidance advises testing the transcription service with a recording similar to your research data, to check that its accuracy is acceptable for your needs.Northwestern University Information Technology, n.d.
Northwestern's guidance says its Zoom recording and transcription option should not be used with data subject to HIPAA, or with data use agreements that prohibit Zoom or AI services.Northwestern University Information Technology, n.d.
6 audio file formatsNuance's Dragon 15 help for Windows lists .wav, .mp3, .wma, .dss, .ds2 and .m4a as the supported audio file formats for transcribing a recording, with a minimum sampling rate of 16 kHz.Nuance Communications, 2018
7 audio file formatsNuance's Dragon for Mac help says Dragon can transcribe audio files in mp3, mp4, wav, aif, aiff, m4v and m4a formats.Nuance Communications, n.d.
NVivo 14 for Windows says it can import PDF files but cannot edit them in NVivo.QSR International, n.d.
NVivo 15 for Windows says that documents imported from Microsoft Word or text files can be edited inside NVivo, for example by applying paragraph styles, formatting text and inserting images. Interview transcripts are listed as an example of this kind of document.QSR International, NVivo help, 2025
Ofcom's code defines subtitling as text on screen representing speech and sound effects, synchronised as closely as possible to the sound, and suitable for people with hearing loss and d/Deaf people.Ofcom, Ofcom (Office of Communications, United Kingdom), n.d.
Oliver and colleagues argue that interview transcription is a powerful act of representation, not just a behind-the-scenes task.Daniel G. Oliver et al., Social Forces, 84(2), 2005
Oliver and colleagues describe denaturalized transcription as removing idiosyncratic elements of speech, such as stutters, pauses and nonverbals.Daniel G. Oliver et al., Social Forces, 84(2), 2005
On a personal Zoom account you can switch on Create audio transcript yourself, without admin permissions, unless an admin has locked the setting.Zoom Help Center, 2026
On iPhone, a recording made in an earlier version of Voice Memos (iOS 17 or earlier) is transcribed automatically when you open it, if it contains recorded speech.Apple Support, 2026
36.4 percent word error rateOn the AMI meeting corpus recorded with a single distant microphone (SDM1), Whisper Large V2 had a word error rate of 36.4 percent.Alec Radford et al., arXiv (eess.AS), 2022
16.9 percent word error rateOn the AMI meeting corpus recorded with individual headset microphones (IHM), Whisper Large V2 had a word error rate of 16.9 percent.Alec Radford et al., arXiv (eess.AS), 2022
20 percent relative DER reductionOn the AMI meeting corpus, assigning overlapped speech with overlap-aware resegmentation reduced the diarization error rate of the baseline system by 20 percent relative to that baseline.Latané Bullock et al., arXiv (submitted to ICASSP 2020), 2019
100.7 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 100.7 percent for Malayalam.Alec Radford et al., OpenAI (arXiv), 2022
104.1 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 104.1 percent for Bengali.Alec Radford et al., OpenAI (arXiv), 2022
37.0 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 37.0 percent for Kannada.Alec Radford et al., OpenAI (arXiv), 2022
38.3 percent word error rateOn the Fleurs benchmark, the Whisper large-v2 model had a word error rate of 38.3 percent for Marathi.Alec Radford et al., OpenAI (arXiv), 2022
34.7 timesOn the Fleurs benchmark, Whisper large-v2's word error rate for Bengali was about 34.7 times its rate for Spanish.Calculated from other facts
31.6 times the Spanish error rateOn the FLEURS test, Whisper large-v2's word error rate in Yoruba was about 31.6 times its rate in Spanish.Calculated from other facts
70 percent word accuracy (study measure)On the German interview, Whisper reached above 70 percent on the study's word accuracy measure, while Dragon was below 5 percent.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
46.75 percent word error rateOn the Mozilla Common Voice Hindi test in the Voxtral paper, Whisper large-v3 had a word error rate of 46.75 percent.Mistral AI, arXiv (Audio and Speech Processing), 2025
0.7 percentage pointsOn the Whisper Large V3 Turbo experiment, labeling filled pauses lowered the word error rate by 0.7 percentage points.Calculated from other facts
Open, burned-in or hard-coded captions are visible to all viewers because they are embedded in the video.Wikipedia, n.d.
