Model card and the original license

46f99ac verified about 14 hours ago

20.6 kB

---
license: mit
pipeline_tag: audio-to-audio
library_name: onnx
tags:
- onnx
- moonbox
- onnxruntime
- deepfilternet
- speech-enhancement
- denoising
- noise-suppression
- 48khz
- real-time
---
# DeepFilterNet3 ONNX

ONNX builds of [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet), made by Moonbox for on-device use with ONNX Runtime (CPU, CUDA, DirectML). Each
precision is one folder; what the graphs take and return, and the signal processing that goes around them, is
under *How to run*.

## Files

The last column is the benchmark on the VoiceBank-DEMAND test set (824 noisy utterances; details under *How to
run*): each precision's speech-quality scores, and how far its output is from the original PyTorch model's.

| Precision | File | Size | sha256 | Benchmark |
|---|---|---|---|---|
| fp16 | [`fp16/config.ini`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/config.ini) | 2.1 kB | `415eb925d449` | PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) |
| fp16 | [`fp16/df_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/df_dec.onnx) | 1.7 MB | `a3380e4164a9` | PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) |
| fp16 | [`fp16/enc.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/enc.onnx) | 1.0 MB | `694385ef6060` | PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) |
| fp16 | [`fp16/erb_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/erb_dec.onnx) | 1.7 MB | `72126330232d` | PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) |
| int8 | [`int8/config.ini`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/config.ini) | 2.1 kB | `415eb925d449` | PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) |
| int8 | [`int8/df_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/df_dec.onnx) | 0.9 MB | `7b729627da56` | PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) |
| int8 | [`int8/enc.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/enc.onnx) | 0.5 MB | `41ac70016f7f` | PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) |
| int8 | [`int8/erb_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/erb_dec.onnx) | 0.9 MB | `394ba8bd8f85` | PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) |

## How to run

DeepFilterNet3 (real-time speech enhancement at 48 kHz: ERB-band gains plus deep filtering of the low band) as the
three ONNX graphs of the official export (`models/DeepFilterNet3_onnx.tar.gz` in the GitHub repository): `enc.onnx`
(encoder), `erb_dec.onnx` (ERB-gain decoder) and `df_dec.onnx` (deep-filter decoder), with the model's
`config.ini`. The time axis is free, so a whole file runs in one call. Our `fp16/` and `int8/` are made from those
graphs; the fp32 originals are in the release itself.

The graphs are the neural network only: the STFT, the ERB and complex features, applying the gains and the deep
filter, and the ISTFT go around them. The simplest way to get all of that exactly right is the official
`deepfilternet` package with its three networks swapped for these graphs (how this repository was benchmarked):

```python
import numpy as np, onnxruntime as ort, soundfile as sf, torch
from df.enhance import enhance, init_df                     # pip install deepfilternet==0.5.6 (torch / torchaudio 2.1)
model, df_state, _ = init_df()                              # the official DeepFilterNet3 checkpoint
prov = ["CPUExecutionProvider"]                              # int8: CPU; fp16: ["CUDAExecutionProvider"]
class Graph(torch.nn.Module):
    def __init__(self, path, first_only):
        super().__init__()
        self.s, self.first_only = ort.InferenceSession(path, providers=prov), first_only
        self.names = [i.name for i in self.s.get_inputs()]
    def forward(self, *xs):
        out = [torch.from_numpy(o) for o in self.s.run(None, {n: x.numpy() for n, x in zip(self.names, xs)})]
        return out[0] if self.first_only else tuple(out)
model.enc, model.erb_dec, model.df_dec = (Graph("int8/enc.onnx", False), Graph("int8/erb_dec.onnx", True),
                                          Graph("int8/df_dec.onnx", True))
noisy, sr = sf.read("noisy_48k.wav", dtype="float32")       # 48 kHz mono
with torch.no_grad():
    clean = enhance(model, df_state, torch.from_numpy(noisy)[None])
sf.write("enhanced.wav", clean.squeeze(0).numpy(), 48000)
```

Without Python, the DeepFilterNet project's Rust crate (`libDF`, with its `tract` runtime) implements the same
feature pipeline around these three graphs.

