Ditto

Benchmarks · October 2026

Tested on a real meeting.

Four people around a table, talking over each other, picked up by one microphone in the middle: the audio a phone actually hears. We ran it through Ditto on an iPhone 13 from 2021 and scored every result against a transcript written by people.

The test meeting AMI EN2002a · four people · one table mic
0:00 / 35:42
Audio MP3, 4.3 MB Human transcript JSON, 195 KB
Processing time
53 s
to transcribe an hour-long meeting and work out who spoke
First words
3 s
until the first words of the transcript appear
Speaker accuracy
93.3%
of speech credited to the right person, with High accuracy speakers

Speed

Everything runs on the phone, so there is no upload and no server queue. We test on an older iPhone on purpose: an iPhone 13 finishes an hour-long meeting in 53 seconds with High accuracy.

The whole meeting, and the time Ditto needs for it
53 s with High accuracy 59 min 6 s of meeting
Time to finish, transcript and speakers · shorter is better
High accuracy 53 s
On-device 76 s
Peak memory
150 MB
Thermal state
Fair at peak
Data sent
None

The first recording after you install or update Ditto takes longer, while iOS prepares the speech models for your phone.

Words

Word error rate counts every word missed, added or misheard against the human transcript. High accuracy runs entirely on the phone and matches Ditto’s cloud transcription.

Word error rate, table microphone · lower is better
High accuracy 38.5%
Cloud (Pro) 39.6%
On-device 48.5%

Why the numbers look high

A distant microphone in a four-way conversation is the hardest case for speech recognition. The people who transcribed this meeting wrote down every “yeah” and every word said over someone else, and most of what Ditto misses is exactly that: short replies and crosstalk, not the sentences you read back later. From the headset microphones, the same meeting scores 29.4% with High accuracy.

Speakers

Speaker error is the share of the meeting where Ditto gets wrong who is talking, or whether anyone is. Measured strictly over 16 meetings and 9 hours of audio, overlapping speech included.

Speaker error, 16 meetings · lower is better
High accuracy (beta) 12.1%
Standard 43.8%

Turn on High accuracy in Settings → Speakers. It works out how many people there are by itself; when you set the number of people, Ditto uses Standard. Either way, you can fix a label by hand, and once you name someone, Ditto recognizes their voice in later meetings.

Next to other apps

The same file, given to other transcription apps and services with their default settings. High accuracy makes the fewest word errors of all of them, and the fewest once speakers count too, without leaving the phone. Argmax Pro comes closest.

Word error and words-plus-speakers error on the same meeting, lower is better
App Word error Words + speakers error
Ditto High accuracy iPhone 13 38.5% 39.0%
Ditto Cloud (Pro) Cloud 39.6% 53.4%
Plaud Cloud 40.1% 62.4%
Argmax Pro iPhone 13 40.4% 41.3%
Ditto On-device iPhone 13 48.5% 57.8%
Deepgram Cloud 62.5% 71.0%
Gladia Cloud 82.4% 87.3%
Lower is better. Words + speakers compares each person’s words with what the app attributed to them, so a wrong word and a word given to the wrong person both count. The meeting has four people.

Ask your meetings

112 questions about 12 meetings, each answer checked against notes a person wrote about that meeting. A partly right answer counts as half. Answers take 2 to 3 seconds.

Answers correct · higher is better
Cloud, one meeting 85%
Cloud, 4 meetings 78%
On-device, one meeting 76%
On-device, 4 meetings 62%

On-device answers need an iPhone with Apple Intelligence.

How we tested

The meeting
EN2002a from the AMI Meeting Corpus (CC BY 4.0): 35 minutes 42 seconds, four people, from the microphone array on the table. Every app got the same 16 kHz MP3, made from the original recording.
The phone
iPhone 13, release build of Ditto, October 2026. Speed was measured on a longer meeting from the same corpus and microphone, EN2001d (59 minutes 6 seconds). High accuracy means High accuracy for both the words and the speakers; On-device is the default. Speed and memory come from the app’s own diagnostics.
Scoring words
Both transcripts are lowercased, without punctuation or fillers such as “um” and “uh”, then compared word by word in time order. The human transcript has one entry per utterance, with speaker and start and end times in seconds, so you can score any app the same way.
Scoring speakers
The 16 meetings of the AMI test set, single distant microphone. No tolerance around the edges of each turn, and time when people overlap counts.
Other apps
Run by hand in October 2026, each on its own app or website with default settings, on the same file. Argmax Pro ran on the same iPhone 13.
One meeting
The comparison uses one hard meeting. It shows where each app stands on this kind of audio, not a ranking for every recording.
Scoring answers
Questions on 12 AMI meetings, written from human meeting notes. A language model grades each answer against those notes.
Your meetings
Results depend on the room, the microphone and how often people talk over each other. A headset, or a quiet room with fewer people, does better than this meeting.