Tips
Recommended settings at a glance
For iPhone, iPad and Mac, the transcription and summary settings we recommend for each situation — talking up close, recording a room, importing recordings, summarizing in another language — based on our measurements.
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.
These recommendations are a starting point from the conditions we tested. The size and echo of the room, how many people speak and how far they are from the device, the microphone, and how people talk can change the best setting completely. Before an important meeting, record the same situation briefly with two settings and compare the transcripts.
Three changes worth making first
- On the Mac, switch transcription to Apple Speech Recognition (Apple Intelligence) — the default legacy Apple Speech recognizer was erratic as soon as the voice came from a distance
- Recording a room from a distance? Try WhisperKit large-v3 (quantized) — free on the Mac, Pro on iPhone and iPad; it was the most accurate on distant voices
- For names, list the mishearings in the glossary — not just the correct spelling but the ways each name was actually misheard; that fixed the most names
The tables below give our recommendation for each device and situation. The numbers come from our English measurements of a fictional meeting read aloud. The full measurements are inWhich transcription engine is most accurate for meetings?, Getting names right in meeting transcripts andWhich languages can on-device summary AI handle?
How to read the tables
- Percentages are word error rates — how far the transcript was from the correct text. 10% means about 10 words in every 100 were wrong, missing or extra. Lower is better
- “Close up”, “distant” and “built-in mic” — close up is a clean audio file; distant is a recording made by an iPhone away from the speakers; built-in mic is a recording made with a MacBook Pro's built-in microphone
- “Misses” in the summary rows — answers that did not come back in the requested language (another language, mixed scripts, or empty)
iPhone and iPad
| Situation | Transcription | Summaries | What we measured (English) |
|---|---|---|---|
| Talking close to the device (one-on-one, voice notes) | AI (Apple Intelligence) (the default) | Apple Intelligence (the default) | Close-up word error rate 4.1%, the lowest; also the fastest |
| Recording a room, phone away from the speakers | WhisperKit large-v3 (quantized) (Pro, 6 GB memory or more) | Apple Intelligence | Distant-voice word error rate: AI 9.1% → WhisperKit 5.8% |
| Start right away, no download | Standard | Apple Intelligence | Word error rate 14.0% close up and 14.2% at a distance — usable, but the least accurate |
| Many names of people and companies | Any of the above + list mishearings in the glossary | The same glossary is passed to summaries | WhisperKit: 5 → 9 of 10 names |
| Summaries in other major languages | Match the meeting language | Apple Intelligence (the default) | 3 misses in 54 requests, all keyword lists |
| Summaries in Arabic, Russian or Thai | Match the meeting language | MLX Qwen3.5 2B or 4B (Pro) | Apple Intelligence does not support them; with English meetings both MLX models answered in the right language every time |
Word error rate: wrong, missing or extra words per 100 (lower is better).
Choose the transcription engine under Recognition Engine in Settings and the summary engine under Summary Engine (Pro); the glossary is under Glossary. See Choosing engines and models andSummaries and the glossary.
Mac
| Situation | Transcription | Summaries | What we measured (English) |
|---|---|---|---|
| Talking close to the Mac, or online-meeting audio | Apple Speech Recognition (Apple Intelligence) (change from the default Apple Speech) | Apple Intelligence (the default) | Close-up word error rate 4.1%; an hour of audio in under a minute |
| Recording a room with the Mac's built-in mic | Apple Speech Recognition (Apple Intelligence) or WhisperKit large-v3 quantized (both free) | Apple Intelligence | Built-in-mic word error rate: Apple Speech Recognition 9.3% / WhisperKit 9.0%; the default Apple Speech varied between 51.2% and 18.0% |
| Importing a meeting recorded on iPhone | WhisperKit large-v3 quantized (free) | Apple Intelligence | Word error rate on an imported iPhone recording: Apple Speech Recognition 9.1% / WhisperKit 5.8% |
| Cloud transcription | Not needed for accuracy in English | — | OpenAI gpt-4o-mini-transcribe: word error rate 9.8% on the imported iPhone recording, behind the on-device engines |
| Many names of people and companies | Any of the above + list mishearings in the glossary | The same glossary is passed to summaries | Listing mishearings helped most; recognizer hints did not help |
| Summaries in Arabic, Russian or Thai | Match the meeting language | Any MLX model (Pro) or a cloud model (Pro) | With English meetings, every MLX model from 2B to 27B and every cloud model but one answered in the right language |
| A Mac without Apple Intelligence | As above | MLX Qwen3.5 2B (free) | 4 misses in 72 requests with English meetings; use a larger model for non-English meetings |
Word error rate: wrong, missing or extra words per 100 (lower is better). Cloud engines (OpenAI, Claude, ChatGPT, Gemini) are part of Pro on the Mac and use your own API key; audio or text is sent to that provider.
Pick engines from the engine names at the top of the transcript and summary panes; open the glossary from the session in the sidebar. See Choosing engines andSummaries and translation. For the apps, see Otonota for Mac and Otonota for iPhone.
True in every situation
- The defaults stay on the device — with the default settings, neither audio nor text leaves your device
- Summaries are only as good as the transcript — a name the transcript lost rarely comes back in the summary, so fix the transcript first
- Full translations always came back in the requested language — language mix-ups happened in one-line summaries and keyword lists
How we measured
This page summarizes three measurements made on the same day: a fictional weekly meeting (176 words of English) read aloud, recorded twice each by an iPhone Duo and a MacBook Pro's built-in microphone, and processed with the same code as Otonota 1.1.4. The procedures and all numbers are in each article's “How we measured” section.
Questions and answers
- What are the best Otonota settings for recording meetings on iPhone?
- Keep the default “AI (Apple Intelligence)” when people speak close to the phone; it was the most accurate on clear audio (4.1%). When the phone sits away from the speakers, WhisperKit large-v3 (quantized, Pro) did a little better: 5.8% against 9.1%. Keep Apple Intelligence for summaries.
- What are the best Otonota settings for recording meetings on a Mac?
- Switch transcription from the default Apple Speech 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. For recordings imported from an iPhone, WhisperKit large-v3 quantized was the most accurate.
- How do I get names right in transcripts?
- Use the glossary rather than switching engines. List each name, then add the ways it got misheard in your first meeting, for example
Quillon(Quillen). See Getting names right in meeting transcripts. - Can Otonota summarize an English meeting in Thai or Arabic?
- Yes, with an MLX model (Pro); Apple Intelligence does not support those languages. With English meetings, MLX Qwen3.5 2B and 4B on iPhone and iPad, and every MLX model on the Mac, answered in the requested language every time.
- Can I use Otonota without sending meeting audio anywhere?
- Yes. With the default settings, transcription and summaries run on the device. Audio or text leaves the device only if you choose a cloud engine (OpenAI, Claude, ChatGPT or Gemini) in the Mac app with Pro and your own API key.
