cruizesync/zipa-small-crctc-500k-w8

ZIPA-CR small (500k) — weight-only int8 ONNX

A converted copy of ZIPA-CR small (anyspeech/zipa-small-crctc-500k, revision dfe87c97662606fa6db90ad06f627d4363314744), a multilingual phone recognition model. Used by Cruize Music's hum and sing search.

Credit

ZIPA phone recognition model by Jian Zhu et al. (2025), CC BY 4.0 — https://github.com/lingjzhu/zipa

Paper: ZIPA: A family of efficient speech models for multilingual phone recognition, ACL 2025 — https://aclanthology.org/2025.acl-long.961/

What was changed

This file was made from the authors' FP32 model.onnx (SHA-256 ec62e6f5…280d96). Every MatMul weight is stored as int8, with one symmetric scale per output column, and turned back to float by a DequantizeLinear right before its MatMul. Activations stay float. Nothing else was changed: same inputs (x, x_lens), same outputs, same tokens.txt.

Why: in the published dynamic-int8 model, activation quantisation rounds slightly differently on different ONNX Runtime engines, such as WebAssembly in a browser versus native ARM or x86. With activations in float, this file gives the same results on every engine: log-probabilities agree within about 2e-5. It is about the same size (70.5 MB).

To rebuild it, run python convert_w8.py model.onnx model.w8.onnx (onnx 1.17). It produces this exact file.

Files

File	SHA-256	Size
model.w8.onnx	32fbb4f79d84d02e991a913b3a161b3c40a39908de472cf9cccd17d8188f5535	70,481,402 bytes
tokens.txt	f8e042a0c9130532b22d03ec7cae2f75a23fbec70c450c31a8efb51787b2b8fe	769 bytes
convert_w8.py	—	the conversion script

Licence

The model weights are licensed under CC BY 4.0, the same as the original. The changes are described above. The original authors do not endorse this copy.