Uni2005 ONNX
An ONNX export of the Uni2005 universal phoneme-recognition acoustic model
from Allosaurus. It is published by
KitsuMate as a separately downloadable runtime model for
ai.kitsumate.onnx.lipsync; it is not bundled in the Unity package.
Contents
| File | Purpose | SHA-256 |
|---|---|---|
onnx/model_fp32.onnx |
Float32 ONNX acoustic model | 517b07ccd197b8acb2c18c3f6ca19e1e8fd489b81853e386720917e01d43de1e |
onnx/model_fp16.onnx |
Float16 weights with float32 inputs and outputs | c41312a802cb91cbec62d168a295566c691dc69530a879480fd73f9a086343ac |
onnx/model_quantized.onnx |
Dynamic signed INT8 weights for ONNX Runtime CPU | b078a37029c2b9e074b0efe699b751d80fc70b325c1648295f290a239399b8af |
onnx/model_uint8.onnx |
Dynamic unsigned INT8 weights for ONNX Runtime CPU | 2a5fad69d2bf112b4ef2ca510ef17d8c2d134f5e7ab2b1ae505ac51ffaea0d04 |
phone.txt |
Universal phone inventory | b5fc99bae89d49f97212305a17901f73f4c80ab8671adf5540ec1af6b63d0037 |
vocab.json |
CTC token-to-phone mapping | 8dbf53951b7ac3379b4f6dcfdb9f6b659f588fac7d4fa04c1395876f0462a338 |
checksums.sha256 is the machine-readable form of these hashes.
Model contract
- Input:
mfcc,float32, shape[batch, time, 120](dynamic batch/time). - Output:
logits,float32, shape[batch, time, 230]. - Audio preprocessing: 8 kHz mono audio; 25 ms window; 10 ms hop; 40 MFCC coefficients with speaker CMVN, three-frame context windows, and a factor-3 subsample. The model does not perform resampling or MFCC extraction.
- Decode the output with CTC using
vocab.json.
The FP16 file stores reduced-precision weights while retaining float32 model
I/O. The two 8-bit files use ONNX Runtime dynamic quantization for the LSTM and
MatMul weights. They retain float32 inputs and outputs. model_quantized.onnx
is the default CPU-oriented variant; model_uint8.onnx is provided for
runtimes where unsigned weights perform better.
The variants were checked with ONNX 1.16.1, ONNX Converter Common 1.16.0, and ONNX Runtime 1.24.4 on CPU. Across deterministic inputs of 8, 31, and 100 frames, both variants preserved 100% frame-level argmax agreement with the float32 model and cosine similarity above 0.997. The Unity integration test uses the signed INT8 variant.
Run python tools/quantize.py with those versions to rebuild the variants.
The models have not been qualified for every runtime or Unity AI Inference
operator set.
Provenance and licence
The model is derived from the uni2005 model in
xinjli/allosaurus, recorded here at
commit d9f1adaf47d3b3765b41f4177da62a051516d636. See NOTICE for
attribution and the included GPL-3.0 licence for terms.
KitsuMate does not claim authorship of the original Allosaurus model and does not imply endorsement by its authors. The source model and these redistributed artifacts are covered by GPL-3.0 as confirmed for this publication.
Use with KitsuMate ONNX
Pin a Hugging Face commit revision and verify the published SHA-256 before placing the three runtime files in your application's model cache. Do not add them to a UPM package archive.