Real-time speech enhancement for recordings, live streams, and edge devices.
Pretrained 8, 16, and 48 kHz models for the CLI, Python API, and stateful streaming.
Live demo · Audio examples · Pretrained models · Paper
| Model | Params [M] | MACs [G] | TFLite Size [MB] | ONNX Size [MB] |
|---|---|---|---|---|
| dpdfnet2_8khz | 2.51 | 1.29 | 10.5 | 9.7 |
| dpdfnet8_8khz | 3.56 | 3.99 | 16.5 | 13.8 |
| Model | Params [M] | MACs [G] | TFLite Size [MB] | ONNX Size [MB] |
|---|---|---|---|---|
| baseline | 2.31 | 0.36 | 8.5 | 8.3 |
| dpdfnet2 | 2.49 | 1.35 | 10.7 | 9.7 |
| dpdfnet4 | 2.84 | 2.36 | 12.9 | 11.1 |
| dpdfnet8 | 3.54 | 4.37 | 17.2 | 13.9 |
| Model | Params [M] | MACs [G] | TFLite Size [MB] | ONNX Size [MB] |
|---|---|---|---|---|
| dpdfnet2_48khz_hr | 2.58 | 2.42 | 11.6 | 10.0 |
| dpdfnet8_48khz_hr | 3.63 | 7.17 | 18.7 | 14.2 |
For CPU-only ONNX inference using the packaged CLI and Python API:
pip install dpdfnet
# Enhance one file dpdfnet enhance noisy.wav enhanced.wav --model dpdfnet4 --attn-limit-db 12 # Enhance a directory (uses all CPU cores by default) dpdfnet enhance-dir ./noisy_wavs ./enhanced_wavs --model dpdfnet2 --attn-limit-db 12 # Enhance a directory with a fixed worker count dpdfnet enhance-dir ./noisy_wavs ./enhanced_wavs --model dpdfnet2 --workers 4 --attn-limit-db 12 # Download models dpdfnet download dpdfnet download dpdfnet8 dpdfnet download dpdfnet2_8khz dpdfnet download dpdfnet4 --force
import soundfile as sf import dpdfnet # In-memory enhancement: audio, sr = sf.read("noisy.wav") enhanced = dpdfnet.enhance(audio, sample_rate=sr, model="dpdfnet4", attn_limit_db=12) sf.write("enhanced.wav", enhanced, sr) # Enhance one file: out_path = dpdfnet.enhance_file("noisy.wav", model="dpdfnet2", attn_limit_db=12) print(out_path) # Model listing: for row in dpdfnet.available_models(): print(row["name"], row["ready"], row["cached"]) # Download models: dpdfnet.download() # All models dpdfnet.download("dpdfnet4") # Specific model dpdfnet.download("dpdfnet2_8khz")
Install sounddevice (not included in dpdfnet dependencies):
pip install sounddevice
StreamEnhancer processes audio chunk-by-chunk, preserving RNN state across
calls. Any chunk size works; enhanced samples are returned as soon as enough
data has accumulated for the first model frame (20 ms).
import numpy as np import sounddevice as sd import dpdfnet INPUT_SR = 48000 # Use one model hop (10 ms) as the block size so process() returns # exactly one hop's worth of enhanced audio on every callback. BLOCK_SIZE = int(INPUT_SR * 0.010) # 480 samples at 48 kHz enhancer = dpdfnet.StreamEnhancer(model="dpdfnet2_48khz_hr") def callback(indata, outdata, frames, time, status): mono_in = indata[:, 0] if indata.ndim > 1 else indata.ravel() enhanced = enhancer.process(mono_in, sample_rate=INPUT_SR) n = min(len(enhanced), frames) outdata[:n, 0] = enhanced[:n] if n < frames: outdata[n:] = 0.0 # silence while the first window accumulates with sd.Stream( samplerate=INPUT_SR, blocksize=BLOCK_SIZE, channels=1, dtype="float32", callback=callback, ): print("Enhancing microphone input - press Ctrl+C to stop") try: while True: sd.sleep(100) except KeyboardInterrupt: pass # Optional: drain the final partial window at the end of a recording tail = enhancer.flush()
Notes:
Latency - the first enhanced output arrives after one full model window (~20 ms) has been buffered. All subsequent blocks are returned with ~10 ms additional delay.
Sample rate -StreamEnhancerresamples internally. Pass your device's native rate assample_rate; the return value is at the same rate.
Block size - usingBLOCK_SIZE = int(SR * 0.010)(one model hop) gives one enhanced block per callback. Other sizes also work but may produce empty returns while the buffer fills.
Multiple streams - create a separateStreamEnhancerper stream. Callenhancer.reset()between independent audio segments to clear RNN state.
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txtModel files are not bundled in this repository.
