# Transcribe a recording, then check it before you quote it Get a fast AI transcript of an interview, lecture, memo or call, then listen to the parts that matter before you quote, search or summarise it. Workflows · v1.0 · draft · https://hayzlabs.com/workflows/transcribe-then-check/ In short: - Get everyone's consent, record in a quiet room, and transcribe with timestamps using a tool you already have. - Spot-check random stretches by listening; if any has a phrase nobody said, treat the whole transcript as a draft. - Listen in full to every quote, number, name and date you will use before you summarise or quote it. **Use it when** you have a recording of an interview, lecture, voice memo or call, and you want to quote it, search it or summarise it. **Not when** anyone in the recording did not agree to it, or the recording is medical, legal or about someone's job. Keep those out of AI tools unless your organisation has approved one for that use. **You need** the recording, the consent of the people in it, a transcription tool you already have, headphones, and a text file for the check sheet below. Remove what the tool does not need before you upload anything: [What not to paste](https://hayzlabs.com/start/what-not-to-paste/). The steps, times and thresholds are our suggestion. > [!NOT ADVICE] > Recording and consent rules differ by place and by setting, and we do not give legal advice. For health, legal or employment recordings, follow your organisation's rules or ask someone qualified. ## Why the check matters - In one study of OpenAI's Whisper as it ran in 2023, about 1% of transcriptions contained whole phrases or sentences that were never spoken, and 38% of those invented passages included explicit harms such as violence, false associations or false authority ([Koenecke et al., 2024](https://arxiv.org/abs/2402.08021), benchmark, peer-reviewed, FAccT 2024). The study used 13,140 audio segments from 437 speakers. - The inventions were more common for speakers with aphasia, a language disorder often caused by a stroke (1.7% of segments against 1.2% for the control group), and audio with a larger share of silence was more likely to get invented text. The authors found the rate fell after a late-2023 update but that the model still "regularly and reproducibly hallucinates". - OpenAI's own Whisper model card (its published fact sheet on the model) says its output may include "texts that are not actually spoken in the audio input", that it is prone to repeating text, that accuracy is uneven across languages, accents and speaker groups, and it recommends "against use in high-risk domains like decision-making contexts" ([model card](https://github.com/openai/whisper/blob/main/model-card.md), vendor). A word error is easy to spot. An invented sentence reads smoothly. That is why the check below is about listening, not reading. ## Steps | # | Step | Who | Time | Check | |---|---|---|---|---| | 1 | Get consent from everyone who will be recorded, before you press record. Say what you will do with the transcript. | Person | 1 min | Each person agreed, and you noted when. **If it fails:** do not record. Take notes by hand. | | 2 | Record well: a quiet room, one microphone per speaker if you can, the microphone near whoever speaks, and no talking over each other. | Person | 2 min setup | A 10-second test sounds clear on headphones. **If it fails:** move closer or to a quieter room. | | 3 | Transcribe with a tool you have. Set the language to the one spoken. Keep the timestamps. | Tool | Varies | You have a text file with timestamps and, where the tool supports it, speaker labels. **If it fails:** split long files or try another tool from the list below. | | 4 | Spot-check: pick three random 1-minute stretches, listen to each against the text, and count the errors in the sheet. | Person | 5 min | Three rows filled. **If it fails:** see the rule below. | | 5 | Listen in full to every passage you will quote, every number, name and date, and every silence, music or crosstalk spot. | Person | Length of those passages | Each one is marked as checked or corrected. **If it fails:** correct the text from the audio. If you cannot hear it, write `[inaudible]`. | | 6 | Mark each checked quote with its timestamp and the word "checked". | Person | 1 min per quote | Every quote you will use has a timestamp. | | 7 | Only now summarise, search or extract. For meetings, use [Meeting notes to decisions](https://hayzlabs.com/workflows/meeting-notes-to-decisions/). | Person and AI | 2 min | Anything you quote from the summary traces back to a checked line. | > [!OUR RULE] > If any spot-check stretch contains a phrase that nobody said, treat the whole transcript as a draft. Use only lines you have listened to. If the three stretches have more than five word errors in total, listen in full to every section you plan to rely on, not only the quotes. ## Tools you may already have Read on each vendor's page on 2026-10-11: - **Zoom** cloud recordings can produce an audio transcript on Pro, Business, Education or Enterprise accounts, saved