Tips
Which transcription engine is most accurate for meetings?
We ran the same fictional meeting through Apple, WhisperKit and OpenAI engines three ways: as an audio file, recorded on iPhone, and recorded on a Mac’s built-in mic. Here are the word error rates.
These are measurements made by the Otonota team, inside Otonota. Results vary with app and engine versions, recording equipment, distance and room, and with how people speak and what they talk about. Treat them as a reference, not a guarantee.
The short answer
- With clear audio, Apple Speech Recognition (Apple Intelligence) was the most accurate engine we tested, at a word error rate of 4.1%. It is the default on iPhone and iPad, and it is also the fastest.
- iPhone and iPad — when the speaker was away from the phone, the default “AI (Apple Intelligence)” stayed around 9.1%. WhisperKit large-v3 quantized (Pro) did better at 5.8%. “Standard” fell behind at 14.2%.
- Mac — on the built-in microphone, Apple Speech Recognition (Apple Intelligence) and the free WhisperKit large-v3 quantized were both around 9%. The default legacy Apple Speech recognizer was erratic, and OpenAI (Pro) was less accurate than both.
Three ways of capturing the same meeting
We wrote a fictional weekly project meeting (about 180 words, 70 seconds), turned it into speech, and transcribed it three ways. The script contains 14 names, dates and amounts that we checked one by one.
- Audio file — the synthesized file imported as is. No noise, no echo: the easiest possible case
- iPhone mic — played through speakers and recorded by an iPhone placed away from them, like a phone left on the table during a meeting
- Mac built-in mic — the same playback recorded by a MacBook Pro's built-in microphone
Each recording was made twice and the error rates averaged. Lower is better. Results are split into the engines you can choose on iPhone and iPad and the ones you can choose on the Mac, using each app's names. “AI (Apple Intelligence)” on iPhone is the same recognizer as “Apple Speech Recognition (Apple Intelligence)” on the Mac, and “Standard” is the same as “Apple Speech”.
iPhone and iPad
| Engine | Edition | Audio file | iPhone mic |
|---|---|---|---|
| AI (Apple Intelligence) | Free, default | 4.1%(best) | 9.1% |
| Standard | Free | 14.0% | 14.2% |
| WhisperKit large-v3 (quantized) | Pro | 5.8% | 5.8%(best) |
| WhisperKit small | Pro | 8.1% | 9.0% |
Word error rate (WER) = substituted, missing and extra words divided by the number of words in the script. Case and punctuation are ignored; “75 percent” and “75%” count as the same. “iPhone mic” is the iPhone's recording processed by the same engine on a Mac.
With the speaker away from the phone, the default “AI (Apple Intelligence)” scored 9.1% and WhisperKit large-v3 (quantized) 5.8% (figure 1). Transcribing live in the iPhone app with “AI (Apple Intelligence)” gave 8.1%, close to processing the recording afterwards. “Standard” trailed at 14.2%.
The highlighted bar is the default engine.
Show the numbers
| Item | Value |
|---|---|
| WhisperKit large-v3 (quantized) | 5.8% |
| WhisperKit small | 9.0% |
| AI (Apple Intelligence) | 9.1% |
| Standard | 14.2% |
Mac
For the Mac we also include iPhone recordings imported into the Mac app.
| Engine | Edition | Audio file | Mac built-in mic | iPhone recording imported |
|---|---|---|---|---|
| Apple Speech (on-device) | Free, default | 14.0% | 34.6% | 14.2% |
| Apple Speech Recognition (Apple Intelligence) | Free | 4.1%(best) | 9.3% | 9.1% |
| WhisperKit large-v3 quantized | Free | 5.8% | 9.0%(best) | 5.8%(best) |
| WhisperKit large-v3-turbo | Pro | 5.2% | 9.0%(best) | 7.5% |
| OpenAI gpt-4o-transcribe | Pro, own API key | 9.3% | 11.0% | 13.7% |
| OpenAI gpt-4o-mini-transcribe | Pro, own API key | 7.6% | 18.3% | 9.8% |
OpenAI engines are part of Pro on the Mac and use your own API key; the audio is sent to OpenAI.
