Cohere releases 2B parameter Arabic speech recognition model with 25.9% average WER

Cohere and Cohere Labs released Cohere Transcribe Arabic, a 2B parameter automatic speech recognition (ASR) model optimized for Arabic dialects and Arabic-English code-switching. The model tops the Open Universal Arabic ASR Leaderboard with a 25.87% average word error rate (WER) and 11.80% character error rate (CER).

Architecture and capabilities

The model uses a Conformer-based encoder-decoder architecture. A large Conformer encoder extracts acoustic representations from audio input, followed by a lightweight Transformer decoder for text generation. Audio inputs are automatically resampled to 16kHz and converted to log-Mel spectrograms during preprocessing.

Cohere Transcribe Arabic supports both Arabic and English transcription. For long-form audio, the feature extractor automatically segments waveforms into chunks, with the processor reassembling transcriptions using audio chunk indices.

Benchmark performance

According to Cohere, the model achieves the following WER scores across Arabic ASR benchmarks:

• SADA: 37.47%
• Common Voice: 5.82%
• MASC clean: 19.60%
• MASC noisy: 27.07%
• MGB-2: 15.54%
• Casablanca: 49.71%

The model outperforms larger competitors including OmniASR LLM 7B (28.32% average WER), Qwen3-Omni 30B (30.71% average WER), and Whisper Large v3 (36.86% average WER).

Integration and deployment

The model is natively supported in Transformers 5.4.0+ and available under Apache 2.0 license. For production deployment, Cohere recommends using vLLM 0.19.0 for serving.

Cohere provides code examples for single-file transcription, long-form audio processing, and vLLM server setup. The model requires the processor to specify language ("ar" or "en") during inference.

Limitations

Cohere notes the model performs best with single-language audio and lacks automatic language detection. Performance on code-switched audio is inconsistent. The model does not support timestamps or speaker diarization. Like most audio encoder-decoder models, it may hallucinate transcriptions for non-speech audio without a voice activity detection preprocessor.

What this means

This release provides the first competitive open-source ASR model specifically optimized for Arabic dialects, an underserved segment in speech recognition. The 2B parameter count makes it deployable on consumer hardware while matching or exceeding models 15x larger. The Apache 2.0 license enables commercial use without restrictions, potentially accelerating Arabic speech applications in regions where proprietary APIs have limited availability or high costs.