voice_audio_multi/docs/models/canary_asr.md

Canary 180M Flash

NVIDIA Canary 180M Flash recognizes English, German, Spanish, and French speech. Source and target languages can be selected independently. NVIDIA documents translation quality for English-centric pairs. The weights use the CC-BY-4.0 license.

Field	Value
Family	canary_asr
Task	asr
Mode	offline
Output	Transcript text
Long-form audio	Fixed chunks, up to 40 seconds each
Timestamps	Not exposed

Install

python3 tools/model_manager_v2.py install canary_180m_flash_f32

For a smaller download, install canary_180m_flash_q8_0 instead and use canary-180m-flash-q8_0.gguf in the commands below. Each GGUF includes the configuration and tokenizers; no conversion is needed.

CLI

audiocpp_cli --task asr --family canary_asr \
  --model models/Canary-180M-Flash-GGUF/canary-180m-flash-f32.gguf \
  --backend cuda --audio speech.wav \
  --text-out transcript.txt --log

Long recordings are chunked automatically. Inputs must contain at least 20 ms of audio. To translate German speech to English:

audiocpp_cli --task asr --family canary_asr \
  --model models/Canary-180M-Flash-GGUF/canary-180m-flash-f32.gguf \
  --backend cuda --audio speech.wav \
  --request-option language=de --request-option target_language=en \
  --text-out translation.txt --log

Request Options (use with --request-option)

Option	Default	Meaning
language	en	Input language: en, de, es, or fr.
target_language	Input language	Output language: en, de, es, or fr; identical source and target select transcription.
punc	true	Punctuation and capitalization together. Legacy pnc remains accepted.
audio_chunk_mode	auto	auto or fixed splits long audio; none requires at most 40 seconds.
audio_chunk_duration_sec	40	Maximum chunk length, at most 40 seconds.
max_tokens	0	Generated tokens per chunk; zero derives a limit from encoder length.

Session Options (use with --session-option)

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

Use --backend cpu --threads 8 for CPU inference or --backend vulkan for Vulkan.