**Precisions.** `int8/`: weights int8 per output channel behind DequantizeLinear (every Conv, grouped linear and GRU
weight), activations fp32; 2.2 MB, runs on the CPU provider, the one to use for a real-time CPU denoiser. `fp16/`:
weights and math in fp16, inputs and outputs fp32; for GPU providers (ONNX Runtime's CPU provider has no fp16 GRU).
No int4 and no bf16 (no 4-bit or bf16 GRU kernels in ONNX Runtime).

**Benchmark.** VoiceBank-DEMAND test set (Valentini et al.: 824 utterances by two unseen speakers, 48 kHz, the set
the DeepFilterNet papers report), enhanced in one call per file with the post-filter off (the package default).
PESQ is wide-band (ITU-T P.862.2, at 16 kHz), STOI at 48 kHz, SI-SDR against the clean speech; "vs original" is
the SI-SDR of each output against the PyTorch model's output (mean / worst file). Real-time factor with 4 CPU threads
on an Intel laptop CPU (int8) or the RTX 3080 Laptop GPU (fp16), whole pipeline.

| | PESQ | STOI | SI-SDR (dB) | vs original (dB) | Real-time factor |
|---|---|---|---|---|---|
| noisy input | 1.97 | 0.921 | 8.39 | - | - |
| original (PyTorch fp32, CPU) | 3.163 | 0.9452 | 18.91 | - | 0.043 |
| official ONNX fp32 (CPU) | 3.163 | 0.9452 | 18.91 | 135 / 130 | 0.017 |
| `fp16/` (CUDA) | 3.163 | 0.9452 | 18.91 | 65.2 / 55.2 | 0.047 |
| `int8/` (CPU) | 3.159 | 0.9448 | 18.89 | 44.4 / 32.2 | 0.018 |

## How these were made

From [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet) at revision `7b14b895fe`, starting from an fp32 ONNX graph (where it came from is said above).
fp16: onnxconverter-common with inputs and outputs kept fp32, or, where the network's activations leave fp16's
range, fp16 weights with fp32 math (said above when so). int8: weight-only, every Conv / MatMul / GRU weight int8
with a per-output-channel scale behind DequantizeLinear and activations in fp32, so it runs on the CPU provider
without a calibration set. int4: ONNX Runtime's MatMulNBits (weight-only, block 32) on the MatMul weights, the rest
int8 as above, where the graph has enough MatMul weights to matter. No bf16: ONNX Runtime has no bf16 convolution or
GRU kernels. Every graph was benchmarked before upload; the `sha256` column above is what the files must hash to.

## License

MIT/Apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this
repository, and it governs your use of these files. The original model card follows unchanged.

---

# Original model card

# DeepFilterNet
A Low Complexity Speech Enhancement Framework for Full-Band Audio (48kHz) using on Deep Filtering.

![deepfilternet3](https://user-images.githubusercontent.com/16517898/225623209-a54fea75-ca00-404c-a394-c91d2d1146d2.svg)

For PipeWire integration as a virtual noise suppression microphone look [here](https://github.com/Rikorose/DeepFilterNet/blob/main/ladspa/README.md).

### Demo

https://github.com/Rikorose/DeepFilterNet/assets/16517898/79679fd7-de73-4c22-948c-891927c7d2ca

To run the demo (linux only) use:
```bash
cargo +nightly run -p df-demo --features ui --bin df-demo --release
```

### News

- New DeepFilterNet Demo: *DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement*
  - Paper: https://arxiv.org/abs/2305.08227
  - Video: https://youtu.be/EO7n96YwnyE

- New Multi-Frame Filtering Paper: *Deep Multi-Frame Filtering for Hearing Aids*
  - Paper: https://arxiv.org/abs/2305.08225

- Real-time version and a LADSPA plugin
  - [Pre-compiled binary](#deep-filter), no python dependencies. Usage: `deep-filter audio-file.wav`
  - [LADSPA plugin](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/ladspa/) with pipewire filter-chain integration for real-time noise reduction on your mic.

