Reproducible synthetic conversational-audio rerun

Lab note: Reproducible synthetic conversational-audio rerun🔗

Run date: 2026-08-23 Status: Completed; automatic result negative/inconclusive

Question🔗

Does the negative Labnote 003 matched-versus-shuffled result persist when every audio clip is generated locally by a known, fingerprinted synthesizer and exact transcript fidelity is verified before downstream generation?

Correction carried forward🔗

Labnote 003 consumed opaque third-party audio whose synthesis model, voice, conditioning, prompt, seed, code, license, and human validation were not disclosed. That audio is quarantined. This rerun uses only the first 20 rows' text and labels from the pinned public dataset revision; it never reads the dataset audio field. Because the dataset license remains undisclosed, the external text snapshot and all generated artifacts stay outside source control.

Synthesis contract🔗

• hexgrad/Kokoro-82M, locally cached and package-fingerprinted;
• voice af_heart, speed 1.0, 24 kHz PCM;
• one complete synthesis call per rewritten_text;
• no segmentation, splicing, gain, pitch shifting, time stretching, or other DSP;
• WAV, source-snapshot, package, and generated-token fingerprints retained; and
• no claim that neutral Kokoro delivery realizes dataset emotion or intent labels.

The whole-utterance/no-DSP choice carries forward the naturalness lesson from Labnote 009. Labnote 010 showed that native stress annotation does not reliably create directional discourse prominence, so stress markup is deliberately excluded.

Preregistered gates🔗

Downstream LLM generation is forbidden unless:

1. independent ASR exactly matches normalized rewritten_text for all 20 clips; and
2. at least 18 of 20 deterministic next-row matched/shuffled cue pairs differ.

If either gate fails, the failure is the result. If both pass, reproduce Labnote 003's five conditions with the same system prompt, temperature 0.4, seed 42, and deterministic next-row cue derangement. The historical qwen2.5:3b model is no longer installed on the local GPU workstation; the rerun therefore preregisters the available qwen3:8b successor and records this as a model-generation change rather than claiming a strict replication. The label condition remains an unfair reference control and must not be interpreted as synthesized prosody.

Execution🔗

The experiment scheduler submitted each phase as a non-preemptive accelerator workload. The synthesis phase completed successfully; its internal identifier is withheld. Independent local faster-whisper small.en transcription then evaluated the frozen WAVs. The first downstream attempt, whose internal identifier is withheld, failed before retaining a response because the historical qwen2.5:3b model was not installed. The explicitly revised, checkpointed qwen3:8b run The corrected downstream attempt completed 100 responses with exit code zero.

Reproducibility facts:

• dataset revision: 3d49be1a3f15b3f58817ea86918584b5656f3a6e;
• source snapshot SHA-256: 7c9f7abb23649b7275ab020792b552ba25d8b21966fe25dddb63437d7235675c;
• source dataset audio used: no;
• Kokoro/Misaki: 0.9.4/0.9.4;
• synthesis voice/speed: af_heart/1.0;
• downstream model: qwen3:8b, local digest

500a1f067a9f782620b40bee6f7b0c89e17ae61f686b92c24933e4ca4b2b8b41.

Gate results🔗

• transcript fidelity: 20/20 exact after lowercase alphanumeric normalization;
• matched/shuffled cue separation: 20/20 bundles differed; and
• frozen WAV integrity: every synthesis hash was reverified before generation.

Both preregistered gates passed.

Automatic response diagnostics🔗

Mean token-set Jaccard similarity was 0.239 for A versus B, 0.299 for B versus C, 0.289 for B versus D, and 0.301 for C versus D. Matched and shuffled cue blocks both changed output relative to rewritten text. The matched condition was not uniquely closer to the text-only response and produced shorter responses than the shuffled condition. These are behavior-change diagnostics, not empathy or quality measures.

Result🔗

The provenance repair succeeded: this run has exact text/audio fidelity, immutable audio hashes, a known synthesizer, and genuinely distinct matched/shuffled cue blocks. The automatic downstream result nevertheless remains negative/inconclusive. Nothing in these proxies demonstrates that correctly matched neutral-synthesis cues help more than shuffled cues. Human review would be required for a conversational-quality claim, but the prior result no longer depends on unknowable third-party audio generation.