OpenAI advises against splitting a long recording in the middle of a sentence, because the cut can remove context and reduce accuracy.OpenAI, OpenAI, 2026
OpenAI says audio uploads in ChatGPT are available with paid subscriptions and workspaces, including Enterprise.OpenAI Help Center (archived copy), 2026
OpenAI says processing of long audio uploads is best-effort, and that very long recordings may time out.OpenAI Help Center (archived copy), 2026
224 tokens in a Whisper promptOpenAI says Whisper does not follow instructions the way a general-purpose text model does, and that its prompt accepts up to 224 tokens.OpenAI, OpenAI, 2026
OpenAI says Whisper follows the style of its prompt rather than any instructions contained in the prompt.OpenAI, OpenAI Cookbook, n.d.
512 MB per audio fileOpenAI's help article says audio files uploaded to ChatGPT can be up to 512 MB.OpenAI Help Center (archived copy), 2026
$0.0045 per minuteOpenAI's price list shows its gpt-transcribe model at $0.0045 per minute and gpt-4o-transcribe at $0.006 per minute.OpenAI, OpenAI, 2026
OpenAI's speech-to-text guide names gpt-transcribe first and says it is the recommended model for transcribing recorded speech in its original language.OpenAI, OpenAI, 2026
7 file formatsOpenAI's transcription API accepts seven file formats: mp3, mp4, mpeg, mpga, m4a, wav and webm.OpenAI, OpenAI, 2026
25 MBOpenAI's Transcriptions API accepts audio files up to 25 MB, and larger recordings must be compressed or split into chunks.OpenAI, OpenAI, 2026
oTranscribe describes itself as a free web app for transcribing recorded interviews.Elliot Bentley, oTranscribe (a project of the MuckRock Foundation), n.d.
oTranscribe says it works on desktop computers only, so it is not meant for phones or tablets.Elliot Bentley, oTranscribe (a project of the MuckRock Foundation), n.d.
oTranscribe says your audio file and transcript never leave your computer.Elliot Bentley, oTranscribe (a project of the MuckRock Foundation), n.d.
30 minutes per conversationOtter's pricing page caps transcription at 30 minutes for one conversation on its Basic plan, against 90 minutes on its Pro plan.Otter.ai, 2026
300 minutes per monthOtter's pricing page lists a monthly limit of 300 meeting and recording transcription minutes per user on its Basic plan, which has no rollover of unused minutes.Otter.ai, 2026
Pixel Recorder's Transcribe again option gives a transcript in languages that its Live Transcription does not support, according to 9to5Google's December 2023 report.Abner Li, 9to5Google, 2023
Plaud describes the NotePin as ideal for people who have most of their conversations in person.PLAUD, 2026
20 hours of continuous recordingPlaud gives the NotePin 20 hours of continuous recording.PLAUD, 2026
159.00 US dollarsPlaud lists the NotePin at a regular price of $159.00 USD, with a magnetic pin and clip included.PLAUD, 2026
300 minutes per monthPlaud's free Starter Plan, which comes with the NotePin, includes 300 minutes of transcription a month.PLAUD, 2026
Rev defines a verbatim transcript as one that types every word heard, including filler words, false starts, self-corrections and stutters.Rev, Rev, n.d.
Rev describes its non-verbatim style, which removes filler words and stammers, as the most common type of transcript.Rev, Rev, n.d.
Rev describes the transcriptionist role as best for people with strong English grammar and fast, accurate typing, but the page gives no typing speed.Rev, 2026
Rev says freelancer earnings depend on the role chosen, the complexity of projects and how many minutes a freelancer completes.Rev, 2026
Rev says full verbatim transcripts need a lot of extra time and attention to detail, and so cost a little more.Rev, Rev, n.d.
