Expand description
whisper-nvidia-cli — one-shot local NVIDIA speech-to-text for agents.
The process is born, executes one verb, prints compact JSON, and dies. Logs stay on stderr. Product environment variables are ignored.
Modules§
- audio
- Decode, resample to 16 kHz mono, inspect, and cut silence.
- catalog
- Persisted model catalog. Official SHA-1 seeds are written once.
- cli
- Clap surface. Side-effect verbs take a named target. No product env.
- commands
- Verb implementations. Each mutating verb echoes target_resolved.
- config
- Persistable knobs. Flags win, then this file, then compiled seeds.
- engine
- Whisper inference. The C motor never writes to stdout.
- error
- Semantic errors and exit codes for the agent contract.
- gpu
- NVIDIA device and CUDA toolkit discovery. No product env is read.
- i18n
- UI locale for stderr. JSON contract keys stay English.
- media_
helper - The external media helper: locating it, asking it, and outliving it safely.
- model
- Model cache, SHA-1 verify, and TLS-only download with retry.
- output
- Compact JSON on stdout. Logs never go here.
- paths
- XDG on Unix and known-folder paths on Windows.
- schema
- Published JSON contracts, one per payload the CLI can emit.
- signal
- Cooperative shutdown: first signal asks, second signal forces.
- surface
- Agent-native payload reduction. The binary cuts JSON before the model sees it.
- transcribe
- The transcription pipeline: argv to samples to text.
- verbs
- The verbs that answer about the product instead of listening to audio.
Functions§
- run
- Parse argv, run one verb, die with a semantic exit code.