LisanMultilingual
meeting intelligence.
The M3 system powers MeetriX: speech, participant context and a traceable record of the conversation.
System design, meeting workflows and measured Arabic and English transcription.
LisanMeasured in Arabic.
Measured in English.
A multilingual system, with transcription evaluated on real business meetings.
After lexical normalization
After English scoring rules
Matched inputs within each language
Speech that stays connected to its context
M3 brings multilingual transcription, available participant identity, timing and enterprise terminology into one meeting record. MeetriX can use that record to prepare summaries, decisions and follow-up actions.
M3 recorded the lowest character error rate against every external ASR configuration in each language-specific evaluation. Google Chirp 3 was the closest comparison: 2.26% in Arabic and 3.18% in English, versus 1.62% and 1.83% for M3.
28%–83% lower Arabic CER
Relative reduction versus the external configurations on the Arabic evaluation material.
42%–77% lower English CER
Relative reduction versus the external configurations on the English evaluation material.
From system design to inspectable evidence
LisanBuilt for real meetings.
Across languages.
Language, speaker identity and enterprise context belong in the same meeting record.
Real meetings move between short replies, overlapping voices, dialects and business vocabulary. A useful transcript keeps those contributions connected to the people, products and decisions being discussed.
MeetriX product coverage. This paper evaluates Arabic and English transcription separately.
Automatically select the primary meeting language and handle Arabic–English switches within one M3 system.
Keep the available participant identity connected to each contribution as the conversation moves.
Recognise business terminology and resolve references to people, products and platforms.
Preserve brief replies and words at speech boundaries, with links back to their original audio.
Spoken content and enterprise terminology
Available participant identity and context
Meeting time and the original audio
LisanEleven components.
One M3 system.
Three preprocessing models and eight connected components carry speech from capture to the meeting report.
Capture with speaker identity
Keep participant audio distinct.
Carry participant-linked identity.
Prepare the speech
3 preprocessing modelsFind spoken content.
Keep brief contributions.
Protect words at the edges.
Recognise speech in context
Turn speech into text.
Unify business terms.
Resolve people and products.
Produce a traceable record
Connect speaker, time and audio.
Refine the meeting record.
Organise each contribution.
Functional architecture. Speech & silence, short turns and speech boundaries describe proposed preprocessing roles, not official model names. Available platform channels and speaker-activity information determine how participant identity is established.
Speaker identity, meeting time, enterprise terminology and source audio remain linked through live and post-meeting workflows.
LisanThe complete
meeting record.
Transcription quality creates value when the words remain connected to people and context.
Accurate Arabic and English transcripts
Lower normalized transcription error supports a clearer record for review, search and analysis. Human correction time was not measured.
Evidence. M3 achieved 1.62% Arabic CER and 1.83% English CER, the lowest value in each matched-language comparison. The results chapters provide the complete tables.
Cleaner delivered text
Delivered transcripts avoid many of the filler-like insertions, repeated punctuation patterns, and repetition loops that reduce readability.
Evidence. Observed in the reviewed outputs from the Arabic evaluation material; examples are provided in the accompanying comparison page.
Participant-linked records
Decisions, statements, and follow-ups can remain associated with the available meeting participant rather than becoming anonymous text.
Evidence. Platform-provided speaker attribution remained correctly linked throughout the evaluated Arabic and English meetings.
Searchable terminology
Normalised business and technical terms improve enterprise search, knowledge retrieval, and analytics across many meetings.
Evidence. Terminology handling was reviewed on the Arabic evaluation material and is illustrated in the matched-audio examples.
LisanFrom delivery to
controlled deployment.
A configurable meeting workflow, with separate evidence for each operating claim.
Post-meeting summary and delivery
Participants receive a shared record of the topic, attendees, summary, key points, recommendations, and next steps without someone manually composing the follow-up email.
Evidence. Product delivery workflow described in the meeting-intelligence chapters; not part of the CER benchmark.
Fast batch processing
Large meeting backlogs can be processed quickly, and completed meetings can become available while they are still operationally relevant.
