Audio Utils Runner
The Audio Utils runner (audio-utils-runner) processes captured or uploaded audio: normalize to PCM WAV, emit JSON metrics, run speech-to-text, and combine metrics with transcripts for voice analysis.
Node template selection
| Template | Purpose | Limitations | Details |
|---|---|---|---|
convert-to-pcm | Decode WAV/AIFF/AIFC uploads and emit PCM WAV plus a per-file acceptance report. | Non-audio blobs are skipped (node does not fail the whole batch); resample/bit-depth changes need numpy/librosa in the image. | convert-to-pcm |
audio-metrics | Compute JSON audio metrics (duration, levels, and related fields) per accepted file. | Tolerates mixed non-audio upstream blobs; marks invalid items with parseError. | audio-metrics |
generate-transcript | Speech-to-text (Whisper, faster-whisper, whisperx, wav2vec) keyed by content hash. | Heavy CPU/GPU and model dependencies; needs audio-metrics and PCM map inputs; air-gap sites must bundle models. | generate-transcript |
voice-analysis | Merge metrics and transcript signals into a combined JSON analysis document. | Expects wired upstream parents; evolving output schema; not a dashboard publisher. | voice-analysis |
Typical pipeline
Related
Memory and CPU
- The runner declares
cpuLimit: 2000mandmemoryLimit: 4096Mi, and itsdeploymentblock asks the scheduler for exactly one pod (minimumPodsandmaximumPods1) with acpuRequestof750m. Transcription is one long task per session, so one pod with two cores finishes sooner than several pods held to a fraction of a core. In total it reserves and may use less CPU than the previous default of several 750m pods. - An operator can change the pod range and requests per environment with the
scheduler.limits.audio-utils-runner.*platform configuration keys. - Set
computeType: "int8"ongenerate-transcriptto cut its processing time by about 40% on CPU.
Opens your browser's print dialog. Choose "Save as PDF" as the destination.