Rev's application asks applicants to complete a short transcription, caption or translation sample to show they meet its quality standards.Rev, 2026
Rev's application for freelancers tests English proficiency, grammar skills and, where it applies, translation ability.Rev, 2026
$0.55 to $2.00+ per audio or video minuteRev's freelancer page lists pay for its legal transcriptionist role at $0.55 to $2.00 or more per audio or video minute.Rev, 2026
$0.40 to $1.10 per audio or video minuteRev's freelancer page lists pay for its transcriptionist role at $0.40 to $1.10 per audio or video minute.Rev, 2026
Rev's freelancer page says it has a waitlist because of high application volume, but applications are still open.Rev, 2026
Rev's legal transcriptionist role asks the transcriptionist to edit a draft to produce a highly accurate transcript.Rev, 2026
$1.99 per minuteRev's price page lists human transcription with a 99 percent plus accuracy guarantee in 12 hours or less at $1.99 per minute.Rev, Rev, 2026
Robert's Rules of Order Newly Revised says minutes should contain mainly a record of what was done at the meeting, not what was said by the members.Michigan State University, Academic Governance, 2026
5.3 minutes per hour of audioScaled to one hour, the 9to5Mac test's 40 seconds for a 7-minute-31-second episode works out to about 5.3 minutes of processing for an hour of audio.Calculated from other facts
72 seconds per hour of audioScaled to one hour, the 9to5Mac test's 9 seconds for a 7-minute-31-second episode works out to about 72 seconds of processing for an hour of audio.Calculated from other facts
SRT timecodes use a comma, not a period, to separate the seconds from the milliseconds.Wikipedia, n.d.
The 2017 Interspeech paper introduces the Montreal Forced Aligner as a new open-source system for aligning speech with text.Michael McAuliffe et al., Proceedings of Interspeech 2017, 2017
The 2024 study's authors found that the more complex the recordings, the more time corrections take.Jan Gorisch and Thomas Schmidt, Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 2024
20.5 times the API file limitThe 512 MB ChatGPT audio upload limit is about 20.5 times the 25 MB file limit of OpenAI's transcription API.Calculated from other facts
The ADA Title II web rule requires state and local government web content and mobile apps to meet WCAG 2.1 Level A and Level AA, unless that would be a fundamental alteration or an undue burden.US Department of Justice, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
100 hours of meeting recordingsThe AMI Meeting Corpus, the benchmark most diarization papers use for meeting audio, consists of 100 hours of meeting recordings.AMI Consortium, AMI Meeting Corpus, University of Edinburgh, n.d.
The authors say that even when each language in a mix is well resourced, the shortage of annotated code-switched audio makes recognition harder.Hexin Liu et al., arXiv, 2025
The authors say that spontaneous switching within a sentence causes language confusion for code-switching recognizers, and that accent bias blurs the sound boundaries between languages.Hexin Liu et al., arXiv, 2025
The authors state that Whisper's training data includes no isiXhosa and no closely related languages, so they chose Swahili as the language token when they fine-tuned on isiXhosa.Christiaan Jacobs et al., arXiv (Stellenbosch University and University of Edinburgh), 2025
The Buzz documentation says the app can import audio and video files and export the transcripts as TXT, SRT or VTT files.Buzz project (Chidi Williams and contributors), 2026
The CESSDA Data Management Expert Guide advises planning anonymisation at the time of transcription or initial write-up.CESSDA, Data Management Expert Guide, n.d.
The CHAT transcription manual says that if you can tell how many words were unintelligible, you may write several xxx strings in a row.Brian MacWhinney, TalkBank, Carnegie Mellon University, n.d.
The CHAT transcription manual says to use the symbol xxx when you cannot hear or understand what the speaker is saying.Brian MacWhinney, TalkBank, Carnegie Mellon University, n.d.
