audio-judge

name: audio-judge description: >- Judge generated Kokoro TTS audio clips with Gemini through llm-workflows. Use when the user asks whether synthesized speech sounds good, intelligible, whispery, corrupt, or acceptable, or when comparing PyTorch reference clips against Core ML / Swift pipeline output.

Audio Judge (Kokoro TTS)

Purpose

Fast perceptual gate for Kokoro TTS WAV clips: Gemini listens and returns a structured verdict. Objective waveform probes run first; listening is for needs_listening candidates and A/B comparisons.

Use When

• The user asks whether TTS output sounds acceptable, intelligible, whispery, or corrupt.
• Comparing PyTorch reference clips against Core ML / Swift pipeline output.
• A listening pack or bakeoff clip needs a smoke gate before human review.

Do Not Use When

• The task is numeric parity only — use coreml-validate first.
• The task is waveform health only with no listening need — use scripts/audio_quality_probe.py alone.
• The FFmpeg worker is healthy and you only need timings — use bakeoff.

Ground Rules

1. Read CLAUDE.md. This skill is for kokoro-coreml (Kokoro TTS), not Crossfade music generation.
2. Do not judge TTS by correlation/SNR alone. Perceptual listening is the gate when metrics and ears disagree.
3. Keep the boundary intact:
   • FFmpeg owns media prep and Gemini file upload on the primary path.
   • llm-workflows owns prompt execution and primary artifacts.
   • kokoro-coreml owns generation, context, and the TTS-native fallback script. The fallback is only for worker-down recovery — never replace listening with metrics.
4. Kokoro rubric (from scripts/audio_quality_probe.py and bakeoff practice): intelligible English speech, natural prosody, no whisper/static/clicks/ dropouts, spoken text matches the input prompt when provided.

Primary-path caveat: audio_judge_v1 was written for Crossfade music. Compensate with strong TTS framing in --prompt, --expected-style, and --context-file. Use the primary path whenever the worker is healthy.

Preflight (before the first run)

1. Objective gate (cheap rejects):

   uv run --no-sync python scripts/audio_quality_probe.py /path/to/clip.wav
   

   If quality_decision is reject_without_listening, fix generation before Gemini. Only send needs_listening or reference clips.

2. Primary path only: invoke with node, not pnpm run judge:audio -- (the -- is forwarded and breaks the script).

3. Primary path only: convert float32 WAVs to 16-bit PCM before upload — the FFmpeg worker probing stage rejects float32:

   afconvert -f WAVE -d LEI16 in.wav out.wav
   

   The fallback script auto-converts WAVs; step 3 is not required there.

4. Env: WORKFLOW_RUNTIME_TOKEN and FFMPEG_CLIENT_TOKEN in llm-workflows/.env (or set LLM_WORKFLOWS_ENV to that file). GEMINI_API_KEY is required for the fallback (same .env or env var). Optional: WORKFLOW_RUNTIME_BASE_URL, FFMPEG_BASE_URL.

Typical clip sources

Render fresh Config F clips:

uv run --no-sync python scripts/bakeoff_listen.py --keys 3s

Bakeoff text: scripts/bakeoff_harness.py (BAKEOFF_INPUTS). Pass matching text as --prompt.

Command (primary path)

From the llm-workflows checkout (sibling of this repo by default):

node scripts/run-audio-judge.mjs \
  --clip pytorch=/absolute/path/to/pytorch_3s.wav \
  --clip coreml=/absolute/path/to/config_f_3s.wav \
  --baseline-label pytorch \
  --prompt "The quick brown fox jumps over the dog." \
  --expected-style "clear intelligible English speech, natural prosody, no whispering or static, no clicks or dropouts, 24 kHz mono" \
  --context-file /absolute/path/to/kokoro_context.md

Options: --context-file (neutral facts only), --no-wait, --resume.

Context-file safety: the entire file is sent to Gemini. Never point it at .env, credentials, or unrelated secrets.

Context file template

# Kokoro TTS audio judge context

- Task: perceptual TTS quality gate (not music).
- Sample rate: 24000 Hz mono 16-bit PCM.
- Voice: af_heart, speed: 1.0.
- Input key: 3s bakeoff sentence.
- Candidate: Config F (Swift + Core ML, compute_units=ALL).
- Baseline: PyTorch eager reference from audio-parity exports.
- Objective probe: needs_listening; rms=4600, active32=78.8%, zcr=9.1%.
- Question: Is Core ML output acceptably close to PyTorch for shipping?

Fallback (worker down only)

uv run --no-sync python scripts/gemini_audio_judge_direct.py \
  --clip pytorch=/absolute/path/pytorch_3s.wav \
  --clip coreml=/absolute/path/config_f_3s.wav \
  --baseline-label pytorch \
  --prompt "The quick brown fox jumps over the dog." \
  --probe-first \
  --context-file /path/to/kokoro_context.md \
  --output outputs/audio-judge-fallback/<slug>.json

Fallback behavior:

• Reads GEMINI_API_KEY from env or LLM_WORKFLOWS_ENV (default sibling ../llm-workflows/.env).
• Auto-converts WAVs to 16-bit PCM; gain-matches non-baseline clips to baseline RMS (disable with --no-gain-match).
• --probe-first runs audio_quality_probe before Gemini.
• Clips must be .wav/.mp3 audio files under size limits — not arbitrary local files.
• --output must be inside this repo; writes a structured envelope with verdict, clip paths, and schema hints.

Clips upload directly to Gemini (no FFmpeg/R2). Mark fallback reports in notes. Switch back to the primary path once the worker is healthy.

Known failure modes

Output

Primary: llm-workflows/outputs/audio-judge/<slug>/<timestamp>/ — result.json, report.md, checkpoints.

Fallback: --output JSON envelope under outputs/audio-judge-fallback/.

Capture conclusions with write-notes (link artifacts, clip paths, config, primary vs fallback).

Judging protocol

1. Neutral context only — facts, never expected outcome.
2. Always include controls: PyTorch reference and a known-bad clip when available. Discard runs that misrank controls.
3. One run = one vote; require 2–3 agreeing lineups for ship/no-ship.
4. Cross-check audio_quality_probe, waveform alignment, bakeoff parity JSON.
5. No ASR/WER gate unless the user explicitly asks.

Interpreting results

• overallVerdict: "pass" is a smoke gate, not bakeoff readiness.
• Failed Core ML + passing PyTorch → Core ML/runtime/glue, not input text.
• Failed PyTorch baseline → fix reference path before ANE debugging.

Primary path (result.json): use comparison.iphoneAcceptablyCloseToMlx as “Core ML acceptably close to PyTorch” when labels are coreml vs pytorch.

Fallback path (verdict envelope): use verdict.comparisons["<label>_vs_<baseline>"].same_quality_class — not iphoneAcceptablyCloseToMlx. The envelope documents both field names.

Related skills

• bakeoff — render Config F clips and timings.
• coreml-validate — numeric parity before listening.
• write-notes — institutional memory after verdicts.