rs35kh CTC FP16 ONNX for UtaLog
This repository hosts an ONNX derivative of
reazon-research/japanese-wav2vec2-base-rs35kh
for UtaLog's local Japanese lyrics timing module.
It is not an official ReazonSpeech release and is not endorsed by the upstream authors.
Artifact
- File:
rs35kh_ctc_fp16.onnx - Size:
203351696bytes - SHA-256:
5c6b2b5f14445e1dd12cfc055e96bbc47dd87d4e874e43636d212bc201280d48 - Inputs:
input_values, float32, shape[batch, samples], 16 kHz mono audio - Outputs:
logits, shape[batch, frames, 3003] - ONNX opset: 17
Source and changes
- Upstream:
reazon-research/japanese-wav2vec2-base-rs35kh - Pinned upstream revision:
46afc596052b612293c8db256b3a69447a2f57dc - Upstream license: Apache-2.0
- Conversion entry point:
tools/alignment/convert_rs35kh_to_onnx.pyin UtaLog - Modification: export
Wav2Vec2ForCTCto ONNX, keep the convolutional feature extractor in FP32, convert the encoder suffix to FP16, and normalizeReshape.allowzerofor DirectML compatibility.
The exact hosted bytes are a recovered development artifact. A later clean conversion passed model validation but was not byte-identical, so this repository does not claim byte-for-byte reproducibility. UtaLog pins this repository's immutable Hub commit and verifies both byte size and SHA-256 before use.
Intended use and limitations
This graph supplies CTC acoustic emissions for alignment against user-provided Japanese lyrics. It is not sufficient by itself to produce an LRC file: UtaLog generates readings, tokenizes the known lyrics, searches monotonic candidates, preserves locked ranges, and exposes low-confidence or unmatched lines for review.
The upstream model card reports poor long-form ASR transcription results. UtaLog therefore uses bounded audio chunks and local-anchor/section alignment rather than decoding a whole song as unrestricted ASR.
License
The upstream model is distributed under Apache-2.0. This converted derivative is distributed under the same license. See LICENSE-2.0.txt and UtaLog's apps/desktop/resources/alignment/THIRD_PARTY_NOTICES.md for attribution and change information.