The companion guide Robert's Rules of Order Newly Revised In Brief, as quoted by Michigan State University, says the most frequent mistakes in minutes are trying to summarize the reports offered and the arguments made in debate.Michigan State University, Academic Governance, 2026
April 26, 2028 compliance dateThe compliance date for public entities with fewer than 50,000 people, and for special districts, is April 26, 2028.US Department of Justice, ADA.gov, 2026
The DKU-LENOVO diarization system is built from several separate modules, including voice activity detection, segmentation, speaker embedding extraction, clustering, resegmentation and overlap detection.Qingjian Lin et al., arXiv, 2020
9.06 percentage pointsThe DKU-LENOVO system's DER on DIHARD II Track 2 was 9.06 percentage points higher than on Track 1.Calculated from other facts
0 milliseconds (default collar)The dscore scoring tool computes diarization error rate with a collar of 0 milliseconds by default and counts regions of overlapping speech.Neville Ryant, GitHub, n.d.
The editor Larry Jordan wrote that Premiere Pro does not reliably identify speakers, so he planned to review the transcript by hand.Larry Jordan, Larry Jordan (larryjordan.com), 2024
The FCC captioning rule says captions must match the spoken words in their original language, in the order spoken.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC requires the four quality standards to be met, except for de minimis captioning errors.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule requires captions to have proper spelling, appropriate punctuation and capitalization.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule says captioning must provide non-speech information such as the identity of speakers, music, sound effects and audience reaction.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule says captioning must run from the beginning to the end of the program, to the fullest extent possible.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule says captions must not block other important visual content, including character faces.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule says captions must start when the speech or sounds start and end about when they end.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The FCC rule says offline (prerecorded) captioning must be a complete textual representation of the audio, including speaker identification and non-speech information.US Federal Communications Commission, Code of Federal Regulations (text via Cornell Legal Information Institute), n.d.
The Federal Trade Commission says that if someone asks you to pay to get a job, that is a scam, and honest employers will never ask you to.Federal Trade Commission, FTC Consumer Advice, 2026
The HTML standard defines captions as a transcription or translation of the dialogue that also covers sound effects, relevant musical cues and other relevant audio.WHATWG, HTML Living Standard, n.d.
The HTML standard defines subtitles as a transcription or translation of the dialogue, for when the sound is available but not understood.WHATWG, HTML Living Standard, n.d.
The Illinois Zoom help page says that by default recordings made to a computer go to a Zoom folder inside the Documents folder: C:\Users\[user name]\Documents\Zoom on a PC and /Users/[user name]/Documents/Zoom on a Mac.University of Illinois Urbana-Champaign IT help desk, undated
The language table in Whisper's source code includes Chinese, Korean and Japanese, each with its own language code.OpenAI, GitHub, openai/whisper, 2026
The language table in Whisper's source code includes Shona, Yoruba, Somali and Afrikaans, each with its own language code.OpenAI, GitHub, openai/whisper, 2026
The Library of Congress says linear PCM captures audio without lossy compression, and that its fidelity is enhanced by higher sample rates and sample sizes.Library of Congress, n.d.
The loudnorm filter in ffmpeg performs EBU R128 loudness normalization, with both dynamic and linear normalization modes.FFmpeg project, 2026
The MacWhisper website says it uses local models to transcribe your files, which it presents as suited to sensitive and private material.MacWhisper, MacWhisper (vendor website), 2026
The MacWhisper website says you can transcribe any file by dragging and dropping it into the app.MacWhisper, MacWhisper (vendor website), 2026
MIT licenseThe Meetily repository is released under the MIT License, which allows free use, copying and modification of the code.Zackriya Solutions, GitHub (Zackriya-Solutions/meetily), 2024
The Minnesota Historical Society guide describes one view that a transcript should read as smoothly as possible, with almost cut-and-paste editing.Minnesota Historical Society, Oral History Office, Minnesota Historical Society, 2001
The Minnesota Historical Society guide says each editor edits differently, so text an editor thinks is unimportant may be valuable to a researcher.Minnesota Historical Society, Oral History Office, Minnesota Historical Society, 2001
The Minnesota Historical Society guide says the first draft of an oral history transcript should be a nearly verbatim copy of the interview.Minnesota Historical Society, Oral History Office, Minnesota Historical Society, 2001
