Otonota

Tips

Getting names right in meeting transcripts

Names of people and companies trip up every transcription engine. We tried a glossary and a short meeting description under several conditions and counted what they fix and what they cannot.

Published Measured:

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

  • What worked best was adding the ways a name actually got misheard (Quillon heard as “Quillen”, Talvane as “Talvein”, Okonkwo as “Conquo”) to the glossary entry for that name. With WhisperKit, the number of invented names transcribed correctly went from 5 to 9 out of 10 on the iPhone recordings and from 5 to 9 on the Mac's built-in microphone.
  • Simply listing the correct spellings changed little in English. It mainly fixes capitalization, which most engines already got right.
  • Some things a glossary cannot fix: names that were not heard at all, and names misheard as ordinary words (“Nguyen” came out as “when”). Passing the glossary to the recognizer as a hint did not change the result for Apple Speech Recognition and barely changed it for OpenAI.

Three places a glossary is used

In Otonota you write one term per line, optionally followed by other ways it is said or heard in brackets. The Mac app keeps a glossary per session; the iPhone and iPad app keeps one for the whole app. Both let you save glossaries under a name. On every device, each entry is used in three places:

  1. Correcting the transcript — words that match an entry, or one of its listed alternatives, are replaced with the entry's spelling. This works with every engine
  2. Hinting the recognizer — the terms are passed to the speech recognizer in advance. Apple's built-in engines and OpenAI (Mac only) receive them; WhisperKit does not
  3. Summaries — the terms are given to the language model as the correct spellings when it writes summaries and translations

We measured how much each one helps, using the same fictional weekly meeting as Which transcription engine is most accurate for meetings? It contains 10 invented names of people, companies and projects, unfamiliar to every engine.

Counting names under each condition

We recorded the meeting twice on an iPhone and twice with a Mac's built-in microphone, and transcribed each recording while changing how the glossary was used. Table 1 covers the engines you can choose on iPhone and iPad, table 2 the ones on the Mac. The glossary correction works the same way on both.

iPhone and iPad

Table 1Engines on iPhone and iPad: names transcribed correctly from the iPhone recordings, out of 10 (English, average of two, higher is better)
EngineNo glossaryCorrectionHintCorrection + hint+ mishearings
AI (Apple Intelligence)6.06.06.06.09.0(best)
WhisperKit large-v3 (quantized)5.05.0——9.0(best)

The highlighted row is the default engine. WhisperKit is part of Pro. The app does not pass hints to it, hence “—”, and its “+ mishearings” uses correction only; for “AI (Apple Intelligence)” it uses correction and hints. Mishearings were taken from the two Mac recordings.

Both engines reached 9 of 10 once the mishearings were listed: WhisperKit large-v3 (quantized) from 5, and the default “AI (Apple Intelligence)” from 6. Listing the correct spellings alone, or passing them as hints, changed nothing.

Mac

Table 2Engines on the Mac: names transcribed correctly from the built-in microphone recordings, out of 10 (English, average of two, higher is better)
EngineNo glossaryCorrectionHintCorrection + hint+ mishearings
Apple Speech Recognition (Apple Intelligence)5.05.05.05.06.5
WhisperKit large-v3 quantized5.05.0——9.0(best)
OpenAI gpt-4o-transcribe6.06.56.06.56.5

WhisperKit columns as in table 1. Mishearings were taken from the other recording. OpenAI is part of Pro and uses your own API key.

Listing the correct spellings alone changed almost nothing: the misheard forms (“Quillen” for Quillon, “Zephyrim” for Zephyrine) are different words, not different capitalizations, so nothing matched. Hints did not help either. Apple Speech Recognition produced exactly the same transcript with and without them, on the iPhone and the Mac recordings alike; OpenAI's results moved by about one name, which is within its normal variation from run to run.

Teaching the glossary its mishearings

The big gain came from adding mishearings. If “Quillon” keeps coming out as “Quillen”, write Quillon(Quillen). From then on, “Quillen” is replaced with “Quillon”.

