M3, measured in Arabic and English.
Lisan’s September 2026 evaluation reports 1.62% Arabic CER and 1.83% English CER on separate business-meeting recordings.
Compare configurations within each language.
Each entry received matched inputs and human-verified references within its language evaluation. Arabic uses the stated lexical-normalisation rules; English has its own scoring rules. The two language sets are separate.
| Evaluated configuration | Arabic CER | English CER |
|---|---|---|
| M3 | 1.62% | 1.83% |
| Google Chirp 3 | 2.26% | 3.18% |
| Cohere Transcribe Arabic | 4.93% | 3.54% |
| ElevenLabs Scribe v1 | 5.01% | 6.41% |
| Deepgram Nova-3 | 6.09% | 5.58% |
| Google Latest Long | 9.01% | 7.98% |
| Sonix | 9.8% | 4.21% |
What the numbers describe.
CER counts character substitutions, deletions and insertions against the reference after the defined normalisation. These results measure transcription, not summary quality, external diarization, customer savings or complete meeting applications.
Operating measurements, separately scoped.
45 seconds per audio hour
Reported batch transcription throughput on one L4 GPU.
82 concurrent meetings
A separate operating measurement on one L4 GPU.
40 languages · 32 Arabic dialects
Product coverage; the CER tables evaluate Arabic and English.
Read the complete method.
The canonical report contains scoring rules, configuration details and evidence boundaries. Read the full M3 research report →
Batch throughput and concurrent capacity are distinct measurements; neither specifies end-to-end live latency or report delivery time.
Make your next meeting count.
Start with your next conversation. Or bring a representative meeting and let us help you plan an enterprise pilot.
Arabic and English · Cloud and private deployment · By Lisan