Hugging Face's logo

parseh
/
aligner-ja

ONNX
Japanese
wav2vec2
ctc
forced-alignment
quantized
parseh
Model card Files
xet
 Community
aligner-ja / README.md
br1-ursino's picture
br1-ursino
Parseh aligner ja v1: ONNX int8 from reazon-research/japanese-wav2vec2-large-rs35kh@2162c9edb97f362f7623d6b3a2e417458e879bca
66eb48c verified
preview code
|
Raw History Blame Contribute Delete
2.44 kB
metadata 
license: apache-2.0
language: ja
library_name: onnx
tags:
  - onnx
  - wav2vec2
  - ctc
  - forced-alignment
  - quantized
  - parseh
base_model: reazon-research/japanese-wav2vec2-large-rs35kh
base_model_relation: quantized
datasets:
  - reazonspeech

Parseh Japanese forced-alignment network

This CTC network aligns known Japanese caption text to 16 kHz mono audio in Parseh. It is not a general-purpose speech-recognition model. Parseh runs it locally with NumPy and ONNX Runtime, then uses CTC Viterbi alignment for word spans or character spans.

Provenance and conversion

Derived from reazon-research/japanese-wav2vec2-large-rs35kh at 2162c9edb97f362f7623d6b3a2e417458e879bca, based on reazon-research/japanese-wav2vec2-large. It was converted to ONNX opset 17 and dynamically quantized to int8 (QUInt8, per-channel MatMul weights); no retraining was performed. Exact source and output hashes are in meta.json and SHA256SUMS.

Verification

  • ONNX Runtime CPU versions: 1.23.2, 1.30.0; dynamic 0.5 s and 60 s inputs passed.
  • fp32 ONNX/PyTorch log-softmax max absolute difference: 9.918212890625e-05; frame-argmax agreement: 1.0.
  • int8/fp32 frame-argmax agreement: 1.0; generated-speech span-start delta median/p95/max: 0.0 / 0.0 / 0.0 ms.
  • Generated 30-second CPU alignment cost: 4.144841 s; peak RSS: 2903474176 bytes.

espeak-ng generated audio was used only for pipeline monotonicity, not a real-speech accuracy claim. Unknown characters use a wildcard score to preserve a monotonic path.

Limitations

Input must be 16 kHz mono audio and known caption text. Noise, accents, text errors, vocabulary gaps, and quantization can affect timestamps. No accuracy guarantee is made. This language returns character spans because captions are not space-delimited.

Licence and attribution

The source tag is apache-2.0. LICENSE and NOTICE retain source, base-model, dataset attribution, and changes. No endorsement by original authors, Meta, Mozilla, or Parseh is implied.

Citation from the source card 

@misc{reazon-research-japanese-wav2vec2-large-rs35kh,
  title={japanese-wav2vec2-large-rs35kh},
  author={Sasaki, Yuta},
  url = {https://huggingface.co/reazon-research/japanese-wav2vec2-large-rs35kh},
  year = {2024}
}

Contact

Contact: development@parseh.io.