as a VTT file (a plain-text subtitle format) with a timestamp on each section. Zoom's tips: keep background noise low, place the microphone near active speakers, and prefer an external microphone ([Zoom](https://support.zoom.com/hc/en/article?id=zm_kb&sysparm_article=KB0064927)). - **Microsoft Teams** transcripts show speaker names and timestamps, all participants are notified when transcription starts, and organisers can download a .docx or .vtt file. Microsoft says to "make sure the meeting spoken language reflects the language everyone is speaking" ([Microsoft](https://support.microsoft.com/en-us/office/view-live-transcription-in-microsoft-teams-meetings-dc1a8f23-2e20-4684-885e-2152e06a4a8b)). - **Pixel Recorder** on Google Pixel phones transcribes recordings, and you can edit the text. Google says re-transcribing "may process audio files on Google servers" ([Google](https://support.google.com/pixelphone/answer/16267698?hl=en)). - **Voice Memos** on iPhone can show a transcription of a recording ([Apple](https://support.apple.com/guide/iphone/view-a-transcription-iph00953a982/ios)). - **Whisper** is open source (its code is public and free to use) under the MIT License and installs from a terminal (the window where you type commands) with `pip install -U openai-whisper` plus `ffmpeg` ([GitHub](https://github.com/openai/whisper)). It installs on your own computer. The model sizes range from `tiny` (about 1 GB of video memory, the memory on a graphics card) to `large` (about 10 GB). ## Check sheet Copy this into a text file next to the recording and fill it as you go: ```text title="Transcript check sheet · v1.0" Recording: Date recorded: Consent: Tool and language setting: SPOT-CHECK (3 random 1-minute stretches) Start time | Word errors | Wrong speaker | Invented phrase (Y/N) | | | | | | | | | Total word errors: Any invented phrase: Y / N LISTENED IN FULL Timestamp | What (quote / number / name / date / silence / crosstalk) | Checked or corrected | Note | | | Ready to summarise only after every quote above says checked or corrected. ``` Replace the following: - ``: the file name, such as `interview-03.m4a`. - ``: the day it was recorded. - ``: names and how they agreed, such as `both, aloud at 00:00`. - ``: the tool and the language you set, such as `Zoom, English`. - `` to ``: three start times you picked at random, such as `07:00`. - ``, ``, ``, ``: one row per passage you listened to in full. > [!CHECK] > Before you use the transcript, read your quotes aloud while the audio plays. The words should match exactly, and the speaker should be the person you name. Pay most attention after a pause, where an extra sentence can appear. ## How to tell it worked - Every quote you publish or send has a timestamp and the word "checked" in the sheet. - Every number, name and date you use matches the audio, not only the text. - The sheet shows a spot-check result, so you know how much to trust the lines you did not hear. - It has failed if you quoted a line you only read. Go back and listen to it. > [!LIMIT] > Three random minutes cannot prove that the rest is right. They only tell you how careful to be. Invented text can appear anywhere, and the evidence on it comes from one study of one model as it ran in 2023; newer tools may do better or worse, and we have not measured them. Speaker labels can be wrong even when the words are right. Uploading a recording sends other people's voices to a vendor, so check where it goes first. Some meeting tools also train on what you upload: Otter's privacy policy (effective 16 June 2026) lists "training our proprietary AI technology on de-identified audio recordings" and on transcriptions, which it says may contain personal information ([Otter privacy policy](https://otter.ai/privacy-policy), vendor policy). ## Sources 1. Koenecke et al., 2024. "Careless Whisper: Speech-to-Text Hallucination Harms". 13,140 audio segments from 437 speakers, FAccT 2024. Peer-reviewed benchmark. [Link](https://arxiv.org/abs/2402.08021). Summary: [Cornell Chronicle](https://www.news.cornell.edu/stories/2024/06/ai-speech-text-can-hallucinate-violent-language). 2. OpenAI. Whisper model card and repository. Vendor. [Model card](https://github.com/openai/whisper/blob/main/model-card.md), [repository](https://github.com/openai/whisper). 3. Zoom. Audio transcripts for cloud recordings. Vendor help page. [Link](https://support.zoom.com/hc/en/article?id=zm_kb&sysparm_article=KB0064927). 4. Microsoft. Live transcription in Teams meetings. Vendor help page. [Link](https://support.microsoft.com/en-us/office/view-live-transcription-in-microsoft-teams-meetings-dc1a8f23-2e20-4684-885e-2152e06a4a8b). 5. Google. Pixel Recorder transcription. Vendor help page. [Link](https://support.google.com/pixelphone/answer/16267698?hl=en). 6. Apple. View a Voice Memos transcription on iPhone. Vendor help page. [Link](https://support.apple.com/guide/iphone/view-a-transcription-iph00953a982/ios). 7. Otter.ai. Privacy policy, effective 16 June 2026. Vendor policy, read 2026-10-19. [Link](https://otter.ai/privacy-policy)