{"course":{"slug":"ai-for-small-business","title":"AI for a Small Business","url":"https://addaly.com/learn/ai-for-small-business","level":"beginner","track":"work"},"lesson":{"slug":"meetings-calls-and-notes","title":"Transcription, and the consent question nobody asks","url":"https://addaly.com/learn/ai-for-small-business/meetings-calls-and-notes","position":26,"of":88,"minutes":9,"answer":"Transcription errors concentrate on names, numbers and unusual words, the very content you wanted, so a transcript is a memory aid rather than a record.","keyIdea":"Transcription errors concentrate on names, numbers and unusual words, the very content you wanted, so a transcript is a memory aid rather than a record, and the short written confirmation you send afterwards is the thing that counts."},"body":{"format":"markdown","text":"## Accuracy is not one number\n\nSpeech-to-text is measured by **word error rate**: the proportion of words wrong, inserted or missed. The figure varies enormously with conditions, and the variation matters more than the headline.\n\nClean audio, one speaker, a common accent in a widely-spoken language: error rates in the low single digits. Effectively a good transcript.\n\nNow change one thing at a time. Two people talking over each other. A café in the background. A phone speaker. A strong regional accent. A language with less training data behind it. Technical vocabulary and product names. Each of these degrades the result, and several together can push a transcript from useful to actively misleading.\n\nThe pattern worth remembering: errors are not spread evenly. They concentrate on **names, numbers and unusual words**, which is to say, exactly the content of a business conversation that you needed the transcript for. A summary of a supplier call can be broadly right and have the quantity wrong.\n\n## What to do about it\n\n**Improve the audio before you improve the tool.** A cheap clip-on microphone, or putting the phone closer, does more than switching vendors. Recording each person separately, where the platform allows it, does more still.\n\n**Say names and numbers clearly, and repeat the important ones.** \"Four hundred and twenty units. Four-two-zero.\" This feels stilted and it is the single most effective thing you can do.\n\n**Never take a number from a transcript.** Confirm it in writing afterwards. This one rule prevents most of the damage.\n\n**Read the action items against your memory.** A summary invents plausible next steps when the conversation was vague, and a meeting that ended without a decision often produces a summary containing one.\n\n## The consent question\n\nThis is the part that gets skipped, and it is not a footnote.\n\nRecording law varies enormously by country and sometimes within one. Broadly, some places require only one party to consent: you, if you are on the call. Others require **all** parties. Some distinguish between a private conversation and a business one. Some require notice rather than consent. Penalties in the strict jurisdictions are not trivial, and a recording made unlawfully may also be unusable if you ever needed it.\n\nTwo practical positions:\n\n**Ask, always.** \"I'm recording this so I can get the notes right. Is that alright?\" Nobody has ever objected to that sentence and it settles the question at almost no cost.\n\n**Find out your local rule once.** Twenty minutes of reading, or one question to an adviser. If you take calls from customers in other countries, the stricter rule is the safer one to apply to everybody.\n\nNote that many meeting tools now join as a visible participant and announce recording. That is a good design and it is not automatically the same as consent. It is notice; whether notice is enough depends on where you are.\n\n## Free paths\n\n- **Whisper**, released openly by OpenAI, runs on your own computer. Excellent accuracy for a free tool, no upload, nothing leaves the machine. Front-ends such as `whisper.cpp` and Buzz make it usable without programming.\n- **Live captions** built into Android, iOS, Windows and Google Meet. Free, rough, and often enough.\n- **The meeting tool you already pay for.** Zoom, Teams and Meet all have recording and summaries, usually included above the free tier.\n\n## What a transcript is for\n\nNot a record. A transcript is a search aid and a memory prompt.\n\nThe record is the short message you send afterwards: *as discussed, 420 units, delivery 14 March, price held at last year's rate.* That message is what both parties agreed to, it is what you will rely on in a dispute, and writing it is exactly the sort of small task worth drafting from a transcript and then reading properly before you send.\n\n## Speaker labels and who said what\n\nA second, quieter failure: **diarisation**, the business of deciding which speaker said which sentence. It is a separate problem from recognising the words, and it is harder.\n\nTwo people with similar voices, an interruption, a speakerphone in the middle of a table, and lines get attributed to the wrong person. The transcript then says the supplier offered the discount when in fact you asked for it.\n\nThis matters because attribution is often the point. If you are transcribing a negotiation, a complaint or a disciplinary conversation, check who said what before relying on any of it. Recording each participant on their own channel, which several platforms support, removes the problem almost entirely.\n\n## A use worth more than the meeting notes\n\nThe strongest use of transcription in a small business is not meetings. It is talking to yourself.