README.md

---
license: apache-2.0
pipeline_tag: audio-to-audio
tags:
  - speech_enhancement
  - noise_suppression
  - real_time
  - fullband
---


# 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 TensorFlow Lite (TFLite) models optimized for mobile and edge deployment:

**16kHz models**
* `baseline.tflite`
* `dpdfnet2.tflite`
* `dpdfnet4.tflite`
* `dpdfnet8.tflite`

**48kHz model**
* `dpdfnet2_48khz_hr.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.
* Fullband option: A dedicated 48kHz model is provided for fullband enhancement.

---

## Model Variants and Footprint

### 16kHz models

| 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        |

### 48kHz model

| Model        | Params [M] | MACs [G] | TFLite Size [MB] |
| ------------ | :--------: | :------: | :--------------: |
| DPDFNet-2 HR |    2.58    |   2.42   |      11.6        |

---

## Intended Use

Primary task: Real-time, single-channel speech enhancement (noise suppression).

Deployment targets: Mobile devices, embedded NPUs, and edge platforms.

Input and Output:

* **16kHz models**
  * Input: 16kHz mono noisy speech waveform
  * Output: 16kHz mono enhanced speech waveform
* **48kHz model**
  * Input: 48kHz mono noisy speech waveform
  * Output: 48kHz 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 inference script for running the TFLite models on WAV files using streaming-style, frame-by-frame inference: `run_tflite.py`.

> **Note:** When using `dpdfnet2_48khz_hr`, the inference script automatically switches to the 48kHz processing pipeline.

### 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`, `dpdfnet2_48khz_hr`.

---

## 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