|
|
--- |
|
|
license: apache-2.0 |
|
|
pipeline_tag: audio-to-audio |
|
|
tags: |
|
|
- speech_enhancement |
|
|
- noise_suppression |
|
|
- real_time |
|
|
--- |
|
|
|
|
|
|
|
|
# DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN |
|
|
|
|
|
DPDFNet is a family of causal, single-channel speech enhancement models for real-time noise suppression in challenging everyday environments. It extends the DeepFilterNet2 enhancement framework by inserting Dual-Path RNN (DPRNN) blocks into the encoder, strengthening long-range temporal and cross-band modeling while preserving a compact, streaming-friendly design. |
|
|
|
|
|
This repository provides four TensorFlow Lite (TFLite) models optimized for mobile and edge deployment: |
|
|
|
|
|
* `baseline.tflite` |
|
|
* `dpdfnet2.tflite` |
|
|
* `dpdfnet4.tflite` |
|
|
* `dpdfnet8.tflite` |
|
|
|
|
|
--- |
|
|
|
|
|
## Key Features |
|
|
|
|
|
* Causal and low-latency: Designed for streaming use cases such as telephony, conferencing, and embedded devices. |
|
|
* Dual-Path RNN integration: Improves temporal context and frequency-domain interactions for more robust enhancement in difficult noise conditions. |
|
|
* Scalable family: Choose baseline or dpdfnet2/4/8 to balance quality vs. compute. |
|
|
* Edge deployment focus: Demonstrated on Ceva NeuPro Nano NPUs in the accompanying work. |
|
|
|
|
|
--- |
|
|
|
|
|
## Model Variants and Footprint |
|
|
|
|
|
| Model | Params [M] | MACs [G] | TFLite Size [MB] | |
|
|
| --------- | ---------: | -------: | ---------------: | |
|
|
| Baseline | 2.31 | 0.36 | 8.5 | |
|
|
| DPDFNet-2 | 2.49 | 1.35 | 10.7 | |
|
|
| DPDFNet-4 | 2.84 | 2.36 | 12.9 | |
|
|
| DPDFNet-8 | 3.54 | 4.37 | 17.2 | |
|
|
|
|
|
--- |
|
|
|
|
|
## Intended Use |
|
|
|
|
|
Primary task: Real-time, single-channel speech enhancement (noise suppression). |
|
|
|
|
|
Deployment targets: Mobile devices, embedded NPUs, and edge platforms. |
|
|
|
|
|
Input and Output: |
|
|
|
|
|
* Input: 16 kHz mono noisy speech waveform |
|
|
* Output: 16 kHz mono enhanced speech waveform |
|
|
|
|
|
Typical applications: |
|
|
|
|
|
* Voice calls and VoIP |
|
|
* Video conferencing |
|
|
* Always-on voice interfaces |
|
|
* Wearables, earbuds, and embedded audio devices |
|
|
|
|
|
--- |
|
|
|
|
|
## Inference |
|
|
|
|
|
This repo includes a reference script for running the TFLite models on WAV files using streaming-style, frame-by-frame inference: `run_tflite.py`. |
|
|
|
|
|
### Setup |
|
|
|
|
|
Install dependencies: |
|
|
|
|
|
```bash |
|
|
pip install numpy soundfile librosa tqdm |
|
|
pip install tflite-runtime |
|
|
``` |
|
|
|
|
|
### Model placement |
|
|
|
|
|
By default, the script loads models from: |
|
|
|
|
|
* `./<model_name>.tflite` |
|
|
|
|
|
Create the folder and place the `.tflite` files there (or edit `TFLITE_DIR` in the script to match your layout). |
|
|
|
|
|
### Run enhancement on a folder of WAVs |
|
|
|
|
|
The script processes `*.wav` files non-recursively and writes enhanced outputs as 16-bit PCM WAVs: |
|
|
|
|
|
```bash |
|
|
python run_tflite.py --noisy_dir /path/to/noisy_wavs --enhanced_dir /path/to/out --model_name dpdfnet8 |
|
|
``` |
|
|
|
|
|
Available `--model_name` options: `baseline`, `dpdfnet2`, `dpdfnet4`, `dpdfnet8`. |
|
|
|
|
|
--- |
|
|
|
|
|
## Training Data |
|
|
|
|
|
The models were trained using a mixture of public speech and noise datasets, including DNS4 (downsampled), MLS, MUSAN, and FSD50K. |
|
|
|
|
|
--- |
|
|
|
|
|
## Citation |
|
|
|
|
|
If you use these models, please cite: |
|
|
|
|
|
```bibtex |
|
|
@article{rika2025dpdfnet, |
|
|
title = {DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN}, |
|
|
author = {Rika, Daniel and Sapir, Nino and Gus, Ido}, |
|
|
year = {2025} |
|
|
} |
|
|
``` |
|
|
|
|
|
--- |
|
|
|
|
|
## License |
|
|
|
|
|
Apache-2.0 |
|
|
|
