audible focus-control gate

Labnote 008 audible focus-control gate🔗

Status: Complete negative result. Post-hoc focus editing remained perceptibly artifacted and was rejected before listener judgments.

Labnote 004 showed that distinct audio bytes are not evidence of a meaningful prosodic contrast. Its first listener slice was stopped after three blinded judgments: delivery labels were never compiled, and focus was represented only as gain.

This bounded follow-up tests one mechanism before spending more human-review labor: can a compiler place clearly audible emphasis on either of two authored focus spans while holding sentence and voice fixed?

The v1 pilot passed its acoustic gate, but the operator heard an unnaturally high pitch and a brief doubled voice at the edited boundary. No judgments were submitted. The likely causes were a three-semitone shift with shifted formants and hard joins.

The frozen smooth-focus revision:

• deterministically selects two authored contrastive-emphasis pairs;
• evaluates both af_heart and am_adam;
• starts each comparison from the same neutral waveform;
• applies token-aligned +1.5-semitone, +3.5 dB emphasis with preserved formants,

quality pitch processing, smooth transient handling, and 15 ms boundary overlap;

• admits a comparison only when both directions clear energy, token-local F0,

waveform-separation, and duration-stability gates.

An initial v2 pass at +2.5 dB preserved pitch and duration but failed the frozen local energy threshold. V3 changed only gain to +3.5 dB and qualified all four trials for review.

The runner creates analysis artifacts only. prepare_review.py is a separate step that refuses to produce a blinded listener bundle unless every frozen trial passed.

python experiments/labnote_008/run_focus_gate.py \
  --source-run artifacts/labnote-004/oracle-synthesis \
  --output-dir artifacts/labnote-008/focus-gate

Result🔗

The v3 four-question bundle passed every numerical acoustic intake gate. On the pre-judgment listening check, however, the operator immediately recognized the same unnatural voice effect that motivated the later native-synthesis experiments. Smooth transient handling and preserved formants did not remove the perceptible post-hoc DSP artifact. The review was stopped with zero judgments submitted.

This is a negative gate result: audibility and bounded waveform metrics were insufficient proxies for natural, usable emphasis. Continuing the blinded preference test would have measured tolerance for an obvious synthesis artifact rather than focus recovery. Labnote 009 therefore removed token-local DSP entirely and tested native whole-utterance lexical stress instead.