Live transcription of a business call is accurate enough to follow along and flag action items. It makes noticeably more mistakes than a transcript produced after the call from the full recording, especially with names, numbers and people talking over each other. For confidential calls, the most accurate private method is to record with everyone’s knowledge, then transcribe on your own device with a large speech model once the call ends. Treat live captions as a guide, not the record. The audio path (phone line, video app or speakerphone) usually matters more to accuracy than whether the model runs locally or in the cloud.
What limits transcription accuracy on business calls? #
- Phone-line audio. Traditional phone calls carry a narrow band of frequencies, roughly 300 to 3,400 Hz. That cuts off much of the high-frequency detail in consonants like “s” and “f,” so “fifty” and “fifteen” start to sound alike.
- Compression and dropouts. Video and VoIP calls compress audio and occasionally drop fragments when the connection wobbles.
- Speakerphone and echo. A room mic picks up reflections and background noise, and echo cancellation can clip the start of words.
- Crosstalk. Two people talking at once is the single biggest source of errors for every engine.
- Names, numbers and jargon. Company names, product codes and figures are rare in training data and costly to get wrong.
- Streaming itself. Live engines guess from a few seconds of audio. Whisper-family models were trained on 30-second chunks, according to OpenAI’s paper, so they do better with more context.
Real-time vs after-the-call transcription #
| Live, during the call | After the call, from the recording | |
|---|---|---|
| What you get | Captions that may change as they appear | A finished transcript with timestamps |
| Context the model sees | A few seconds at a time | The whole call |
| Model size | Smaller, to keep up | As large as your device handles |
| Names and numbers | Most error-prone | Better, especially with custom vocabulary |
| Best use | Following along, noting action items | Minutes, quotes, records |
A common approach is to use live captions if they help you keep up, but base anything you’ll send, quote or file on the after-the-call transcript.
How to measure accuracy on your own calls #
Vendor accuracy claims are measured on their test audio, not yours. A quick check:
- Take a 3 to 5 minute stretch of a typical call (with permission).
- Type an exact reference transcript.
- Compare each tool’s output and count substituted, missed and added words.
- Divide by the number of words in your reference. That’s the word error rate (WER).
Pay special attention to the errors that matter: names, amounts, dates and “not.” A tool with a slightly higher WER that gets numbers right may be the better choice.
Private ways to transcribe business calls #
Phone calls on iPhone #
On iPhones with Apple Intelligence, you can record a call in the Phone app. Apple says participants are automatically notified when recording starts, and that the Notes and Phone apps can record, transcribe and summarize audio. Apple lists call transcription for English and a handful of other languages on its feature availability page. Check Apple’s documentation for how processing is handled on your device before using it for regulated calls.
Some Android phones include call recording in the dialer, depending on the manufacturer and country.
Video calls on Zoom, Teams or Meet #
Record the call to your computer, then transcribe the file locally with a Whisper app such as MacWhisper, Buzz or whisper.cpp. The steps, including how to get a separate audio track per participant, are in our guide to private Zoom meeting transcription.
In-person meetings and calls on speaker #
Private Transcribe records on an iPhone or Android phone and runs Whisper on the phone itself, so the conversation is never uploaded. There’s no account, and it works in airplane mode after a one-time model download. For business calls, the useful parts are:
- Professional or Best models (539 MB and 574 MB) for important conversations, trading a little speed for accuracy.
- Custom vocabulary for client names, product names and acronyms.
- Import of recordings you already have as audio or video files (up to 90 minutes).
- Timestamps on every line, so you can check a figure against the audio in seconds.
It doesn’t transcribe live and doesn’t label speakers. Record, stop, and read the transcript a moment later.
How to get more accurate call transcripts #
- Use a headset rather than speakerphone for your side of the call.
- Agree on turn-taking at the start. It helps the transcript and the meeting.
- Spell out names the first time (“that’s Nguyen, N-G-U-Y-E-N”) and add them to your vocabulary list.
- Say key numbers twice, differently (“fifteen, one-five”) so a mishearing is obvious.
- Choose the largest model your device handles for anything you’ll rely on.
- Proofread amounts, dates and commitments against the recording before sending minutes.
Our comparison of private transcription accuracy vs online services covers when a cloud service is still worth it.
Consent and compliance #
Recording a call has legal requirements that come before any accuracy question. Consent rules differ between countries and between US states, and some require every participant to agree. Regulated industries, such as financial services and healthcare, may also have recordkeeping or privacy rules about how call recordings and transcripts are stored.
Keep it simple: tell everyone at the start that you’re recording, why, and that it stays on your device. For legal practice specifically, see best offline transcription tools for lawyers.
Keeping the transcript confidential #
A private transcription tool is only part of it:
- Delete the audio once the transcript is checked, unless you’re required to keep it.
- Store the text where your organization keeps other confidential documents, not in a personal cloud folder.
- Share excerpts, not full transcripts, when only a decision or action item is needed.
The full workflow is in how to transcribe confidential meetings without cloud storage.
Frequently asked questions #
How accurate is real-time transcription on business calls? #
There’s no single number, because it depends on the line quality, the speakers and the engine. On a clear headset call in a major language, live captions capture most of what’s said. Phone-line audio, speakerphones and crosstalk cause noticeably more errors. Test on a few minutes of your own calls to know for sure.
Is local transcription as accurate as cloud transcription for calls? #
With a large model, local transcription is close to automated cloud services on clear audio. Cloud tools are usually better at naming speakers automatically. On poor phone-line audio, every engine struggles, so recording quality matters more than where the model runs.
Can I transcribe phone calls on my iPhone? #
Yes, on iPhones that support Apple Intelligence. The Phone app can record a call, and Apple says participants are automatically notified. The recording can be transcribed and summarized. Always make sure recording is lawful where you and the other parties are.
How do I stop transcription errors in numbers and names? #
Add names and terms to the tool’s custom vocabulary, spell out unusual names on the call, and say important numbers in two forms. Then check every figure in the transcript against the recording before you act on it.
Do I need consent to transcribe a business call? #
You need whatever consent the law requires for recording the call, which varies by location, and some places require everyone’s consent. Transcribing a lawful recording afterwards on your own device doesn’t add a new disclosure, but your company’s policies may still apply.