grahama1970/analyze-chatterbox-emotions

analyze-chatterbox-emotions 是做什麼的？

Evaluate a generated Chatterbox render as a voice-quality artifact, not as a claim about the speaker's real emotion.

Use this when a Chatterbox or Persona Dream render needs proof that emotional tags, pauses, pace, and audio quality behaved as requested.

Contract

Input:

./run.sh analyze --audio path/to/render.wav \
  --expected-text "optional source text" \
  --target-label reassuring \
  --target-arousal 0.35 \
  --target-valence 0.4 \
  --out /tmp/chatterbox_voice_eval.json \
  --report /tmp/chatterbox_voice_eval.md

Output JSON schema: analyze_chatterbox_emotions.voice_eval.v1.

The evaluator reports:

• affect: target label plus acoustic arousal/valence proxies and a compatibility score.
• prosody: duration, estimated speech rate, RMS/loudness proxy, F0 estimates when available.
• pauses: silence ratio, detected pause spans, and planned-vs-measured pause comparison when a Chatterbox render plan or conversation turn is supplied.
• quality: clipping, peak amplitude, discontinuity flags, and low-signal warnings.
• intelligibility: expected-text word count and optional transcript fields; this skill does not silently invent ASR.

Optional inputs

• --render-plan path.json: Chatterbox render_plan.json, Persona Dream journal receipt, dynamic conversation receipt, or a conversation turn JSON carrying chatterbox_pause_plan.
• --transcript "text": caller-supplied ASR transcript. If omitted, transcript similarity is null and listed as a non-claim.

Verdicts

• pass: no hard technical failures and score >= 75.
• review: score 50-74 or important missing evidence such as no ASR transcript.
• fail: hard technical issue, unusable audio, severe clipping, missing audio, or score < 50.

The score is a weighted evaluation signal, not ground truth:

0.35 affect_match + 0.25 intelligibility + 0.20 prosody + 0.20 technical_quality

Important boundaries

• Do not say the person or persona "is sad/angry/happy". Say the waveform and available classifiers/proxies are compatible or incompatible with the requested target.
• Speech-emotion recognition models are optional signals. If unavailable, this skill still emits acoustic proxy metrics and records emotion_classifier.available=false.
• Exact pauses are verified from the waveform when possible; pause_after_ms in a render plan is only the requested pause.

Maintenance

Run:

./sanity.sh
../agentic-evals/run.sh run fixtures/agentic_eval.json --output /tmp/analyze-chatterbox-emotions-agentic-eval.json