Evidence. Batch transcription processes an hour of recorded audio in 45 seconds on one L4 GPU. This is a recorded-audio throughput measurement.
Concurrent meeting capacity
Multiple meetings can share a single accelerator within the operating configuration.
Evidence. A separate operating measurement supports 82 concurrent meetings on one L4 GPU. Capacity and batch speed were measured separately.
Customer-controlled processing
Sensitive financial, contractual, employee, and customer conversations can remain within a controlled deployment boundary.
Evidence. The evaluated transcription path called no external transcription service.
LisanFrom transcript to
participant follow-up.
A structured meeting record makes the conversation easier to review and act on.
After transcription, M3 combines the transcript with available participant identities and the meeting timeline. MeetriX uses that structured record to prepare a concise summary and the participant follow-up.
Words, participants and timestamps
Discussion, decisions and next steps
Registered meeting attendees
Delivery to intended participants
What the structured summary can contain
Title, date and language
Attendees and available participant details
A concise account of the discussion
Important topics and supporting context
Recommendations and agreed direction
Action items and follow-up work
One record, a configurable delivery workflow
The summary template can match the organisation’s workflow. The recipient list is derived from meeting participant information, so the identity context used in the transcript also identifies the intended follow-up recipients.
LisanA record that stays
connected to its context.
Meeting knowledge becomes more useful when people, terminology and source audio stay linked.
A flat text transcript loses part of the meeting. M3 preserves relationships between the meeting, original audio segment, available participant identity, timeline position, transcript and generated summary.
Audio + participant + time + transcript + summary
Consistent terminology supports knowledge retrieval.
Return from a statement to its source audio.
Feed approved knowledge and workflow systems.
Accountability
Keep decisions and commitments connected to the participant and moment from which they came.
Operational insight
Analyse topics, recurring issues, customer signals and meeting activity over time.
Knowledge retrieval
Canonical business terms make the record easier to find in search and retrieval workflows.
Approved automation
Structured records can feed a CRM, case record, knowledge base or compliance archive without reconstructing context from flat text.
LisanLanguage-specific material.
A shared reference within each test.
Each language has its own source recordings. Every compared system is scored against the same references within that language.
| Evaluation | Material | What it supports |
|---|---|---|
| Arabic | Real, multi-dialect Arabic business meetings with English technical terms. | Normalized Arabic CER and qualitative review of the matched system outputs. |
| English | Real English-language business meetings with human-verified references. | English CER and a separate qualitative review of the matched outputs. |
| Product workflow | Participant context, timestamps, structured records and delivery. | Descriptions of meeting capabilities; these are not extra CER results. |
Reference preparation
A human reviewer checked each scored reference against the source audio. The reference retains spoken content, names, numbers and business terminology. Only segments with a verified reference enter the CER calculation. Non-speech material can inform diagnostic review without becoming an additional scored speech result.
Evaluation-data separation
The meetings and human-verified references used in both the Arabic and English evaluations were not used to train M3. References were used only for scoring and were not supplied to any system during transcription.
LisanHold the audio constant.
Isolate transcription quality.
M3 is a meeting system compared with external ASR configurations, including distinct Google configurations.
The M3 system and the external ASR configurations received the same source audio and segment boundaries within each language evaluation. Every system was explicitly configured for Arabic or English, as appropriate, and scored against the same human-verified reference. Automatic language selection was evaluated separately from CER. External results came from API runs on the evaluation audio.
Capture
Participant-linked audio and timing context.
Transcribe
Independent outputs from matched audio.
Score
Shared references and normalization.
Participant identity is not diarization accuracy
The capture workflow already supplied participant-separated audio. External speaker diarization—the step that separates a mixed recording by who spoke—was not used or evaluated. M3 kept every speaker correctly linked to their transcript throughout the evaluated Arabic and English meetings. This checks continuity of available identity context, not independent speaker recognition.
Short utterances stay inside the review
Brief acknowledgements, greetings and short questions can carry a business decision. A speech-bearing reference paired with an empty output contributes deletions; unsupported additional output contributes insertions. Reviewed examples make these behaviors visible without creating a separate aggregate short-utterance score.
