Cohere Transcribe

#1 on the Open ASR Leaderboard — 5.42 average WER across 9 benchmarks vs. Whisper large-v3’s 7.44, at ~3× the throughput. Benchmark details.

Supported languages: English, German, French, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Arabic, Vietnamese, Chinese, Japanese, Korean

This model has no language detection, so language must be set: null transcribes as English, and any code outside the list above is refused rather than transcribed. hyprwhspr status shows which language is in effect.

Memory: 4 GB VRAM (bfloat16), or 8 GB RAM on CPU (float32).

Setup

Cohere Transcribe is a gated model on HuggingFace — you must accept the license before downloading.

1. Accept the license agreement at: huggingface.co/CohereLabs/cohere-transcribe-03-2026
2. Generate a read token at: huggingface.co/settings/tokens
3. Run hyprwhspr setup and select Cohere — you will be prompted for your token

The model (~4 GB) is downloaded during setup. Your token is securely stored locally in ~/.config/hyprwhspr/credentials.json and never shared.

Configuration

{
    "transcription_backend": "cohere-transcribe",
    "cohere_transcribe_device": "auto",      // auto | cuda | cpu
    "cohere_transcribe_dtype": "bfloat16",   // bfloat16 (GPU default) | float32 (CPU)
    "cohere_transcribe_compile": false       // torch.compile for faster throughput (adds warmup on first call)
}

On GPU the model always runs in bfloat16 (its native precision). A float16 setting is accepted but coerced to bfloat16 — true float16 (IEEE half) overflows the model’s attention mask, so it is not usable here. Use float32 (default on CPU) to force full precision.

Model stored in: ~/.cache/huggingface/hub/models--CohereLabs--cohere-transcribe-03-2026/