New tool + generated artefacts so the on-device voice spinner can now
hot-swap between all 8 voices — previously only Damien's prefix/suffix
were present in the model dir, and the tablet fell back to him
regardless of selection.
scripts/export_voice_prefix_suffix.py runs Qwen3TTS's voice-clone
path under a forward hook, captures the first prefill call's 1024-dim
talker input embeddings, aborts the rest of the (very slow on CPU)
decode via a sentinel exception, and slices out the first 9 vectors
as <name>_voice_prefix.bin and the last 2 as <name>_voice_suffix.bin.
Validated against the shipped damien_voice_prefix.bin: using
damien_15s_24k.wav as the reference audio, max|diff| = 0, so the
extraction matches the original tooling bit-for-bit.
Generated and adb-pushed to
/data/local/tmp/kazeia/models/qwen3-tts-npu/:
amir / didier / elodie / jerome / richard / sid / zelda
(+ re-generated damien from the canonical 15s_24k reference)
Qwen3TtsEngine.setVoice (already wired) reads <voice>_voice_prefix.bin
/ <voice>_voice_suffix.bin by basename, so voice changes now take
effect from the next synthesized segment with no app restart.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
db281002d9