LisanExplicit language settings.
Clear scoring rules.
The published configurations define the benchmark; they do not establish a universal ranking of providers.
| Evaluated entry | Arabic setting | English setting |
|---|---|---|
| M3 | Arabic explicitly selected; participant-linked capture | M3 + soxr + 1 s silence; --language en |
| Google Chirp 3 | chirp_3 · ar-XA | chirp_3 · en-US |
| Google Latest Long | latest_long · ar-SA | latest_long · en-US |
| ElevenLabs Scribe v1 | scribe_v1 · ara | scribe_v1 · eng |
| Cohere Transcribe Arabic | cohere-transcribe-arabic-07-2026 · language=ar | cohere-transcribe-arabic-07-2026 · language=en |
| Deepgram Nova-3 | Nova-3; Arabic explicitly selected | nova-3 · en |
| Sonix | Arabic explicitly selected | en |
English normalization
English CER counts character substitutions, deletions and insertions against the reference. Punctuation, letter case and whitespace are ignored, except that a missing space which fuses separate words counts as one error. Remaining spelling differences count. The Arabic lexical-equivalence policy is not applied to the English comparison.
Why Scribe v1 remains in the comparison
The same Arabic evaluation audio was rerun with Scribe v2 under the same input and scoring conditions. It did not improve accuracy on this material; v1 performed slightly better. The table therefore retains v1. This observation does not rank the model versions for other workloads.
LisanArabic: define equivalence
before counting errors
Arabic lexical normalization is part of the Arabic result. English uses the separate rules stated in its configuration chapter.
CER after defined lexical normalisation. Character error rate is calculated as substitutions plus deletions plus insertions, divided by the number of characters in the human reference. Equivalent ways of expressing the same lexical content are treated consistently before character-level alignment:
- An Arabic word is counted as correct when it is transcribed in either a dialectal or Modern Standard Arabic form that carries the same meaning.
- An English word is counted as correct whether it is written in Latin characters or represented phonetically in Arabic script.
- An English word translated into an equivalent Arabic word is also counted as correct.
- Equivalent numeric expressions are counted as the same value whether they are written with Western digits, Arabic-Indic digits, or words; for example,
15,١٥, andخمسة عشرare treated as equivalent. - Punctuation marks are excluded from error scoring, whether they are present, absent, or written differently.
Arabic letter variants such as the different forms of alef are also normalised consistently so that equivalent forms are treated as the same character. Latin text is case-insensitive, so uppercase and lowercase differences are not counted as errors. Whitespace is otherwise excluded from error scoring; however, when a missing word-boundary space joins two separate words, that missing boundary is counted as an error. After all accepted equivalence rules and normalisation steps are applied, every remaining missing, additional, or substituted character is counted as an error. Numbers and English business terms remain part of the evaluated content.
LisanArabic transcription
measured against the same inputs
Character error rate on matched Arabic recordings and human-verified references. Lower is better.
Scroll horizontally to inspect all chart values.
| Evaluated entry | CER | Character accuracy | Gap vs M3 (points) | M3 relative reduction |
|---|---|---|---|---|
| M3 | 1.62% | 98.38% | — | — |
| Google Chirp 3 | 2.26% | 97.74% | 0.64 | 28% |
| Cohere Transcribe Arabic | 4.93% | 95.07% | 3.31 | 67% |
| ElevenLabs Scribe v1 | 5.01% | 94.99% | 3.39 | 68% |
| Deepgram Nova-3 | 6.09% | 93.91% | 4.47 | 73% |
| Google Latest Long | 9.01% | 90.99% | 7.39 | 82% |
| Sonix | 9.80% | 90.20% | 8.18 | 83% |
Character accuracy = 100 − CER. Gap and relative reduction compare each entry with M3; relative reductions are rounded to whole percentages. Arabic lexical-equivalence rules apply.
LisanEnglish transcription
measured against the same inputs
Character error rate on matched English recordings and human-verified references. Lower is better.
