speech-enhancement-models

k2-fsa/sherpa-onnxspeech-enhancement-modelsMar 10, 2025by csukuangfj

AI Summary

This release documents and provides model details for GTCRN and DPDFNet speech enhancement models, including parameter counts and intended use cases for various hardware tiers.

Key Highlights

Documentation for GTCRN models
Documentation for DPDFNet models (16kHz and 48kHz variants) with performance metrics
Comparison of model sizes and intended use cases

New Features

Model documentation and availability (GTCRN, DPDFNet)

Full Release Notes

## gtcrn

`gtrcn_simple.onnx` is from https://github.com/Xiaobin-Rong/gtcrn

`speech_with_noise.wav` is from https://modelscope.cn/models/iic/speech_zipenhancer_ans_multiloss_16k_base/file/view/master?fileName=examples%252Fspeech_with_noise.wav&status=0

## DPDFNet

[DPDFNet](https://github.com/ceva-ip/DPDFNet) models are from
https://huggingface.co/Ceva-IP/DPDFNet/tree/main/onnx

### 16 kHz models

| Model | Params [M] | MACs [G] | TFLite Size [MB] | ONNX Size [MB] | Intended Use |
| --- | :---: | :---: | :---: | :---: | --- |
| dpdfnet_baseline | 2.31 | 0.36 | 8.5 | 8.5 | Fastest / lowest resource usage |
| dpdfnet2 | 2.49 | 1.35 | 10.7 | 9.9 | Real-time / embedded devices |
| dpdfnet4 | 2.84 | 2.36 | 12.9 | 11.2 | Balanced performance |
| dpdfnet8 | 3.54 | 4.37 | 17.2 | 14.1 | Best enhancement quality |

### 48 kHz model

| Model | Params [M] | MACs [G] | TFLite Size [MB] | ONNX Size [MB] | Intended Use |
| --- | :---: | :---: | :---: | :---: | --- |
| dpdfnet2_48khz_hr | 2.58 | 2.42 | 11.6 | 10.3 | High-resolution 48 kHz audio |