Cohere has officially released Cohere Transcribe Arabic, a highly specialized, open-source Automatic Speech Recognition (ASR) model tailored to tackle the complexities of Arabic dialects.
The new 2-billion-parameter model is available globally under the permissive Apache 2.0 license, launching directly on Hugging Face alongside a public hosted API tier.
1. Conquering the “Dialect Flattening” Problem
Arabic is spoken by more than 300 million people worldwide, fractured across roughly 30 recognized dialect varieties. Traditional speech-to-text models systematically flatten regional accents—collapsing localized colloquial speech back into formal Modern Standard Arabic (MSA) or hallucinating words entirely.
Cohere Transcribe Arabic addresses this head-on by focusing training on three tough audio environments:
- Dialect Faithfulness: Preserving regional colloquialisms accurately without forcing them into a formal written register.
- Bilingual Code-Switching: Seamlessly transcribing rapid, conversational switching between Arabic and English, which is highly common in professional and everyday speech.
- Noisy Environments: Maintaining high transcription accuracy across specialized corporate vocabularies, video meetings, and poor acoustic settings.
2. Setting a New SOTA on the Leaderboard
Despite its compact 2B-parameter footprint, the model outpaces systems more than three times its size, officially claiming the number one spot on the Hugging Face Arabic ASR Leaderboard:
| Model | Size (Parameters) | Average Word Error Rate (WER) | Performance vs. Cohere |
| Cohere Transcribe Arabic | 2 Billion | 25.87% | — |
| Meta OmniASR-LLM | 7 Billion | 28.32% | ↓ 2.45 points lower accuracy |
| OpenAI Whisper Large V3 | ~1.5 Billion | 36.86% | ↓ 11.0 points lower accuracy |
Human Preference: In head-to-head blind testing, human native speakers preferred transcripts generated by Cohere Transcribe Arabic over OpenAI’s Whisper Large V3 in 95.8% of evaluated tests, praising its handling of regional phrasing and heavy code-switching.
3. Architecture and Local Deployment Options
The model is built on an efficient Conformer-based encoder-decoder design. For maximum developer flexibility, it can be implemented across multiple deployment pathways:
Plaintext
[ COHERE TRANSCRIBE ARABIC INFRASTRUCTURE ]
├── Hugging Face ──► Download raw weights or run the Web Space Demo (Up to 25MB files)
├── Cohere API ──► Free developer tier with rate limits; fully managed cloud inference
└── Model Vault ──► Enterprise-grade, dedicated instance-hour hosting for private data
Because it doesn’t carry massive hardware requirements, the 2B model is light enough to be quantized into INT8 or Q4 formats and run entirely locally on commodity CPUs or edge devices via ONNX Runtime. This makes it an ideal option for enterprise workflows that process sensitive, highly confidential audio records without risking data leaks via external cloud servers.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