Scroll horizontally to inspect all chart values.
| Evaluated entry | CER | Character accuracy | Gap vs M3 (points) | M3 relative reduction |
|---|---|---|---|---|
| M3 | 1.83% | 98.17% | — | — |
| Google Chirp 3 | 3.18% | 96.82% | 1.35 | 42% |
| Cohere Transcribe Arabic | 3.54% | 96.46% | 1.71 | 48% |
| Sonix | 4.21% | 95.79% | 2.38 | 57% |
| Deepgram Nova-3 | 5.58% | 94.42% | 3.75 | 67% |
| ElevenLabs Scribe v1 | 6.41% | 93.59% | 4.58 | 71% |
| Google Latest Long | 7.98% | 92.02% | 6.15 | 77% |
Character accuracy = 100 − CER. Gap and relative reduction compare each entry with M3; relative reductions are rounded to whole percentages. English case, punctuation and whitespace rules apply; remaining spelling differences count.
LisanHow the English
outputs behaved
These review levels describe observed output behavior. They are not additional error-rate measurements.
| Criterion | M3 | Cohere Transcribe Arabic | ElevenLabs Scribe v1 | Deepgram Nova-3 | Google Latest Long | Google Chirp 3 | Sonix |
|---|---|---|---|---|---|---|---|
| Overall English transcription | |||||||
| Short utterances | |||||||
| Text omission | |||||||
| Punctuation | |||||||
| Names, numbers and terminology | |||||||
| Word repetition | |||||||
| Speaker identification | |||||||
| Unsupported fillers and insertions |
The other rows summarize the reference-error review, including approximate token alignment. They do not establish statistical significance or independent aggregate rates. M3’s Strong speaker rating reflects correct participant-to-transcript linkage throughout the evaluated English meetings. External speaker identification was not evaluated; the comparison uses participant-separated clips.
LisanNames that remain
recognisable in the record.
English example · Core42 and Compass in a brief contribution.
access to core 42 compass models, right?
| Evaluated entry | Exact transcript |
|---|---|
| M3 | access to core 42 compass models, right? |
| Google Chirp 3 | Access to code 42 combos, model 3. |
| Cohere Transcribe Arabic | access to core 42 compass models, right? |
| ElevenLabs Scribe v1 | Access to Core 42 combo small drive |
| Deepgram Nova-3 | Access to code 42 Compass model drive. |
| Google Latest Long | Access to Code 42. Composed model Drive. |
| Sonix | Access to code 42 campus models, right? |
Why this matters. M3 and Cohere match the reference exactly. Other outputs change one or more product terms or the closing phrase. The difference affects what a reader can retrieve from the meeting record.
M3 and Cohere matched the reviewed short references. M3 showed consistent preservation of names and terminology; Deepgram and Sonix had recurrent losses.
M3, Chirp 3 and ElevenLabs showed the lowest concentration of omissions. ElevenLabs also showed recurrent filler-like insertions.
Punctuation is excluded from CER. Repeated punctuation and unsupported repetition remain part of the qualitative review.
The viewer also contains “clear” versus “good” or “great,” and the role “note taker” with the participant name “Adam.”
This selected case illustrates the evaluated outputs; it does not replace aggregate CER or establish a general vendor ranking.
LisanTechnical terms.
Inside a Saudi-dialect update.
Arabic example · A routine status update with Kubernetes, scaling and paper.
صباح الخير كنت أمس أشتغل على موضوع في الـ Kubernetes من ناحية scaling كان في مشكلة بسيطة قاعد أدور لها حل وأيضا قاعد أشوف الموضوع البيبر ويعطيكم العافية
صباح الخير كنت أمس أشتغل على موضوع في الـ Kubernetes من ناحية scaling كان في مشكلة بسيطة قاعد أدور لها حل وأيضا قاعد أشوف الموضوع البيبر ويعطيكم العافية
صباح الخير، كنت امس اشتغل على موضوع في الكبرنتيز من ناحيه السكيلينج كان في مشكله بسيطه قاعد ادور لها حل، وايضا قاعد اشوف الموضوع البيبر ويعطيكم العافيه.