- DeepFilterNet2 Paper: *DeepFilterNet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio*
  - Paper: https://arxiv.org/abs/2205.05474
  - Samples: https://rikorose.github.io/DeepFilterNet2-Samples/
  - Demo: https://huggingface.co/spaces/hshr/DeepFilterNet2

- Original DeepFilterNet Paper: *DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering*
  - Paper: https://arxiv.org/abs/2110.05588
  - Samples: https://rikorose.github.io/DeepFilterNet-Samples/
  - Demo: https://huggingface.co/spaces/hshr/DeepFilterNet
  - Video Lecture: https://youtu.be/it90gBqkY6k

## Usage

### deep-filter

Download a pre-compiled deep-filter binary from the [release page](https://github.com/Rikorose/DeepFilterNet/releases/).
You can use `deep-filter` to suppress noise in noisy .wav audio files. Currently, only wav files with a sampling rate of 48kHz are supported.

```bash
USAGE:
    deep-filter [OPTIONS] [FILES]...
ARGS:
    <FILES>...
OPTIONS:
    -D, --compensate-delay
            Compensate delay of STFT and model lookahead
    -h, --help
            Print help information
    -m, --model <MODEL>
            Path to model tar.gz. Defaults to DeepFilterNet2.
    -o, --out-dir <OUT_DIR>
            [default: out]
    --pf
            Enable postfilter
    -v, --verbose
            Logging verbosity
    -V, --version
            Print version information
```

If you want to use the pytorch backend e.g. for GPU processing, see further below for the Python usage.

### DeepFilterNet Framework

This framework supports Linux, MacOS and Windows. Training is only tested under Linux. The framework is structured as follows:

* `libDF` contains Rust code used for data loading and augmentation.
* `DeepFilterNet` contains DeepFilterNet code training, evaluation and visualization as well as pretrained model weights.
* `pyDF` contains a Python wrapper of libDF STFT/ISTFT processing loop.
* `pyDF-data` contains a Python wrapper of libDF dataset functionality and provides a pytorch data loader.
* `ladspa` contains a LADSPA plugin for real-time noise suppression.
* `models` contains pretrained for usage in DeepFilterNet (Python) or libDF/deep-filter (Rust)

### DeepFilterNet Python: PyPI

Install the DeepFilterNet Python wheel via pip:
```bash
# Install cpu/cuda pytorch (>=1.9) dependency from pytorch.org, e.g.:
pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
# Install DeepFilterNet
pip install deepfilternet
# Or install DeepFilterNet including data loading functionality for training (Linux only)
pip install deepfilternet[train]
```

To enhance noisy audio files using DeepFilterNet run
```bash
# Specify an output directory with --output-dir [OUTPUT_DIR]
deepFilter path/to/noisy_audio.wav
```

### Manual Installation

Install cargo via [rustup](https://rustup.rs/). Usage of a `conda` or `virtualenv` recommended.
Please read the comments and only execute the commands that you need.

Installation of python dependencies and libDF:
```bash
cd path/to/DeepFilterNet/  # cd into repository
# Recommended: Install or activate a python env
# Mandatory: Install cpu/cuda pytorch (>=1.8) dependency from pytorch.org, e.g.:
pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
# Install build dependencies used to compile libdf and DeepFilterNet python wheels
pip install maturin poetry
#  Install remaining DeepFilterNet python dependencies
# *Option A:* Install DeepFilterNet python wheel globally within your environment. Do this if you want use
# this repos as is, and don't want to develop within this repository.
poetry -C DeepFilterNet install -E train -E eval
# *Option B:* If you want to develop within this repo, install only dependencies and work with the repository version
poetry -C DeepFilterNet install -E train -E eval --no-root
export PYTHONPATH=$PWD/DeepFilterNet # And set the python path correctly
# Build and install libdf python package required for enhance.py
maturin develop --release -m pyDF/Cargo.toml
# *Optional*: Install libdfdata python package with dataset and dataloading functionality for training
# Required build dependency: HDF5 headers (e.g. ubuntu: libhdf5-dev)
maturin develop --release -m pyDF-data/Cargo.toml
# If you have troubles with hdf5 you may try to build and link hdf5 statically:
maturin develop --release --features hdf5-static -m pyDF-data/Cargo.toml
```

