CAM++ zh-CN common — LiteRT
3D-Speaker CAM++ (campplus_cn_common, 192-d) converted PyTorch → LiteRT
via litert-torch, static input (1, 500, 80), fp32.
Front end (matches the validated ggml/rapidspeech pipeline): kaldi log-mel
fbank — 25 ms / 10 ms, povey window, 80 bins (20 Hz–8 kHz, kaldi mel scale),
dither 0, waveform ×2¹⁵, natural log — then global CMN over the utterance,
sliced into 500-frame windows; per-window embeddings are L2-normalized,
mean-pooled and L2-normalized again.
Parity gates (2026-07-24):
- tflite vs PyTorch source, 14-input battery (9 real-speech fbank windows + 5
random): worst element-wise max|Δ| 1.7e-05, worst cosine 0.99999994.
- Windowed tflite pipeline vs the rapidspeech ggml CAM++ reference embedding
(10 s real speech): cosine 0.9924 — cross-window speaker-linking
thresholds tuned on the ggml pipeline carry over.
Converted for
VoxSumDroid's
all-LiteRT migration.