صباح الخير، آآآ، كنت أمس أشتغل على موضوع في الـ Kubernetes من ناحية scaling، كان فيه مشكلة بسيطة قاعد أدور لها حل، وأيضًا قاعد أشوف الـ موضوع الـ paper، ويعطيكم العافية.
Why this matters. M3 matches the verified reference and keeps Kubernetes and scaling in Latin script. ElevenLabs also preserves the English terms, with added filler-like content. Chirp 3 uses phonetic Arabic forms, which the Arabic scoring rules accept when equivalent.
Readable terminology and CER are related but different: a canonical spelling can help search even when its phonetic equivalent is accepted during scoring. The full case retains every compared output, including omissions and substitutions in other configurations.
Listen and inspect every Arabic system output ↗LisanWhat the outputs
showed in practice
Observed behaviours explain the measured comparison without creating separate category scores.
The evaluated Arabic meeting outputs were reviewed for content omission, handling of names and English terminology, filler-like insertions, unexpected language changes, repeated punctuation, and repetition loops. These observations explain visible output behaviours and are supported by matched-audio examples. They are not presented as separately scored aggregate metrics unless a numerical result is explicitly stated.
| System | Character error rate (CER) | Observed output characteristics |
|---|---|---|
| M3 | 1.62% | Lowest measured CER, with reliable preservation of business-critical content, numbers, and technical terminology. |
Google Cloud Speech-to-Text V2 (chirp_3) | 2.26% | Strong handling of English technical terminology within Arabic speech, with low content loss in the reviewed output. |
| Cohere | 4.93% | Highly accurate on Arabic speech. The review identified omissions, inconsistent terminology, repeated punctuation and repetition loops. |
| ElevenLabs Scribe v1 | 5.01% | Preserved English technical terminology particularly well. Frequent spurious filler-like vocalisations such as “آآآ” were observed, and one segment switched to an unexpected language or script. |
| Deepgram Nova-3 | 6.09% | Frequent errors were observed in participant names, together with substantial content omissions, particularly at the beginning of source segments. |
Google Cloud Speech-to-Text V1 (latest_long) | 9.01% | Correctly returned no transcript for many non-speech segments, but missed a small number of speech-bearing segments. English technical terms were frequently omitted and numbers were occasionally omitted. |
| Sonix.ai | 9.8% | Frequent errors were observed in names and English technical terms, together with segment-start omissions and excessive sentence-final punctuation. |
LisanM3 and Google Chirp 3
M3 and Google Cloud Speech-to-Text V2 (chirp_3).
Detailed qualitative criteria. Category observations describe the reviewed outputs; aggregate rates were not calculated unless explicitly stated.
| Review criterion | M3 | Google Cloud Speech-to-Text V2 (chirp_3) |
|---|---|---|
| Arabic transcription | Highly accurate. | Accurate. |
| English inside Arabic | Preserved English technical and business terminology more consistently within Arabic speech. | English technical terminology was preserved strongly within Arabic speech. |
| Text omission | Preserves business-critical content more reliably and reduces the loss of important numbers, terminology, and organisation or product names. | Low content loss was observed, with substantially fewer empty outputs than the V1 configuration. |
| Punctuation | Produced clearer sentence boundaries without the same concentration of repeated-punctuation artifacts. | Not separately scored; no recurring punctuation defect was highlighted in the qualitative review. |
| Spelling and terminology | Produces more consistent business terminology for reading and search. | English technical terms were rendered reliably within Arabic speech. |
| Word repetition | Showed substantially fewer repetition problems in the reviewed output. | No recurring repetition issue was isolated in the qualitative review. |
| Filler words and event tags | Did not show the same concentration of filler-like, event-tag, or repeated-punctuation artifacts in the evaluated meeting review. | No recurring filler-word or event-tag pattern was highlighted in the qualitative review. |
LisanCohere and ElevenLabs Scribe v1
Cohere and ElevenLabs Scribe v1, on the same meeting evaluation set.
