multi-utterance prosody candidate pool

Labnote 017: multi-utterance prosody candidate pool🔗

Question🔗

Can a small multi-utterance candidate pool supply natural repeated readings for a held-out fingerprint-retrieval test?

Frozen pool🔗

• Four previously unused conversational utterances.
• Three broad candidate recipes per utterance: calm matter-of-fact delivery, weary

exasperation, and amused disbelief.

• One shared seed, Qwen3-TTS 1.7B CustomVoice, the Aiden speaker, and whole-utterance

synthesis without post-processing.

• Exactly twelve private clips; recipe identity is hidden during annotation.
• Reuse the Prosody 015 listener ontology and 1–5 naturalness, intensity, and confidence

scales, recording what each clip actually expresses rather than recipe compliance.

This is corpus construction, not a test of instruction following. Retrieval evaluation will be frozen only after annotation establishes which labels the candidate pool really contains; no held-out claim is made here.

Result🔗

All twelve outputs passed mechanical integrity and were blindly annotated. Six met the 4/5 naturalness floor, spanning all four new utterances. Every natural clip carried a disbelief label; none was neutral-only. The nominally matter-of-fact recipe produced two natural clips, but both were heard as expressive. Weary exasperation produced three natural clips, and amused disbelief produced one.

The pool successfully expanded the natural expressive gallery across new lexical content, but it cannot support a balanced held-out neutral-versus-expressive retrieval test. A minimal follow-up should generate one uninstructed control per utterance and stop after four clips rather than repeating the full pool.