To keep this honest, the mishearings always came from a different recording than the one measured: for the Mac, from the other take; for the iPhone, from the two Mac takes. That is how you would grow a glossary from earlier meetings, not by looking at the answers.

Names transcribed correctly with WhisperKit large-v3 quantized, out of 10 (English, average of two recordings)No glossary: iPhone recording 5.0, Mac built-in mic 5.0; Names in the glossary: iPhone recording 5.0, Mac built-in mic 5.0; Plus the mishearings: iPhone recording 9.0, Mac built-in mic 9.0No glossary: iPhone recording 5.0 / Mac built-in mic 5.0No glossary5.05.0Names in the glossary: iPhone recording 5.0 / Mac built-in mic 5.0Names in the glossary5.05.0Plus the mishearings: iPhone recording 9.0 / Mac built-in mic 9.0Plus the mishearings9.09.00246810Names transcribed correctly with WhisperKit large-v3 quantized, out of 10 (English, average of two recordings)No glossary: iPhone recording 5.0, Mac built-in mic 5.0; Names in the glossary: iPhone recording 5.0, Mac built-in mic 5.0; Plus the mishearings: iPhone recording 9.0, Mac built-in mic 9.0No glossary: iPhone recording 5.0 / Mac built-in mic 5.0No glossary5.05.0Names in the glossary: iPhone recording 5.0 / Mac built-in mic 5.0Names in the glossary5.05.0Plus the mishearings: iPhone recording 9.0 / Mac built-in mic 9.0Plus the mishearings9.09.00246810Names transcribed correctly with WhisperKit large-v3 quantized, out of 10 (English, average of two recordings)No glossary: iPhone recording 5.0, Mac built-in mic 5.0; Names in the glossary: iPhone recording 5.0, Mac built-in mic 5.0; Plus the mishearings: iPhone recording 9.0, Mac built-in mic 9.0No glossary: iPhone recording 5.0 / Mac built-in mic 5.0No glossary5.05.0Names in the glossary: iPhone recording 5.0 / Mac built-in mic 5.0Names in the glossary5.05.0Plus the mishearings: iPhone recording 9.0 / Mac built-in mic 9.0Plus the mishearings9.09.00246810
Fig. 1Names transcribed correctly with WhisperKit large-v3 quantized, out of 10 (English, average of two recordings)
Show the numbers
ItemiPhone recordingMac built-in mic
No glossary5.05.0
Names in the glossary5.05.0
Plus the mishearings9.09.0

No glossary

…last week's action item of the 12 Quillen prototypes. … did you reorder the parts from Talvein Electronics? … I'll attach the quote from a Conquo fabrication. Also, we plan to renew the warehouse contract with Zephyrim Logistics…

With mishearings

…last week's action item of the 12 Quillon prototypes. … did you reorder the parts from Talvane Electronics? … I'll attach the quote from a Okonkwo Fabrication. Also, we plan to renew the warehouse contract with Zephyrine Logistics…

Fig. 2The same recording transcribed by WhisperKit large-v3 quantized on the Mac (excerpts from the second built-in-mic take)

Correct names and how they were transcribed without a glossary: Quillon → “Quillen”, Talvane Electronics → “Talvein Electronics”, Okonkwo Fabrication → “Conquo fabrication”, Zephyrine Logistics → “Zephyrim Logistics”. All four were fixed by mishearings seen in the first recording. The leftover “a” before Okonkwo is the engine's; the glossary replaces only the matched words.

On the Mac, Apple Speech Recognition gained one or two names and OpenAI none. Their mishearings changed from one recording to the next, so last week's list did not always match this week's errors.

What a glossary cannot fix

  • Names that were not heard — if a name is missing from the transcript, there is nothing to correct
  • New mishearings — only the variants you have listed are replaced. The glossary gets better as you add what each meeting reveals
  • Mishearings that are ordinary words — “Nguyen” was transcribed as “when”. Listing “when” as a variant would replace every “when” in the meeting, so don't

Effect on summaries

The glossary and the meeting description (one to three lines about what the meeting is) are also passed to the summarizer. We took an iPhone recording of the meeting, applied the same correction as the app, had Apple Intelligence write a one-line summary every five lines, and counted the names that appeared correctly (table 3). We measured this with the iPhone and iPad app; the Mac app passes the glossary and the description to Apple Intelligence summaries in the same way.

