# Actual fast dialogue handoff

Final original Ukrainian speech: **65.596s,183words,25turns,4characters,167.39WPM**. This is **69.4% faster** by full-track word rate than iteration02(98.79WPM). No music or singing.

|Character|New voice|Heard delivery arc|
|---|---|---|
|Oksana|Kore|Guarded request, dry skepticism, self-directed choice, relieved photo callback|
|Iryna|Zephyr|Surprised concern, quick reassurance, practical explanation, precise proof, smiling invitation|
|Larysa|Aoede|Brief warm peer advice|
|Serhii|Charon|Bright practical family call|

Every submitted multisyllabic Ukrainian occurrence has contextual stress evidence and U+0301. Actual source takes and selected processed PCM were heard occurrence by occurrence. Five contiguous final-master windows verify complete words, joins, voices and emotional turns. Independent focused checks caught defects that target-guided green reviews missed: mis-stressed залиш/себе, a substituted adjective, and an authored залишилися stress. New requests repair those. D11 now says невеличкій and D23 uses the equivalent imperative лиши́. All failed versions remain available.

Provider Rapid Fire alone did not meet measured pace. The edit removes only measured silence(-45dB threshold), keeps short phrase breaks, then applies pitch-preserving atempo1.00–1.35 before a fresh acoustic review. Finalmaster simply concatenates the accepted48kmono16bit PCM turns; no later speed changes. Loudness: -18.17LUFS, peak -1.50dBTP.

Use assets/audio/audio-lock.json, words.json and parts.json with vocal-score.json. Every D01–D25 line carries character_id and mode. parts.json contains exact48k sample boundaries. D01/D02 pilot PCM remains unchanged. Word times come from actual-waveform faster-whisper estimates; short difficult turns use explicit accepted-text attention alignment after free ASR errors. This alignment does not certify stress: separate acoustic reviews do that.

Validation: pronunciation CLI preflight183occurrences/25parts; listening CLI accepted183occurrences; all5continuous windows pass; focused repairs pass. Model acoustic descriptions are qualitative, not a human linguistic certificate or measured phoneme timing. Actual Seedance lip synchronization remains a separate moving-picture gate.

[Provider fields](https://docs.kie.ai/market/google/gemini-3-1-flash-tts) · [Final listening](speech-listening.json) · [Measured pace](speech-measurements.json) · [Focused adjudications](pronunciation-adjudications.json) · [Rejected master](voice-rejections.json)