### Use DeepFilterNet from command line

To enhance noisy audio files using DeepFilterNet run
```bash
$ python DeepFilterNet/df/enhance.py --help
usage: enhance.py [-h] [--model-base-dir MODEL_BASE_DIR] [--pf] [--output-dir OUTPUT_DIR] [--log-level LOG_LEVEL] [--compensate-delay]
                  noisy_audio_files [noisy_audio_files ...]
positional arguments:
  noisy_audio_files     List of noise files to mix with the clean speech file.
optional arguments:
  -h, --help            show this help message and exit
  --model-base-dir MODEL_BASE_DIR, -m MODEL_BASE_DIR
                        Model directory containing checkpoints and config.
                        To load a pretrained model, you may just provide the model name, e.g. `DeepFilterNet`.
                        By default, the pretrained DeepFilterNet2 model is loaded.
  --pf                  Post-filter that slightly over-attenuates very noisy sections.
  --output-dir OUTPUT_DIR, -o OUTPUT_DIR
                        Directory in which the enhanced audio files will be stored.
  --log-level LOG_LEVEL
                        Logger verbosity. Can be one of (debug, info, error, none)
  --compensate-delay, -D
                        Add some paddig to compensate the delay introduced by the real-time STFT/ISTFT implementation.
# Enhance audio with original DeepFilterNet
python DeepFilterNet/df/enhance.py -m DeepFilterNet path/to/noisy_audio.wav
# Enhance audio with DeepFilterNet2
python DeepFilterNet/df/enhance.py -m DeepFilterNet2 path/to/noisy_audio.wav
```

### Use DeepFilterNet within your Python script

```py
from df import enhance, init_df
model, df_state, _ = init_df()  # Load default model
enhanced_audio = enhance(model, df_state, noisy_audio)
```

See [here](https://github.com/Rikorose/DeepFilterNet/blob/main/scripts/external_usage.py) for a full example.

### Training

The entry point is `DeepFilterNet/df/train.py`. It expects a data directory containing HDF5 dataset
as well as a dataset configuration json file.

So, you first need to create your datasets in HDF5 format. Each dataset typically only
holds training, validation, or test set of noise, speech or RIRs.
```py
# Install additional dependencies for dataset creation
pip install h5py librosa soundfile
# Go to DeepFilterNet python package
cd path/to/DeepFilterNet/DeepFilterNet
# Prepare text file (e.g. called training_set.txt) containing paths to .wav files
#
# usage: prepare_data.py [-h] [--num_workers NUM_WORKERS] [--max_freq MAX_FREQ] [--sr SR] [--dtype DTYPE]
#                        [--codec CODEC] [--mono] [--compression COMPRESSION]
#                        type audio_files hdf5_db
#
# where:
#   type: One of `speech`, `noise`, `rir`
#   audio_files: Text file containing paths to audio files to include in the dataset
#   hdf5_db: Output HDF5 dataset.
python df/scripts/prepare_data.py --sr 48000 speech training_set.txt TRAIN_SET_SPEECH.hdf5
```
All datasets should be made available in one dataset folder for the train script.

The dataset configuration file should contain 3 entries: "train", "valid", "test". Each of those
contains a list of datasets (e.g. a speech, noise and a RIR dataset). You can use multiple speech
or noise dataset. Optionally, a sampling factor may be specified that can be used to over/under-sample
the dataset. Say, you have a specific dataset with transient noises and want to increase the amount
of non-stationary noises by oversampling. In most cases you want to set this factor to 1.