The Minnesota Historical Society guide says to delete a false start when the narrator reorganizes their thoughts and resumes.Minnesota Historical Society, Oral History Office, Minnesota Historical Society, 2001
The Minnesota Historical Society guide says to leave a space and type [unclear] when a word or phrase cannot be made out.Minnesota Historical Society, Oral History Office, Minnesota Historical Society, 2001
The Montreal Forced Aligner was evaluated on word and phone boundaries against boundaries that humans had marked by hand.Michael McAuliffe et al., Proceedings of Interspeech 2017, 2017
The openai-whisper command-line tool has a --max_line_count option, which sets the maximum number of lines in a segment. The help text says it requires word timestamps to be turned on.OpenAI, GitHub (openai/whisper), 2026
The openai-whisper command-line tool has a --max_line_width option, which sets the maximum number of characters in a line before the line breaks. The help text says it requires word timestamps to be turned on.OpenAI, GitHub (openai/whisper), 2026
The openai-whisper command-line tool has a --word_timestamps option, which the source marks as experimental. It extracts word-level timestamps and refines the results based on them.OpenAI, GitHub (openai/whisper), 2026
The openai-whisper command-line tool's output format option accepts srt and vtt, along with txt, tsv, json, jsonl and all.OpenAI, GitHub (openai/whisper), 2026
The pyannote.audio paper describes an open-source toolkit written in Python for speaker diarization, with pre-trained models for voice activity detection, speaker change detection, overlapped speech detection and speaker embedding.Hervé Bredin et al., arXiv (submitted to ICASSP 2020), 2019
The second DIHARD challenge used DER as its primary metric, without a collar and without excluding overlapping speech.Neville Ryant et al., arXiv (Interspeech 2019 challenge paper), 2019
38 percentThe study's thematic analysis found that 38 percent of Whisper's hallucinations included explicit harms such as perpetuating violence, making up inaccurate associations or implying false authority.Allison Koenecke et al., arXiv (ACM FAccT 2024), 2024
The supported-language list in Meta's Omnilingual ASR code includes xho_Latn, the code for Xhosa in Latin script.Meta Platforms, Inc., GitHub, facebookresearch/omnilingual-asr, 2025
The supported-language list in Meta's Omnilingual ASR code includes zul_Latn, the code for Zulu in Latin script.Meta Platforms, Inc., GitHub, facebookresearch/omnilingual-asr, 2025
55 percentThe top rate Rev pays a transcriptionist is about 55 percent of the $1.99 per minute it charges customers for human transcription.Calculated from other facts
The University of Utah's guidance says recorded or transcribed content may be subject to public records requests, and lists this as a concern with unapproved tools such as Otter.ai.University of Utah, AI at Utah, 2025
Apache 2.0 licenseThe Vosk speech recognition repository is licensed under the Apache License, version 2.0.Alpha Cephei, GitHub (alphacep/vosk-api), n.d.
The W3C says the SRT file format was used as the basis for the WebVTT text track format.W3C Timed Text Working Group, W3C, 2026
The W3C specification says a WebVTT file consists of a sequence of text segments, each associated with a time interval called a cue.W3C Timed Text Working Group, W3C, 2026
The W3C WebVTT specification says the main use for WebVTT files is captioning or subtitling video content.W3C Timed Text Working Group, W3C, 2026
The W3C WebVTT specification's sample caption file for an interview begins with the word WEBVTT, followed by timed cues.W3C Timed Text Working Group, W3C, 2026
The Whisper command-line tool writes all available output formats when no output format is specified.OpenAI, GitHub (openai/whisper), 2026
The Whisper README says the package also requires the ffmpeg command-line tool to be installed on your system.OpenAI, GitHub (openai/whisper), 2026
The Whisper README says the turbo model is not trained for translation tasks, and that translating non-English speech into English needs one of the multilingual models: tiny, base, small, medium or large.OpenAI, GitHub (openai/whisper), 2026
The Whisper README's example for translating Japanese speech into English uses the medium model, the language flag set to Japanese and the translate task.OpenAI, GitHub (openai/whisper), 2026
The whisper.cpp project shows Whisper running fully offline and on the device on an iPhone 13.Georgi Gerganov, GitHub, 2026
The WhisperX paper adds word-level timestamps to Whisper output, using voice activity detection and forced phoneme alignment.Max Bain et al., arXiv (cs.SD), 2023
BSD 2-Clause licenseThe WhisperX repository is licensed under the BSD 2-Clause License, which permits redistribution and use in source and binary forms with few conditions.Max Bain, GitHub (m-bain/whisperX), 2024
The yt-dlp README lists aac, alac, flac, m4a, mp3, opus, vorbis and wav as audio formats its extraction option can write, with best as the default.yt-dlp contributors, GitHub, yt-dlp project, n.d.
