Transcribe Audio (Faster Whisper)

How it fits

Typical backends

Caveats

• Hard constraintmodel binds a CTranslate2 Faster-Whisper checkpoint bundle, loaded local-files-only; a directory without a converted checkpoint fails to start.
• CompatibilityThe compute_type must match the device — half/int8-half on GPU, int8 or float32 on CPU — and a mismatch errors inside the inference backend.
• Parameter interactionEnabling vad_filter runs voice-activity detection before decoding and substantially reduces hallucinated text over silence, at a small extra cost.
• AccuracyLeaving language empty enables auto-detection, which is slower and less reliable than pinning language to what is being spoken when it is known.
• I/O contractThe plain-text arm joins all segments into one String; the subtitle arms preserve per-segment start/end times, and one of them also attaches a per-segment confidence derived from the decoder log-probability. An empty AudioFrame returns the empty result for the selected arm without running inference.
• CompatibilityThe GPU build expects a matching CUDA and cuDNN runtime; when they are absent the backend can fall back to CPU, which is much slower.
• State lifetimeConfiguration is read once at startup; changing it requires a restart.