Cohere Transcribe Arabic
About this model
Section titled “About this model”Cohere Transcribe Arabic is a finetuned version of Cohere Transcribe that is optimized for Arabic audio inputs. Like Cohere Transcribe, it is a 2B parameters dedicated audio-in, text-out, automatic speech recognition (ASR) model available open source. The model is designed to accurately capture the diversity of dialects, accents, and acoustic conditions among Arabic speakers during transcription.
Model details
Section titled “Model details”- Input: Audio waveform
- Output: Text
- Model name:
cohere-transcribe-arabic-07-2026 - Languages covered: Arabic, English
- Multidialectal support: Yes
- Code-switching support: Yes
- Maximum file size: 25MB
- License: Apache 2.0
Availability
Section titled “Availability”You can access Cohere Transcribe Arabic via our API for free, low-setup experimentation subject to rate limits.
For production deployment without rate limits, provision a dedicated Model Vault. This enables low-latency, private cloud inference without having to manage infrastructure. Pricing is calculated per hour-instance, with discounted plans for longer-term commitments. Contact our team to discuss your requirements.
Strengths
Section titled “Strengths”Cohere Transcribe Arabic achieves state-of-the-art transcription accuracy for Arabic speech. These results are robust to diverse or variable audio inputs, including: bilingual dialogues (Arabic-English); regional dialects and phrasing, and enterprise-specific vocabulary. Like its parent model, Cohere Transcribe Arabic has been optimized for high-throughput, production inference, and is at the frontier for far-field (where the speaker is not close to the receiver) transcription tasks.
Key Limitations
Section titled “Key Limitations”- Timestamps: The model does not output timestamps alongside transcripts.
- Speaker diarization: The model does not automatically identify individual speakers in a multispeaker audio file.
Model architecture
Section titled “Model architecture”Cohere Transcribe is built on a speech-optimized Transformer variant: a Conformer. Input audio waveforms are converted into a Mel spectrogram and then processed by a Conformer encoder that holds the majority of the model’s parameters. The encoder’s representations are then passed to a lightweight Transformer decoder that generates text tokens. Cohere Transcribe is trained using standard supervised cross-entropy.