<details>
  <summary>Dataset config example:</summary>
<p>
  
`dataset.cfg`

```json
{
  "train": [
    [
      "TRAIN_SET_SPEECH.hdf5",
      1.0
    ],
    [
      "TRAIN_SET_NOISE.hdf5",
      1.0
    ],
    [
      "TRAIN_SET_RIR.hdf5",
      1.0
    ]
  ],
  "valid": [
    [
      "VALID_SET_SPEECH.hdf5",
      1.0
    ],
    [
      "VALID_SET_NOISE.hdf5",
      1.0
    ],
    [
      "VALID_SET_RIR.hdf5",
      1.0
    ]
  ],
  "test": [
    [
      "TEST_SET_SPEECH.hdf5",
      1.0
    ],
    [
      "TEST_SET_NOISE.hdf5",
      1.0
    ],
    [
      "TEST_SET_RIR.hdf5",
      1.0
    ]
  ]
}
```

</p>
</details>

Finally, start the training script. The training script may create a model `base_dir` if not
existing used for logging, some audio samples, model checkpoints, and config. If no config file is
found, it will create a default config. See
[DeepFilterNet/pretrained_models/DeepFilterNet](https://github.com/Rikorose/DeepFilterNet/blob/main/DeepFilterNet/pretrained_models/DeepFilterNet/config.ini)
for a config file.
```py
# usage: train.py [-h] [--debug] data_config_file data_dir base_dir
python df/train.py path/to/dataset.cfg path/to/data_dir/ path/to/base_dir/
```

## Citation Guide

To reproduce any metrics, we recomend to use the python implementation via `pip install deepfilternet`.

If you use this framework, please cite: *DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering*
```bibtex
@inproceedings{schroeter2022deepfilternet,
  title={{DeepFilterNet}: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering}, 
  author = {Schröter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
  booktitle={ICASSP 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year={2022},
  organization={IEEE}
}
```

If you use the DeepFilterNet2 model, please cite: *DeepFilterNet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio*

```bibtex
@inproceedings{schroeter2022deepfilternet2,
  title = {{DeepFilterNet2}: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio},
  author = {Schröter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
  booktitle={17th International Workshop on Acoustic Signal Enhancement (IWAENC 2022)},
  year = {2022},
}
```

If you use the DeepFilterNet3 model, please cite: *DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement*

```bibtex
@inproceedings{schroeter2023deepfilternet3,
  title = {{DeepFilterNet}: Perceptually Motivated Real-Time Speech Enhancement},
  author = {Schröter, Hendrik and Rosenkranz, Tobias and Escalante-B., Alberto N. and Maier, Andreas},
  booktitle={INTERSPEECH},
  year = {2023},
}
```

If you use the multi-frame beamforming algorithms. please cite *Deep Multi-Frame Filtering for Hearing Aids*

```bibtex
@inproceedings{schroeter2023deep_mf,
  title = {Deep Multi-Frame Filtering for Hearing Aids},
  author = {Schröter, Hendrik and Rosenkranz, Tobias and Escalante-B., Alberto N. and Maier, Andreas},
  booktitle={INTERSPEECH},
  year = {2023},
}
```

## License

DeepFilterNet is free and open source! All code and model weights in this repository are dual-licensed under either:

* MIT License ([LICENSE-MIT](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/LICENSE-MIT) or [http://opensource.org/licenses/MIT](http://opensource.org/licenses/MIT))
* Apache License, Version 2.0 ([LICENSE-APACHE](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/LICENSE-APACHE) or [http://www.apache.org/licenses/LICENSE-2.0](http://www.apache.org/licenses/LICENSE-2.0))

at your option. This means you can select the license you prefer!

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

	---
	license: mit
	pipeline_tag: audio-to-audio
	library_name: onnx
	tags:
	- onnx
	- moonbox
	- onnxruntime
	- deepfilternet
	- speech-enhancement
	- denoising
	- noise-suppression
	- 48khz
	- real-time
	---

	# DeepFilterNet3 ONNX

	ONNX builds of [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet), made by Moonbox for on-device use with ONNX Runtime (CPU, CUDA, DirectML). Each
	precision is one folder; what the graphs take and return, and the signal processing that goes around them, is
	under How to run.

	## Files

	The last column is the benchmark on the VoiceBank-DEMAND test set (824 noisy utterances; details under *How to
	run*): each precision's speech-quality scores, and how far its output is from the original PyTorch model's.