On the built-in microphone, Apple Speech Recognition (Apple Intelligence) and the WhisperKit models all landed around 9%, and the cloud engines did worse (figure 2). The default legacy Apple Speech recognizer was the outlier: its two takes came out at 51.2% and 18.0% — the same setup, very different results.
The highlighted bar is the default engine.
Show the numbers
| Item | Value |
|---|---|
| WhisperKit large-v3 quantized | 9.0% |
| WhisperKit large-v3-turbo | 9.0% |
| Apple Speech Recognition (Apple Intelligence) | 9.3% |
| OpenAI gpt-4o-transcribe | 11.0% |
| OpenAI gpt-4o-mini-transcribe | 18.3% |
| Apple Speech (on-device) | 34.6% |
The legacy Apple Speech recognizer is unpredictable at a distance
Apple Speech (the Mac default, “Standard” on iPhone) scored 14.0% on the audio file but fell behind as soon as the voice came from a distance. In our Japanese test the same recognizer missed almost everything. For meetings, choose Apple Speech Recognition (Apple Intelligence) or WhisperKit.
Names are where every engine slips
Even accurate engines misspell unfamiliar names. Tables 3 and 4 count how many of the 14 checked terms (10 names of people, companies and projects, plus two dates, an amount and a percentage) came out right.
| Engine | Audio file | iPhone mic |
|---|---|---|
| AI (Apple Intelligence) | 10.0(best) | 10.0(best) |
| Standard | 7.0 | 8.0 |
| WhisperKit large-v3 (quantized) | 9.0 | 9.0 |
| WhisperKit small | 7.0 | 9.0 |
The highlighted row is the default engine.
| Engine | Audio file | Mac built-in mic | iPhone recording imported |
|---|---|---|---|
| Apple Speech (on-device) | 7.0 | 6.5 | 8.0 |
| Apple Speech Recognition (Apple Intelligence) | 10.0 | 9.0 | 10.0(best) |
| WhisperKit large-v3 quantized | 9.0 | 9.0 | 9.0 |
| WhisperKit large-v3-turbo | 10.0 | 8.0 | 8.0 |
| OpenAI gpt-4o-transcribe | 11.0 | 9.5(best) | 8.5 |
| OpenAI gpt-4o-mini-transcribe | 12.0(best) | 8.5 | 8.5 |
The highlighted row is the default engine. Recorded columns are the average of two takes.
No engine got all 14, even from the clean file: names like Okonkwo and Nguyen are hard for every model. A glossary helps more than switching engines; we measured how much inGetting names right in meeting transcripts.
Speed
Figure 3 plots processing time against accuracy. The horizontal axis is how long one hour of audio takes to process, as a fraction of an hour; the lower left is fast and accurate.
Measured on a MacBook Pro (M4 Max) processing the audio file. Apple Speech and the cloud engines receive audio in real time, so they are left out.
Show the numbers
| Item | Processing time (fraction of real time, on a Mac) | Word error rate (%, iPhone mic) |
|---|---|---|
| Apple Speech Recognition (Apple Intelligence) | 0.01 | 9.1 |
| WhisperKit large-v3 quantized | 0.12 | 5.8 |
| WhisperKit small | 0.04 | 9.0 |
| WhisperKit large-v3-turbo | 0.22 | 7.5 |
Apple Speech Recognition processes an hour of audio in well under a minute. WhisperKit large-v3 quantized needs about seven minutes for the same hour. Both keep up with live recording; the difference shows when you import a long recording afterwards.
