1.0 KiB
speech.transcription@1
speech.transcription@1 accepts canonical base64 PCM WAV, at most 16 MiB and 120 seconds. Language may be explicit or automatic. The response contains transcription, optional detected language, ordered timestamp segments where supported, duration and processing metadata.
Audio is transient: it is excluded from database records, audit events, metrics and logs; it is deleted from Redis after completion; inference is offline; and no cloud fallback exists.
The M9 runtime adapter uses the exact verified Whisper Large v3 Turbo artifact with AutoModelForSpeechSeq2Seq. WAV decoding uses the Python standard library, bounded PCM validation and deterministic resampling. Repository code and runtime network egress remain disabled.
ASR evaluation stores fixture digests and ground truth rather than audio. Typed metrics are WER, latency, real-time factor and peak VRAM. Dutch and English are required where locally generated or licensed fixtures are available; absent fixtures are reported rather than synthesized into false evidence.