The yt-dlp README says its option that converts video to audio-only files requires ffmpeg and ffprobe.yt-dlp contributors, GitHub, yt-dlp project, n.d.
10 minutesTo apply to TranscriptionStaff.com, you fill out a form and take a 10-minute transcription test, and the company contacts only the applicants who pass.Top Score Ventures, LLC, 2026
95 percentTo pass the practical transcript portion of the CET exam, a candidate must demonstrate 95 percent accuracy.American Association of Electronic Reporters and Transcribers (AAERT), 2026
0.75 US dollars per audio minuteTranscription Hub's homepage offers transcription to customers starting at $0.75 per minute, and it says nothing about what transcriptionists are paid.e24 Technologies, 2026
26.50 US dollars per audio hourTranscriptionStaff.com advertises earnings of up to US$26.50 per audio hour for its transcriptionists, and says the price depends on the difficulty and urgency of the audio.Top Score Ventures, LLC, 2026
2 audio hours per weekTranscriptionStaff.com says a transcriptionist must complete at least 2 audio hours every week to keep the account active.Top Score Ventures, LLC, 2026
3 US dollars minimum paymentTranscriptionStaff.com says it pays transcriptionists weekly by PayPal, with a minimum payment of $3 USD.Top Score Ventures, LLC, 2026
Under 28 U.S.C. 753(b), each session of a federal court must be recorded verbatim, by shorthand, mechanical means, electronic sound recording or another method.Legal Information Institute, Cornell Law School (text of the United States Code), n.d.
20 to 100 milliseconds tolerance (stricter test)Under a stricter tolerance of 20 to 100 milliseconds, the 2025 study's word alignments from Whisper were more accurate than those of earlier work.Sung-Lin Yeh et al., arXiv (ASRU 2025), 2025
Under the federal statute on court reporters, a transcript certified by the reporter or other person designated to produce the record is deemed prima facie a correct statement of the testimony and proceedings.Legal Information Institute, Cornell Law School (text of the United States Code), n.d.
Under the federal statute on court reporters, no transcript of a federal court proceeding counts as official unless it is made from records certified by the reporter or other designated person.Legal Information Institute, Cornell Law School (text of the United States Code), n.d.
Under US copyright law, the first factor in deciding whether a use is fair is its purpose and character, including whether the use is commercial or for nonprofit educational purposes.Legal Information Institute, Cornell Law School, n.d.
US copyright law says that fair use of a work for purposes such as criticism, comment, news reporting, teaching, scholarship or research is not an infringement of copyright.Legal Information Institute, Cornell Law School, n.d.
Way With Words defines verification as checking the transcript against the audio and the style specification.Way With Words, n.d.
Way With Words says higher-risk transcription work typically requires a full second pass by another person.Way With Words, n.d.
Way With Words says verification should focus on predictable failure points, such as speaker attribution, names, numbers, jargon and sensitive content handling.Way With Words, n.d.
