SenseVoiceSmall-ONNX-CPU-INT8

An unchanged mirror of Alibaba's official SenseVoiceSmall ONNX CPU model. Uses dynamic activation quantization and INT8 matrix multiplication.

官方 CPU 动态 INT8 模型原文件镜像,权重未作修改。

Usage

Tested with onnxruntime-node / onnxruntime-common 1.30.0 on Windows x64. Create and release sessions inside a Node Worker:

import * as ort from 'onnxruntime-node';

const session = await ort.InferenceSession.create('model_quant.onnx', {
  executionProviders: ['cpu'],
  intraOpNumThreads: 4,
  interOpNumThreads: 1,
});
// Run with the tensors below; consume and dispose tensors after use.
await session.release();

Use default graph optimizations. The model consumes features from the official 16 kHz / 80-bin fbank / LFR m=7,n=6 frontend, normalized with am.mvn. Decode CTC using tokens.txt and valid output lengths.

Tensor Type Shape
speech float32 [batch, frames, 560]
speech_lengths int32 [batch]
language int32 [batch]
textnorm int32 [batch]
ctc_logits float32 [batch, frames+4, 25055]
encoder_out_lens int32 [batch]

Language IDs: auto 0, zh 3, en 4, yue 7, ja 11, ko 12. Text normalization: with ITN 14, without ITN 15. Encoded lengths include four control positions.

Source and integrity

Original SenseVoiceSmall: Alibaba / FunAudioLLM / FunASR. The upstream ONNX repository declares Apache-2.0; see LICENSE.

model_quant.onnx is copied byte-for-byte from official revision 4e95991ddb34c70e9b94f026d62ac82a9d941ae1.

Size: 241,216,270 bytes. SHA-256: 21dc965f689a78d1604717bf561e40d5a236087c85a95584567835750549e822

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for neavo/SenseVoiceSmall-ONNX-CPU-INT8

Quantized
(24)
this model