Which one to choose
iPhone and iPad
- Keep the default — “AI (Apple Intelligence)” was the most accurate on clear audio and the fastest by far
- Recording a room, with the phone away from the speakers — WhisperKit large-v3 quantized (Pro) was a few points more accurate. It needs a device with 6 GB of memory or more; we did not measure its speed on iPhone
- Skip “Standard” for meetings — it needs no download, but it was the least accurate on recorded audio
Mac
- Switch away from the legacy Apple Speech recognizer — Apple Speech Recognition (Apple Intelligence) or WhisperKit large-v3 quantized, both free, were close to each other and far steadier
- Cloud transcription (Mac, Pro) — in English, OpenAI was less accurate than WhisperKit large-v3 quantized and Apple Speech Recognition on every recording, and gpt-4o-mini-transcribe varied widely between takes on the Mac's microphone (25.0% and 11.6%). Sending meeting audio out of the device and paying per minute bought no accuracy here
How to switch engines and how much memory each needs: Choosing engines (Mac) andEngines and models (iPhone and iPad). For the apps themselves, see Otonota for Mac andOtonota for iPhone.
For our recommendation by device and situation on one page, see Recommended settings at a glance.
How we measured
- Audio
- A fictional weekly project meeting (176 words, about 70 seconds), synthesized with the macOS voice Samantha. All company, person and project names are invented; no real meeting was used
- Recording
- Played through desk speakers and recorded twice each by an iPhone Duo (in Otonota) and by a MacBook Pro's built-in microphone, both on the same desk in the same room
- Processing
- MacBook Pro (M4 Max, 36 GB, macOS 27.3). The live result was transcribed on the iPhone Duo (iOS 27.1)
- Engines
- The same code as Otonota 1.1.4, called from a measurement tool. Apple Speech and OpenAI received audio in real time; the others received the whole file. No glossary was used
- Metric
- Word error rate (WER): lowercased, punctuation removed, numbers normalized (“28th” → “28”, “80,000” → “80000”, “percent” → “%”), then word-level edit distance divided by the number of words in the script
- Measured
- 20 November 2026
This is one short script in one room. Several speakers, a noisy room, accents or fast speech can change the ranking. WhisperKit ran on the Mac; we did not measure its speed on iPhone.
Questions and answers
- What is the most accurate transcription engine for meetings on iPhone?
- With clear audio, the default “AI (Apple Intelligence)” was the most accurate engine we tested (word error rate 4.1%). When the phone was away from the speaker, WhisperKit large-v3 quantized (Pro) did better: 5.8% against 9.1%.
- Which transcription engine should I use on a Mac?
- Switch from the default legacy Apple Speech recognizer to Apple Speech Recognition (Apple Intelligence) or WhisperKit large-v3 quantized. Both are free and scored 9.3% and 9.0% on the built-in microphone, while the legacy recognizer varied between 51.2% and 18.0%.
- What is the difference between Apple Speech Recognition (Apple Intelligence) and Apple Speech?
- Apple Speech Recognition is the newer recognizer introduced with macOS 26 and iOS 26 (SpeechAnalyzer); it downloads a model per language the first time. Apple Speech is the older recognizer (SFSpeechRecognizer) and needs no download. On iPhone and iPad they appear as “AI (Apple Intelligence)” and “Standard”. In our tests the newer one handled distant voices far better.
- Is cloud transcription with OpenAI more accurate than on-device engines?
- Not in English. OpenAI was less accurate than WhisperKit large-v3 quantized and Apple Speech Recognition on every recording; on the iPhone recording gpt-4o-mini-transcribe scored 9.8% against 5.8%. OpenAI is available only in the Mac app with Pro and your own API key, and the audio is sent to OpenAI.
- How fast is on-device transcription?
- On a MacBook Pro (M4 Max), Apple Speech Recognition processed an hour of English audio in well under a minute and WhisperKit large-v3 quantized in about seven minutes. Both keep up with live recording.
- Why does my transcript misspell names?
- Every engine struggles with names it has never heard; none got all 14 of our checked terms right. A glossary that lists the correct spelling and the ways each name gets misheard helps more than switching engines. See Getting names right in meeting transcripts.