\n\nNinety seconds into your phone at the end of a job (what happened, what the customer said, what to order) becomes a job note, a marketing post and a reminder. You would not have typed it. Speaking is roughly three times faster than typing for most people, and this closes the gap between what you noticed and what got written down, which is where most small-business knowledge quietly evaporates."},"check":{"question":"A distributor relies on AI meeting summaries for supplier calls. A summary records an agreed unit price of 12.50 when the supplier said 12.15. What does this failure most directly illustrate?","options":[{"text":"Summarisation compresses detail, so precise figures are frequently lost when a long conversation is reduced to a few lines of text.","correct":false,"tempting":"Blames compression, but the figure was not lost; it was present and wrong, which is a transcription error carried faithfully into the summary."},{"text":"The model hallucinated a plausible round number because it had no strong signal for the actual value in the conversation.","correct":false,"tempting":"Reaches for fabrication when the simpler explanation is a misheard digit, and it would not predict that the same error concentrates on names and quantities."},{"text":"Speech recognition errors cluster on numbers and names, so the transcript was wrong before any summarising happened, and no summary can recover what was misheard.","correct":true},{"text":"The audio quality was inadequate for business use, and the distributor should have been using a dedicated recording device with separate channels for each speaker rather than relying on a conference platform's built-in capture.","correct":false,"tempting":"Prescribes an equipment upgrade that helps at the margin, when even good audio produces this error class and the fix is confirming numbers in writing."}]},"prev":{"slug":"contracts-and-long-documents","title":"Reading a contract you were going to skim anyway","url":"https://addaly.com/learn/ai-for-small-business/contracts-and-long-documents"},"next":{"slug":"keeping-the-knowledge-current","title":"The stale document problem","url":"https://addaly.com/learn/ai-for-small-business/keeping-the-knowledge-current"},"authorship":"Founder and series editor: Dr Mustafa Mun. Written by AI models to a written standard. Exam keys checked blind by a second model: 3,665 of 3,666 agreed. Academic review invited.","review":{"status":"Draft, open for academic review"},"cite":"Addaly (2026) 'Transcription, and the consent question nobody asks', in AI for a Small Business. Edited by M. Mun. Addaly. Available at: https://addaly.com/learn/ai-for-small-business/meetings-calls-and-notes (Accessed: 13 October 2026).","licence":{"name":"CC BY-NC-SA 4.0","url":"https://creativecommons.org/licenses/by-nc-sa/4.0/"},"reader":{"schema":1,"title":"Transcription, and the consent question nobody asks","courseTitle":"AI for a Small Business","module":{"index":3,"slug":"your-own-knowledge","title":"Your own documents, and getting straight answers out of them"},"answer":null,"keyIdea":"Transcription errors concentrate on names, numbers and unusual words, the very content you wanted, so a transcript is a memory aid rather than a record, and the short written confirmation you send afterwards is the thing that counts.","hero":{"kind":"plate","image":null,"alt":null,"caption":null},"blocks":[{"t":"h2","id":"accuracy-is-not-one-number","runs":[{"text":"Accuracy is not one number"}]},{"t":"p","runs":[{"text":"Speech-to-text is measured by "},{"text":"word error rate","b":true},{"text":": the proportion of words wrong, inserted or missed. The figure varies enormously with conditions, and the variation matters more than the headline."}]},{"t":"p","runs":[{"text":"Clean audio, one speaker, a common accent in a widely-spoken language: error rates in the low single digits. Effectively a good transcript."}]},{"t":"p","runs":[{"text":"Now change one thing at a time. Two people talking over each other. A café in the background. A phone speaker. A strong regional accent. A language with less training data behind it. Technical vocabulary and product names. Each of these degrades the result, and several together can push a transcript from useful to actively misleading."}]},{"t":"p","runs":[{"text":"The pattern worth remembering: errors are not spread evenly. They concentrate on "},{"text":"names, numbers and unusual words","b":true},{"text":", which is to say, exactly the content of a business conversation that you needed the transcript for. A summary of a supplier call can be broadly right and have the quantity wrong."}]},{"t":"h2","id":"what-to-do-about-it","runs":[{"text":"What to do about it"}]},{"t":"p","runs":[{"text":"Improve the audio before you improve the tool.","b":true},{"text":" A cheap clip-on microphone, or putting the phone closer, does more than switching vendors. Recording each person separately, where the platform allows it, does more still."}]},{"t":"p","runs":[{"text":"Say names and numbers clearly, and repeat the important ones.","b":true},{"text":" “Four hundred and twenty units. Four-two-zero.” This feels stilted and it is the single most effective thing you can do."}]},{"t":"p","runs":[{"text":"Never take a number from a transcript.","b":true},{"text":" Confirm it in writing afterwards. This one rule prevents most of the damage."}]},{"t":"p","runs":[{"text":"Read the action items against your memory.","b":true},{"text":" A summary invents plausible next steps when the conversation was vague, and a meeting that ended without a decision often produces a summary containing one."}]},{"t":"h2","id":"the-consent-question","runs":[{"text":"The consent question"}]},{"t":"p","runs":[{"text":"This is the part that gets skipped, and it is not a footnote."