Detailed qualitative criteria. Category observations describe the reviewed outputs; aggregate rates were not calculated unless explicitly stated.
| Review criterion | Cohere | ElevenLabs Scribe v1 |
|---|---|---|
| Arabic transcription | Highly accurate. | Accurate. |
| English inside Arabic | English technical and business terms were sometimes omitted, mistranscribed, or rendered inconsistently within Arabic speech. | English technical terminology was preserved particularly well in the reviewed output. |
| Text omission | May omit or distort business-critical content, including numbers, technical terms, and organisation or product names. | No separate omission rate was calculated; the principal qualitative issue was inserted filler-like content rather than a recurring segment-start omission pattern. |
| Punctuation | Repeated punctuation patterns that reduce readability were observed in the reviewed output. | Not separately scored; readability was affected more by filler-like insertions than by a recurring punctuation defect. |
| Spelling and terminology | Technical and business terms may appear inconsistently or in phonetic form. | English technical terms were handled well, although filler-like insertions reduced overall transcript cleanliness. |
| Word repetition | Token and phrase repetition loops were observed in the reviewed output. | No separate repetition rate was calculated; frequent filler-like vocalisations were the more prominent insertion issue. |
| Filler words and event tags | Filler-like insertions, event-style tags, and repeated-punctuation artifacts were observed in the reviewed output. | Frequent spurious filler-like vocalisations such as “آآآ” were observed, and one segment switched to an unexpected language or script. |
LisanDeepgram and Google Latest Long
Deepgram Nova-3 and Google Cloud Speech-to-Text V1 (latest_long).
Detailed qualitative criteria. Category observations describe the reviewed outputs; aggregate rates were not calculated unless explicitly stated.
| Review criterion | Deepgram Nova-3 | Google Cloud Speech-to-Text V1 (latest_long) |
|---|---|---|
| Arabic transcription | Accurate. | Many errors. |
| English inside Arabic | English-term performance was not scored separately; reviewed examples showed inconsistent rendering alongside segment-start omissions. | English technical terms were frequently omitted. |
| Text omission | Substantial content loss was observed, particularly at the beginning of source segments. | Correctly rejected many non-speech segments, but missed a small number of speech-bearing segments and sometimes omitted English terms or numbers. |
| Punctuation | Not separately scored; no recurring punctuation defect was isolated in the qualitative review. | Not separately scored; no recurring punctuation defect was isolated in the qualitative review. |
| Spelling and terminology | Participant names were frequently transcribed incorrectly. | English technical terms were frequently missed, and some numbers were omitted. |
| Word repetition | No recurring repetition issue was isolated in the qualitative review. | No recurring repetition issue was isolated in the qualitative review. |
| Filler words and event tags | No recurring filler-word or event-tag pattern was isolated in the qualitative review. | Handled many non-speech segments correctly, although a small number of speech-bearing segments were also rejected. |
LisanSonix and shared conditions
The final system profile and the common dialect and speaker-context conditions.
| Review criterion | Sonix.ai |
|---|---|
| Arabic transcription | Accurate. |
| English inside Arabic | Names and English technical terminology were frequently transcribed incorrectly or omitted. |
| Text omission | Content loss was observed particularly at the beginning of source segments. |
| Punctuation | Punctuation quality was weak, with excessive sentence-final marks observed in the reviewed output. |
| Spelling and terminology | Performance was weak on participant names and English technical terminology. |
| Word repetition | No recurring word-repetition issue was isolated; punctuation was the more prominent readability problem. |
| Filler words and event tags | No recurring event-tag pattern was isolated; excessive punctuation was the more prominent artifact. |
Dialect coverage
All evaluated entries produced Arabic transcripts across the dialects represented in this evaluation set.
Speaker identification
Not evaluated for the external systems. Each system received the same participant-separated audio segments prepared by the M3 capture workflow and was scored only on transcription.
M3: Uses participant-linked audio channels and platform participant information. Platform-provided speaker attribution remained correctly linked throughout the evaluated Arabic and English meetings.
LisanFast processing.
Connected meeting context.
