Newsroom
Introducing M3
The multilingual meeting transcription system inside MeetriX, with an evaluation you can read: 1.62% character error rate on real multi-dialect Arabic meetings and 1.83% on English meetings, lower than each of the six external ASR model configurations compared.
M3, the multilingual meeting transcription system inside MeetriX, recorded a 1.62% character error rate on real multi-dialect Arabic business meetings with human-verified reference transcripts, lower than each of the six external ASR model configurations it was benchmarked against. Today we are publishing the evaluation report behind that number: the scoring rules, the transcript examples and an evidence map, measured against six external ASR model configurations on the same audio and the same references. M3 is the part of MeetriX that turns what people said in a meeting into a speaker-linked, time-aligned transcript.
The full field, measured on the same audio and scored by the same rules: Google Chirp 3 at 2.26%, then Cohere Transcribe Arabic at 4.93%, ElevenLabs Scribe v1 at 5.01%, Deepgram Nova-3 at 6.09%, Google Latest Long at 9.01% and Sonix at 9.8%. The headline figure is measured after defined lexical normalization, and the report says what one evaluation set does and does not support.
English results. The report also covers an English benchmark on real English-language business meetings, run separately from the Arabic one. M3 recorded a 1.83% character error rate, lower than each of the six external ASR model configurations (the same six models, set to English), with Google Chirp 3 next at 3.18%. The two languages use separate evaluation sets, so each figure is read within its own table. Read the English results.
The M3 system records the lowest character error rate against six external ASR model configurations on the same human-verified meetings
Lower is better. Character error rate (CER) after defined lexical normalization.
What goes wrong, and what does not
A meeting record fails in specific ways, and the report shows them segment by segment. In a daily stand-up full of developer vocabulary, the human-verified reference keeps SDK, rendering, highlights, editor, configuration and Chrome extension in Latin script, and M3 keeps all of them except editor, which it writes in Arabic script. Google Chirp 3 transcribed the passage completely but phonetically, in Arabic letters; Cohere Transcribe Arabic did the same and garbled the ending; Sonix dropped the partner name twice. A phonetic spelling is not an error under the scoring rules, so the difference is what a search across the archive will find later.
In another, a greeting shorter than a second is the whole segment. M3, Cohere Transcribe Arabic, Sonix and Google Chirp 3 matched the reference كيفكم؟ exactly. Deepgram Nova-3 and Google Latest Long returned nothing, which the scoring rules count as deletions, and ElevenLabs Scribe v1 produced an English word that was not said. Short turns are where meetings carry their yes, no and agreed, and the report treats them as content.
The comparison also credits what the external configurations did well. Google Chirp 3 handled English technical terminology strongly within Arabic speech and showed low content loss, and on several example segments it is as complete as M3, with the difference limited to phonetic rather than Latin spellings. ElevenLabs Scribe v1 preserved English terminology particularly well. The report's qualitative observations are labeled as such: no separate aggregate rates were calculated for them.
How to read the numbers
CER counts every character that is wrong, missing or extra, divided by the characters in the human-verified reference, after a set of published equivalences: a dialect spelling and its standard form count the same, an English term counts whether written in Latin letters or phonetically in Arabic script, numbers are compared by value, and punctuation is ignored. In practical terms, 1.62% CER corresponds to roughly 16 character edits per 1,000 reference characters after the stated normalization rules. The throughput figure, 45 seconds for a one-hour meeting on an L4 GPU, describes batch processing of recorded audio; live latency is measured separately. The evaluation set is a single set of real meetings; the report says what that supports and what it does not, and the next phase expands to additional organizations, platforms, dialects and acoustic conditions.
M3 is not perfect, and the examples show that too: a spoken first name omitted, a singular where the reference has a plural, and words added that the reference does not contain. We publish those alongside the cases where M3 keeps a term or a name that other systems drop.
Where to go next
The full report, with both leaderboards, the qualitative matrices, nine transcript examples (six Arabic, three English), the methodology and the evidence maps, is at lisan.com/company/research/m3. The product view, including what the result means for a government or enterprise meeting record, is on the MeetriX site. M3 is live in MeetriX today, and organizations can evaluate it on their own meetings against a human-verified reference through a pilot. Write to support@lisan.com.
Start with one product.
Keep the whole platform.
Open a free account and solve today's problem in the next ten minutes. When you are ready for more, six flagships and a 20+ app workspace are already on your account. Or talk to us and we will scope it with you.
Free to start, no card · Your data exportable, always · Trusted by more than 24 government entities