Table 3Names appearing correctly in the transcript and in the summaries, out of 10 (English, recorded on iPhone, summarized by Apple Intelligence)
ConditionTranscriptSummaries
Nothing54
Meeting description only53
Glossary only54
Glossary + description55
D + mishearings75

The mishearings in E include ones seen in this recording itself.

The summaries contained 3 to 5 correct names whatever we did. A summary is written from the transcript, so a name the transcript lost usually does not reappear; with the glossary and description (D), “AquaTech” showed up in the summaries without being in the transcript, but that was a single name. The meeting description on its own (B) even cost one name. Fix the transcript first; the summary follows.

How to build a glossary

  1. List the people, companies, products and abbreviations that come up, one per line. Put the important ones first (summaries receive the first 30)
  2. After the first meeting, read the transcript and note how each name was misheard
  3. Add those mishearings in brackets, separated by commas: Zephyrine Logistics(Zepherin logistics,Zephyrim Logistics). Skip any that are ordinary words
  4. It works from the next meeting on. Save glossaries under a name to switch between them per meeting

The glossary screens are described in Summaries and translation (Mac) and Summaries and the glossary (iPhone and iPad). If the glossary seems to have no effect, seeTroubleshooting. For the apps, see Otonota for Mac and Otonota 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) with 10 invented names, synthesized with the macOS voice Samantha, played through desk speakers and recorded twice by a MacBook Pro's built-in microphone and twice by an iPhone Duo
Engines
WhisperKit large-v3 quantized, Apple Speech Recognition (Apple Intelligence) and OpenAI gpt-4o-transcribe, using the same code as Otonota 1.1.4, run on a MacBook Pro (M4 Max, macOS 27.3)
Glossary
The 10 names. For the mishearings condition, 0 to 3 variants per name taken from the other recording's transcripts (all three engines)
Summaries
The second iPhone recording, transcribed by Apple Speech Recognition, corrected as in the app, then summarized one line per five lines by Apple Intelligence (run in the iOS Simulator on a Mac)
Counting
Whether each of the 10 names appeared at least once with the correct spelling, ignoring case. “Kestrel” counts for “Project Kestrel”
Measured
20 November 2026

Ten invented names and one script. Hints may behave differently with other engine versions or other kinds of words.

Questions and answers

How do I get names of people and companies right in a meeting transcript?
Add each name to the glossary, then add the ways it actually got misheard in brackets, for example Quillon(Quillen). In our test with WhisperKit, the number of invented names transcribed correctly went from 5 to 9 out of 10 on iPhone recordings, and from 5 to 9 on a Mac's built-in microphone.
Is listing the correct spelling in the glossary enough?
Rarely, in English. It fixes differences in capitalization, but a misheard name such as “Zephyrim” for Zephyrine is a different word and is only replaced once you list that variant. Listing spellings alone left WhisperKit at 5 of 10.
Does passing the glossary to the recognizer as a hint improve recognition?
Not in our tests. Apple Speech Recognition produced exactly the same transcript with and without hints, and OpenAI got 6 and 6 of 10 names right without and with them. The correction applied after transcription is what made the difference.
Does a meeting description make names in summaries more accurate?
Not noticeably. Summaries contained 3 to 5 of the 10 names whatever we did, and the description on its own went from 4 to 3. Summaries are written from the transcript, so fix the transcript first.
What can a glossary not fix?
Names that were not heard at all, new mishearings you have not listed yet, and mishearings that are ordinary words (“Nguyen” heard as “when”). Do not list ordinary words as variants; every occurrence would be replaced.
Does the glossary work the same on iPhone and Mac?
The correction after transcription and the hand-off to summaries work the same. The Mac app keeps a glossary per session and the iPhone and iPad app one for the whole app; both let you save glossaries under a name. Recognizer hints go to Apple’s built-in engines and to OpenAI (Mac only), not to WhisperKit.