}]},{"t":"p","runs":[{"text":"Recording law varies enormously by country and sometimes within one. Broadly, some places require only one party to consent: you, if you are on the call. Others require "},{"text":"all","b":true},{"text":" parties. Some distinguish between a private conversation and a business one. Some require notice rather than consent. Penalties in the strict jurisdictions are not trivial, and a recording made unlawfully may also be unusable if you ever needed it."}]},{"t":"p","runs":[{"text":"Two practical positions:"}]},{"t":"p","runs":[{"text":"Ask, always.","b":true},{"text":" “I’m recording this so I can get the notes right. Is that alright?” Nobody has ever objected to that sentence and it settles the question at almost no cost."}]},{"t":"p","runs":[{"text":"Find out your local rule once.","b":true},{"text":" Twenty minutes of reading, or one question to an adviser. If you take calls from customers in other countries, the stricter rule is the safer one to apply to everybody."}]},{"t":"p","runs":[{"text":"Note that many meeting tools now join as a visible participant and announce recording. That is a good design and it is not automatically the same as consent. It is notice; whether notice is enough depends on where you are."}]},{"t":"h2","id":"free-paths","runs":[{"text":"Free paths"}]},{"t":"ul","items":[[{"text":"Whisper","b":true},{"text":", released openly by OpenAI, runs on your own computer. Excellent accuracy for a free tool, no upload, nothing leaves the machine. Front-ends such as "},{"text":"whisper.cpp","code":true},{"text":" and Buzz make it usable without programming."}],[{"text":"Live captions","b":true},{"text":" built into Android, iOS, Windows and Google Meet. Free, rough, and often enough."}],[{"text":"The meeting tool you already pay for.","b":true},{"text":" Zoom, Teams and Meet all have recording and summaries, usually included above the free tier."}]]},{"t":"h2","id":"what-a-transcript-is-for","runs":[{"text":"What a transcript is for"}]},{"t":"p","runs":[{"text":"Not a record. A transcript is a search aid and a memory prompt."}]},{"t":"p","runs":[{"text":"The record is the short message you send afterwards: "},{"text":"as discussed, 420 units, delivery 14 March, price held at last year’s rate.","i":true},{"text":" That message is what both parties agreed to, it is what you will rely on in a dispute, and writing it is exactly the sort of small task worth drafting from a transcript and then reading properly before you send."}]},{"t":"h2","id":"speaker-labels-and-who-said-what","runs":[{"text":"Speaker labels and who said what"}]},{"t":"p","runs":[{"text":"A second, quieter failure: "},{"text":"diarisation","b":true},{"text":", the business of deciding which speaker said which sentence. It is a separate problem from recognising the words, and it is harder."}]},{"t":"p","runs":[{"text":"Two people with similar voices, an interruption, a speakerphone in the middle of a table, and lines get attributed to the wrong person. The transcript then says the supplier offered the discount when in fact you asked for it."}]},{"t":"p","runs":[{"text":"This matters because attribution is often the point. If you are transcribing a negotiation, a complaint or a disciplinary conversation, check who said what before relying on any of it. Recording each participant on their own channel, which several platforms support, removes the problem almost entirely."}]},{"t":"h2","id":"a-use-worth-more-than-the-meeting-notes","runs":[{"text":"A use worth more than the meeting notes"}]},{"t":"p","runs":[{"text":"The strongest use of transcription in a small business is not meetings. It is talking to yourself."}]},{"t":"p","runs":[{"text":"Ninety seconds into your phone at the end of a job (what happened, what the customer said, what to order) becomes a job note, a marketing post and a reminder. You would not have typed it. Speaking is roughly three times faster than typing for most people, and this closes the gap between what you noticed and what got written down, which is where most small-business knowledge quietly evaporates."}]}],"check":{"question":[{"text":"A distributor relies on AI meeting summaries for supplier calls. A summary records an agreed unit price of 12.50 when the supplier said 12.15. What does this failure most directly illustrate?"}],"options":[{"text":[{"text":"Summarisation compresses detail, so precise figures are frequently lost when a long conversation is reduced to a few lines of text."}],"correct":false,"tempting":[{"text":"Blames compression, but the figure was not lost; it was present and wrong, which is a transcription error carried faithfully into the summary."}]},{"text":[{"text":"The model hallucinated a plausible round number because it had no strong signal for the actual value in the conversation."}],"correct":false,"tempting":[{"text":"Reaches for fabrication when the simpler explanation is a misheard digit, and it would not predict that the same error concentrates on names and quantities."}]},{"text":[{"text":"Speech recognition errors cluster on numbers and names, so the transcript was wrong before any summarising happened, and no summary can recover what was misheard."}],"correct":true,"tempting":null},{"text":[{"text":"The audio quality was inadequate for business use, and the distributor should have been using a dedicated recording device with separate channels for each speaker rather than relying on a conference platform’s built-in capture."}],"correct":false,"tempting":[{"text":"Prescribes an equipment upgrade that helps at the margin, when even good audio produces this error class and the fix is confirming numbers in writing."}]}],"order":[3,1,2,0],"why":null},"prev":{"slug":"contracts-and-long-documents","title":"Reading a contract you were going to skim anyway","minutes":10},"next":{"slug":"keeping-the-knowledge-current","title":"The stale document problem","minutes":8}}}