WCAG 2.2 success criterion 1.2.2 requires captions for all prerecorded audio content in synchronized media, and it is a Level A criterion.WebAIM, WebAIM, Utah State University, 2026
WCAG 2.2 success criterion 1.2.4 requires captions for all live audio content in synchronized media, and it is a Level AA criterion.WebAIM, WebAIM, Utah State University, 2026
When you transcribe a recording of someone else's voice in Dragon, you can only correct the output by hand, with the mouse and keyboard.Nuance Communications, 2018
Whisper's load_audio function decodes an audio file by starting the ffmpeg command-line tool as a subprocess, and it needs ffmpeg in the PATH to do so.OpenAI, GitHub, openai/whisper repository (main branch), 2022
2.15 times the headset word error rateWith the same model, Whisper Large V2 had about 2.15 times the word error rate on the single distant microphone recording of AMI meetings as on the headset recording.Calculated from other facts
Wollin-Giering and colleagues recommend reviewing the text of any transcript, whatever tool produced it, to ensure accuracy.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
Wollin-Giering and colleagues tested nine automatic transcription tools on two interviews and conclude that Whisper performs best overall.Susanne Wollin-Giering et al., Forum: Qualitative Social Research, 25(1), Art. 8, 2024
YouTube's terms allow you to view or listen to its content for your personal, non-commercial use.YouTube (Google LLC), n.d.
YouTube's terms say you may not download any part of the service or its content unless the service expressly authorizes it.YouTube (Google), n.d.
YouTube's terms say you may not download, reproduce or otherwise use any part of the service or its content, except where the service specifically permits it.YouTube (Google LLC), n.d.
5-20 percentZayats and colleagues report that a quality checking pass improved the average transcription error rate by 5 to 20 percent.Vicky Zayats et al., arXiv, 2019
Zoom converts a computer recording after the meeting has ended, so the files can be opened only once that conversion is done.University of Illinois Urbana-Champaign IT help desk, undated
zoom_0.mp4 file nameZoom names the audio and video file from a computer recording zoom_0.mp4 by default, and later recordings are numbered zoom_1, zoom_2 and so on.University of Illinois Urbana-Champaign IT help desk, undated
audio_only.m4a file nameZoom names the audio-only file from a computer recording audio_only.m4a by default, and later recordings are numbered audio_only_0, audio_only_1 and so on.University of Illinois Urbana-Champaign IT help desk, undated
Zoom says a downloaded audio transcript is a VTT file that you can open in a text editor or a word processing application.Zoom Help Center, 2026
Zoom says a second email tells you when the audio transcript is ready, and its emails link to your recordings or transcript.Zoom Help Center, 2026
Zoom says an account admin can lock the Create audio transcript setting so that no user in the account can change it.Zoom Help Center, 2026
Zoom says audio transcripts of cloud recordings are saved in VTT format.Zoom Help Center, 2026
Zoom says automated captions can run live during a meeting, so you do not have to wait for the transcript that is generated afterwards.Zoom Help Center, 2026
Zoom says computer recording, which saves a meeting to your own device, is available to free and paid accounts.Zoom Help Center, 2026
Zoom says its audio transcription applies only to cloud recordings, not to recordings saved to your computer.Zoom Help Center, 2026
Zoom says selecting Create audio transcript turns on automatic transcription of cloud recordings, and unselecting it turns that transcription off.Zoom Help Center, 2026
Zoom says that after processing, the audio transcript appears as a separate VTT file in the list of recorded meetings.Zoom Help Center, 2026
Zoom says that when speech-to-text is needed for compliance or accessibility, it recommends a manual captioner or service that can guarantee accuracy.Zoom Help Center, 2026
Zoom says the audio transcript may take additional time to process after the cloud recording becomes available.Zoom Help Center, 2026
Zoom says the audio transcript setting requires a Pro, Business, Education or Enterprise account with cloud recording enabled.Zoom Help Center, 2026
Zoom says you cannot enable or disable audio transcription while you are in a meeting or webinar, so the setting must be right before the recording starts.Zoom Help Center, 2026
Zoom's article on enabling audio transcription says the feature supports English only.Zoom Help Center, 2026
19 languagesZoom's article on using audio transcription lists 19 supported languages for cloud recording transcripts, from English to Danish.Zoom Help Center, 2026
Zoom's automated captions article says the full transcript of a session can be downloaded as a TXT file from the transcript panel.Zoom Help Center, 2026
Zoom's whitepaper says that after a session ends, participants cannot retain, save or download captions as a separate post-meeting file.Zoom, 2026
Zoom's whitepaper says that when captions are enabled, Zoom uses automatic speech recognition to turn spoken meeting audio into caption text that participants see during the meeting.Zoom, 2026