Batch speed and concurrent capacity are separate measurements on one L4 GPU.
For an hour of recorded audio
Processing time on one L4 GPU.
Concurrent meetings
In a separate operating measurement on one L4 GPU.
These figures describe different workloads. They are not a combined speed-and-capacity benchmark or an end-to-end live-latency measurement.
Participant identity travels with the audio
M3 uses participant-linked audio channels and available platform information to keep identity attached to transcript segments. Each speaker remained correctly linked to their transcript throughout the evaluated Arabic and English meetings. External diarization and a separate speaker-attributed error rate were not evaluated.
Available participant identity
Position in the meeting
The original audio segment
From processing to a usable record
Fast batch processing makes recorded conversations available for review sooner. Participant context and source alignment support navigation and follow-through. The measured CER results establish transcription accuracy under the stated scoring rules; summary quality and delivery time are separate evaluation questions.
LisanOne system.
The right language, automatically.
Automatic primary-language selection, with Arabic–English switching inside the same meeting.
Before transcription, M3 identifies the meeting’s primary language from its audio and applies the appropriate language handling. Participants do not need to choose a language manually.
Use the meeting audio to identify its primary language.
Apply the appropriate language handling within M3.
Handle Arabic–English switches as the discussion moves.
Arabic and English. One M3 system.
Arabic and English are handled by the same M3 system, rather than separate models. Selecting the primary language does not prevent M3 from handling the other language when it appears in the conversation.
M3 identified the correct primary language for every evaluated Arabic and English meeting. This was a separate workflow check on the same meeting material.
For CER, M3 and every comparison system were explicitly configured for the language being evaluated. Automatic selection did not contribute to the CER score.
Evaluate on the meetings that matter
Customer evaluations can measure language selection, transcription, participant attribution and meeting intelligence separately, using matched inputs and human-verified references.
LisanResults with
clear boundaries
Keep benchmark findings, product coverage and operational capabilities distinct.
| Finding | Evidence boundary |
|---|---|
| Arabic transcription: M3 1.62% CER | Matched Arabic audio and human-verified references, after the stated lexical equivalences. |
| English transcription: M3 1.83% CER | Matched English audio and human-verified references, after English normalization. |
| Qualitative output behavior | Criterion-level review of actual outputs. No separate aggregate rates or significance tests. |
| Automatic primary-language selection | Correct for every evaluated Arabic and English meeting in a separate workflow check; outside CER. |
| Participant linkage | Correct participant-to-transcript linkage throughout evaluated Arabic and English meetings. External diarization was not tested. |
| Batch processing: 45 seconds per recorded hour on one L4 GPU | Recorded-audio transcription throughput. Separate from concurrent capacity, live latency and downstream delivery. |
| Concurrent capacity: 82 meetings on one L4 GPU | A separate operating measurement; not a combined batch-speed or live-latency benchmark. |
| Product coverage: 40 languages and 32 Arabic dialects | MeetriX product scope. The CER results in this paper apply to the separate Arabic and English evaluations. |
| Customer-controlled deployment | The evaluated transcription path called no external transcription service. Deployment policy and controls must be assessed in their own scope. |
The paper brings together methods, measured results and inspectable transcript evidence.
LisanHear the input.
Read the exact output.
Inspect Arabic and English examples alongside their human-verified references.
Listen to the audio
Source clips make the reviewed examples inspectable. The same clip is the input for every system shown. Selected examples do not become a new aggregate benchmark.
Compare exact transcripts
Each case keeps the human reference and named system outputs. Corrected Arabic display examples preserve the existing benchmark results; the corrections do not add new scored audio or alter the aggregate CER.
Apply the right scoring rule
Arabic uses the stated lexical equivalences. English ignores punctuation, case and whitespace while retaining the merged-word exception. Readability and searchability differences can remain even when a variant is equivalent under CER scoring.
The published evaluations establish the matched-language results shown here. The source material does not support confidence intervals, general multilingual error rates or independent certification of the benchmark.
Open audio and transcript comparisons ↗Read the evaluation protocol ↗