instruction-conditioned contrastive focus

Labnote 013: instruction-conditioned contrastive focus🔗

Question🔗

Can reference-free natural-language instructions produce natural, directionally correct contrastive focus?

Motivation🔗

Prosody 012 generated eight mechanically valid F5-TTS outputs but stopped when every output failed the owner-local naturalness floor. This pilot removes recorded reference audio entirely instead of attempting to clean or tune around that failure. It tests a different control channel: natural-language delivery instructions applied to a fixed preset voice.

Frozen pilot🔗

• Qwen3-TTS 12 Hz 1.7B CustomVoice, English preset voice Aiden.
• The same two contrastive sentence pairs used in Prosody 012.
• Symmetric instructions that name the intended focus and exclude its competitor.
• Two fixed seeds per condition; eight outputs total.
• Whole-utterance generation without reference audio, cloning, fine-tuning, or DSP.
• Generated audio, reports, and review state remain private.

The run fails on missing conditions, invalid focus pairing, duplicate audio, or output outside the frozen 0.8–8.0 second integrity range. Mechanical passage does not admit blinded review. Every output must first sound natural and place contrastive emphasis on the instructed word. Only a fully passing set may become a blinded context-matching comparison.

Result🔗

The isolated Qwen3-TTS run generated all eight frozen outputs on an RTX 4070. Every waveform was distinct and passed the mechanical duration gate. In owner-local admission review, the outputs sounded substantially more natural than Prosody 012's reference-conditioned audio, clearing the qualitative naturalness floor.

The set nevertheless failed directional admission. The operator reported that emphasis appeared to fall on the wrong word, so the experiment stopped without a blinded review. Generated audio, the run report, and admission state remain private.

This separates synthesis quality from focus control: reference-free instruction control repaired the audible-quality failure but did not validate the frozen paired instruction template. Both instructions explicitly named the intended word and its competitor. A future experiment may test whether mentioning the competitor—especially inside a negative instruction—made that word salient, but that explanation was not isolated here.

Interpretation boundary🔗

A positive result would validate only this model, preset voice, instruction template, and tiny English corpus as worthy of expansion. It would not establish arbitrary prosody control, speaker independence, or a production context-to-speech compiler.