X-ASR zh-en punct zipformer2 on LiteRT
Conversion of csukuangfj2/sherpa-onnx-x-asr-zipformer-transducer-zh-en-punct-2026-06-03
(offline zipformer2 transducer) to LiteRT, via torch rebuild (vendored icefall) + ONNX
weight transplant. Shipped as the default ASR engine of
VoxSumDroid since v0.29.0 —
xasr_q8_octav.tflite (OCTAV dynamic-int8, 295 MB), measured 7.4× real-time on a
Samsung SM-A5360 CPU (2×A78), on par with the sherpa-onnx/ORT engine it replaced.
Signatures: enc_375/750/1500/3000 (x[1,T,80] normalized-sample povey fbank,
x_lens[1]) → (enc_out[1,T',512], out_lens[1]); decoder (y[1,2] i32, -1 pad) →
[1,512]; joiner (enc,dec) → logit[1,5000]. Greedy: blank 0, ≤1 sym/frame,
init [-1,0], suppress =4015 without context update; ts = frame×0.04 s.
Gates vs source ONNX: encoder max|Δ| 3.1e-06 (cos 1.0); decoder 1.1e-06; joiner
2.3e-05. End-to-end vs sherpa fp32 (repo test wavs): 4/5 byte-identical text,
timestamps ≤0.08 s. q8-octav CER is IDENTICAL to the fp32 tflite (quantization
adds no measurable error; min-max/MSE recipes measurably worse — use OCTAV).
xasr_encoder_torch.pt = transplanted PyTorch encoder weights (rebuild source);
xasr_punct_fp32.tflite = unquantized master.