voice_audio_multi/docs/models/cohere_asr.md

Cohere Transcribe

Cohere Transcribe is a multilingual speech recognition model released under Apache-2.0.

Field	Value
Family	cohere_asr
Task	asr
Mode	offline
Output	Transcript text
Long-form audio	Quiet-energy chunks, up to 35 seconds each
Languages	English, French, German, Spanish, Italian, Portuguese, Dutch, Polish, Greek, Arabic, Japanese, Chinese, Vietnamese, Korean

Install

python3 tools/model_manager_v2.py install cohere_transcribe_bf16

For a smaller download, install cohere_transcribe_q8_0 or cohere_transcribe_q4_0 instead and use the corresponding -q8_0.gguf or -q4_0.gguf filename below. Each GGUF includes the configuration and tokenizer; no conversion is needed.

CLI

audiocpp_cli --task asr --family cohere_asr \
  --model models/Cohere-Transcribe-GGUF/cohere-transcribe-03-2026-bf16.gguf \
  --backend cuda --audio speech.wav \
  --request-option language=en --text-out transcript.txt --log

Long recordings are split automatically at quiet points. Use --backend cpu --threads 8 for CPU inference or --backend vulkan for Vulkan.

To transcribe French without punctuation and capitalization:

audiocpp_cli --task asr --family cohere_asr \
  --model models/Cohere-Transcribe-GGUF/cohere-transcribe-03-2026-bf16.gguf \
  --backend cuda --audio french.wav \
  --request-option language=fr --request-option punc=false \
  --text-out transcript.txt --log

Request Options (use with --request-option)

Option	Default	Meaning
language	en	en, fr, de, es, it, pt, nl, pl, el, ar, ja, zh, vi, or ko.
punc	true	Request punctuation and capitalization together. Legacy pnc remains accepted.
audio_chunk_mode	auto	auto and quiet_energy use quiet boundaries; fixed uses equal-length chunks; none requires at most 35 seconds.
audio_chunk_duration_sec	35	Maximum audio chunk length, from 0.04 to 35 seconds.
max_tokens	256	Maximum generated tokens per chunk, up to 1014. Increase if generation reaches the limit before EOS.

Session Options (use with --session-option)

Option	Default	Meaning
cohere_asr.weight_type	native	Weight storage type override; native keeps the precision stored in the model file.

Audio chunks must contain at least 20 ms. Translation, diarization, timestamps, and live audio streaming are not supported.