Download PyTorch checkpoints, TFLite, and ONNX models from Hugging Face:
pip install -U "huggingface_hub[cli]" # create target dirs mkdir -p model_zoo/{checkpoints,onnx,tflite} # PyTorch checkpoints (HF path: checkpoints/* -> local: model_zoo/checkpoints/*) hf download Ceva-IP/DPDFNet \ --include "checkpoints/*.pth" \ --local-dir model_zoo \ # ONNX models (HF path: onnx/* -> local: model_zoo/onnx/*) hf download Ceva-IP/DPDFNet \ --include "onnx/*.onnx" \ --local-dir model_zoo \ # TFLite models (HF path: *.tflite at repo root -> local: model_zoo/tflite/*) hf download Ceva-IP/DPDFNet \ --include "*.tflite" \ --local-dir model_zoo/tflite \
Put one or more *.wav files in ./noisy_wavs, then choose one:
python -m tflite_model.infer_dpdfnet_tflite \ --noisy_dir ./noisy_wavs \ --enhanced_dir ./enhanced_wavs \ --model_name dpdfnet4 \ --workers 5 \ --attn-limit-db 12
python -m onnx_model.infer_dpdfnet_onnx \ --noisy_dir ./noisy_wavs \ --enhanced_dir ./enhanced_wavs \ --model_name dpdfnet4 \ --workers 5 \ --attn-limit-db 12
To export ONNX models from checkpoints, use the exporter that matches the sample-rate family.
Set --dprnn-num-blocks to match the checkpoint variant; for 8 kHz and 48 kHz HR exports,
this also determines the ONNX metadata profile:
# 8 kHz python -m onnx_model.export_dpdfnet_8khz_to_onnx \ --checkpoint model_zoo/checkpoints/dpdfnet2_8khz.pth \ --output model_zoo/onnx/dpdfnet2_8khz.onnx \ --dprnn-num-blocks 2 # 16 kHz python -m onnx_model.export_dpdfnet_to_onnx \ --checkpoint model_zoo/checkpoints/dpdfnet4.pth \ --output model_zoo/onnx/dpdfnet4.onnx \ --dprnn-num-blocks 4 # 48 kHz HR python -m onnx_model.export_dpdfnet_48khz_hr_to_onnx \ --checkpoint model_zoo/checkpoints/dpdfnet8_48khz_hr.pth \ --output model_zoo/onnx/dpdfnet8_48khz_hr.onnx \ --dprnn-num-blocks 8
Enhanced files are written as:
<original_stem>_<model_name>.wav
- Project page with examples: https://ceva-ip.github.io/DPDFNet/
- Gradio application: https://huggingface.co/spaces/Ceva-IP/DPDFNetDemo
- Hugging Face model hub: https://huggingface.co/Ceva-IP/DPDFNet
- Evaluation dataset used in the paper: https://huggingface.co/datasets/Ceva-IP/DPDFNet_EvalSet
Run:
python -m real_time_demo
How it works:
- Captures microphone audio in streaming hops.
- Enhances each hop frame-by-frame with ONNX.
- Displays live noisy vs enhanced spectrograms.
- Allows you to control the noise‑reduction level during playback:
0for the raw stream and1for the fully enhanced stream. - Enables the use of AGC during playback.
To change model, edit MODEL_NAME near the top of real_time_demo.py.
- Run the Hugging Face download commands from the
Run From Sourcesection. - Confirm files are in:
model_zoo/tflite/model_zoo/onnx/model_zoo/checkpoints/
- Both offline scripts scan only the exact folder given by
--noisy_dir(non-recursive). - Ensure input files use
.wavextension.
- Check microphone permissions and default input/output device settings.
- Install host audio dependencies for
sounddevice(PortAudio packages on your OS).
- Ensure Qt dependencies from
requirements.txtinstalled successfully. - On headless servers, run offline enhancement instead.
- Run from repo root and use module form exactly as documented (
python -m ...). - Activate your virtual environment before running commands.
- Try smaller models (
baseline,dpdfnet2). - Benchmark ONNX runtime using
python -m onnx_model.infer_dpdfnet_onnx ...and compare RTF.
To compute intrusive and non-intrusive metrics on our DPDFNet EvalSet, we use the tools listed below. For aggregate quality reporting, we rely on PRISM, the scale‑normalized composite metric introduced in the DPDFNet paper.
We provide a dedicated script, pesq_stoi_sisnr_calc.py, which computes PESQ, STOI, and SI-SNR for paired reference and enhanced audio. The script includes a built-in auto-alignment step that corrects small start-time offsets and drift between the reference and the enhanced signals before scoring, to ensure fair comparisons.
- DNSMOS (P.835 & P.808) - We use the official DNSMOS local inference script from the DNS Challenge repository:
dnsmos_local.py. Please follow their installation and model download instructions in that project before running. - NISQA v2 - We use the official NISQA project: https://github.com/gabrielmittag/NISQA. Refer to their README for environment setup, pretrained model weights, and inference commands (e.g., running
nisqa_predict.pyon a folder of WAVs).
Explore applications, plugins, libraries and research projects built with DPDFNet.
Using DPDFNet in your project? Open an issue or submit a pull request to add it to the list.
@article{rika2025dpdfnet, title = {DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN}, author = {Rika, Daniel and Sapir, Nino and Gus, Ido}, year = {2025}, }
Apache License 2.0. See LICENSE.