On July 7, 2026, Cohere released Cohere Transcribe Arabic — the most accurate open-source Arabic speech-to-text (ASR) model to date, purpose-built for the realities of Arabic in business and developer settings: dialect variation, and bilingual Arabic-English code-switching.
What Is Cohere Transcribe Arabic?
A 2-billion-parameter model built on an encoder-decoder architecture: a FastConformer encoder for acoustic modeling paired with a lightweight autoregressive Transformer decoder. The model is open-source on HuggingFace and accessible via the Cohere API.
Key Numbers
- Word Error Rate (WER): 25.87% on the Hugging Face Arabic ASR Leaderboard — best-in-class among open-source models
- Beats Meta's OmniASR-LLM-7B by 2.45 points (28.32%)
- Beats OpenAI's Whisper v3 Large by 11 full points (36.87%)
- Speed: RTFx 525 vs 146 for Whisper and 66 for OmniASR (3.6x faster than Whisper)
- Human evaluation: Preferred over Whisper in 95.8% of tests
Why This Matters for Arabic Users
Previous ASR models suffered from three problems:
- Dialect flattening: Models convert all dialects to Modern Standard Arabic, losing original meaning
- Weak code-switching: When speakers mix Arabic and English (common in business settings), models stumble
- Dialect neglect: Hundreds of millions speak Gulf, Egyptian, Levantine, and Maghrebi dialects that no model adequately covers
Cohere Transcribe Arabic addresses these through:
- Balanced dialect distribution: Careful resampling across dialect groups
- Real bilingual speech: Training data reflecting how Arabic professionals actually speak
- Diverse recording conditions: Different devices, environments, and noise levels
What Does This Mean for You?
For Developers and Businesses
- Free download: Model weights available on HuggingFace
- Self-hosting: Run on your own infrastructure using vLLM
- Cohere API: Available with free tier subject to rate limits
- Production speed: RTFx 525 means high-throughput audio processing
For Journalists and Researchers
- Transcription: Convert interviews and seminars in dialects to text with high accuracy
- Content archiving: Automatically transcribe podcasts and lectures
Known Limitations
- Requires a language tag (Arabic or English) for correct operation
- Does not support timestamps in the current version
- Does not support speaker diarization
- May hallucinate speech from low background noise — VAD recommended as a preprocessor
Quick Comparison
| Metric | Cohere Transcribe Arabic | Whisper v3 Large | OmniASR-LLM-7B |
|---|---|---|---|
| WER (lower better) | 25.87% | 36.87% | 28.32% |
| Size | 2B | ~1.5B | 7B |
| Speed (RTFx) | 525 | 146 | 66 |
| Open Source | Yes | Yes | Yes |
| Dialect support | Excellent | Weak | Moderate |
| Code-switching | Excellent | Weak | Weak |
How to Get Started
- Test the model on HuggingFace Spaces before downloading
- Download weights from HuggingFace:
CohereLabs/cohere-transcribe-arabic-07-2026 - Use the Cohere API with free tier for testing
- For production: use vLLM with a noise gate for best results
Primary source: Cohere blog on HuggingFace — Published: July 7, 2026
