Transcribe Audio (HuggingFace)

How it fits

Typical backends

Caveats

• Parameter interactionLeaving language empty lets the model auto-detect the spoken tongue; multilingual checkpoints honour any supported locale code, while English-only checkpoints ignore it.
• CostThe duration setting caps the internal chunk length in seconds; longer audio is chunked automatically, and very long inputs can exhaust GPU memory on large checkpoints.
• LimitationThere is no built-in voice-activity gating; pre-filter silence upstream with detect_voice_activity_silero_vad or expect occasional hallucinated text over silence.
• LatencyA warm-up pass runs at startup and any warm-up failure is tolerated rather than fatal, so the first real call can be slow if warm-up did not complete.
• I/O contractThe plain-text arm returns one String; the subtitle arm returns per-chunk start/end times. Empty input returns the empty result for the selected arm without running the model.
• State lifetimeThe model, chunk length, and device are fixed at startup, while language is re-read on every call so it can be changed without a restart.