Transcribe Audio (Canary Qwen)

How it fits

Typical backends

Caveats

• LimitationThe default checkpoint is English only; non-English speech is not supported. For multilingual audio prefer transcribe_audio_faster_whisper.
• Hard constraintThe model id must point to a NeMo SALM checkpoint; any other identifier fails to load at startup.
• LimitationThe model was trained on clips up to about 40 seconds. Long audio is split into fixed windows and the partial transcripts are joined, so window boundaries can clip words and very long windows degrade accuracy.
• Parameter interactionA per-window decoded-token cap bounds latency; dense or fast speech in a window can be truncated when the cap is set too low.
• I/O contractEmpty input returns the empty string without invoking the model, and the single output is always a String.
• CostThe weights are several gigabytes and the toolkit pulls a heavy dependency tree; pre-seed the model cache for offline or air-gapped deployments. A GPU is expected for usable throughput.
• CompatibilityThe model is released under CC-BY-4.0; confirm compatibility with your deployment terms.
• State lifetimeConfiguration is read once at startup; changing it requires a restart.