Commit Graph

4 Commits

Author SHA1 Message Date
Kazeia Team 38c0e9874a Disable C++ pipeline (QNN non-deterministic), keep Java RTF 1.8
Root cause found: QNN HTP level=1 compilation is not bitwise
deterministic. Two loads of the same .pte produce slightly different
hidden states → audible trembling in decoded speech.

Java pipeline uses single QNN instance → no trembling, validated quality.
C++ pipeline code preserved for future use when QNN context caching
is fixed (would make both loads use same compiled graph).

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 11:42:49 +02:00
Kazeia Team 8e536094df Fix C++ pipeline eos/pad + disable for quality (keep Java default)
- Fixed trailing embed handling (use pre-computed as-is)
- Added eos/pad embed params to nativeRun
- Improved C++ PRNG for sampling
- Disabled native pipeline: slight quality regression vs Java
  (two separate QNN instances give different numerical results)
- Java pipeline (RTF 1.8) kept as default for validated quality

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 10:53:19 +02:00
Kazeia Team 3b01302cfb Fix missing eos/pad embeddings in native C++ pipeline
The native pipeline was adding zeros after trailing text tokens
instead of tts_eos_embed then tts_pad_embed. This caused the model
to mispronounce final words (e.g. "développement" → "devopment").

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 10:35:05 +02:00
Kazeia Team 393ce79eb5 Native C++ pipeline: RTF 1.4 (was 3.6 in Java)
Full talker+CP autoregressive loop in C++ via JNI.
Talker 20ms/step, CP 44ms/step, total 6.6s for 4.64s audio.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-09 10:09:32 +02:00