	\| Precision \| File \| Size \| sha256 \| Benchmark \|
	\|---\|---\|---\|---\|---\|
	\| fp16 \| [`fp16/config.ini`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/config.ini) \| 2.1 kB \| `415eb925d449` \| PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) \|
	\| fp16 \| [`fp16/df_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/df_dec.onnx) \| 1.7 MB \| `a3380e4164a9` \| PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) \|
	\| fp16 \| [`fp16/enc.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/enc.onnx) \| 1.0 MB \| `694385ef6060` \| PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) \|
	\| fp16 \| [`fp16/erb_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/fp16/erb_dec.onnx) \| 1.7 MB \| `72126330232d` \| PESQ 3.163, STOI 0.9452, SI-SDR 18.91 dB; vs original 65.2 / 55.2 dB (CUDA) \|
	\| int8 \| [`int8/config.ini`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/config.ini) \| 2.1 kB \| `415eb925d449` \| PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) \|
	\| int8 \| [`int8/df_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/df_dec.onnx) \| 0.9 MB \| `7b729627da56` \| PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) \|
	\| int8 \| [`int8/enc.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/enc.onnx) \| 0.5 MB \| `41ac70016f7f` \| PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) \|
	\| int8 \| [`int8/erb_dec.onnx`](https://huggingface.co/moonbox-dev/DeepFilterNet3-ONNX/blob/main/int8/erb_dec.onnx) \| 0.9 MB \| `394ba8bd8f85` \| PESQ 3.159, STOI 0.9448, SI-SDR 18.89 dB; vs original 44.4 / 32.2 dB (CPU) \|

	## How to run

	DeepFilterNet3 (real-time speech enhancement at 48 kHz: ERB-band gains plus deep filtering of the low band) as the
	three ONNX graphs of the official export (`models/DeepFilterNet3_onnx.tar.gz` in the GitHub repository): `enc.onnx`
	(encoder), `erb_dec.onnx` (ERB-gain decoder) and `df_dec.onnx` (deep-filter decoder), with the model's
	`config.ini`. The time axis is free, so a whole file runs in one call. Our `fp16/` and `int8/` are made from those
	graphs; the fp32 originals are in the release itself.

	The graphs are the neural network only: the STFT, the ERB and complex features, applying the gains and the deep
	filter, and the ISTFT go around them. The simplest way to get all of that exactly right is the official
	`deepfilternet` package with its three networks swapped for these graphs (how this repository was benchmarked):

	```python
	import numpy as np, onnxruntime as ort, soundfile as sf, torch
	from df.enhance import enhance, init_df # pip install deepfilternet==0.5.6 (torch / torchaudio 2.1)

	model, df_state, _ = init_df() # the official DeepFilterNet3 checkpoint
	prov = ["CPUExecutionProvider"] # int8: CPU; fp16: ["CUDAExecutionProvider"]

	class Graph(torch.nn.Module):
	def __init__(self, path, first_only):
	super().__init__()
	self.s, self.first_only = ort.InferenceSession(path, providers=prov), first_only
	self.names = [i.name for i in self.s.get_inputs()]
	def forward(self, *xs):
	out = [torch.from_numpy(o) for o in self.s.run(None, {n: x.numpy() for n, x in zip(self.names, xs)})]
	return out[0] if self.first_only else tuple(out)

	model.enc, model.erb_dec, model.df_dec = (Graph("int8/enc.onnx", False), Graph("int8/erb_dec.onnx", True),
	Graph("int8/df_dec.onnx", True))
	noisy, sr = sf.read("noisy_48k.wav", dtype="float32") # 48 kHz mono
	with torch.no_grad():
	clean = enhance(model, df_state, torch.from_numpy(noisy)[None])
	sf.write("enhanced.wav", clean.squeeze(0).numpy(), 48000)
	```

	Without Python, the DeepFilterNet project's Rust crate (`libDF`, with its `tract` runtime) implements the same
	feature pipeline around these three graphs.

	Precisions. `int8/`: weights int8 per output channel behind DequantizeLinear (every Conv, grouped linear and GRU
	weight), activations fp32; 2.2 MB, runs on the CPU provider, the one to use for a real-time CPU denoiser. `fp16/`:
	weights and math in fp16, inputs and outputs fp32; for GPU providers (ONNX Runtime's CPU provider has no fp16 GRU).
	No int4 and no bf16 (no 4-bit or bf16 GRU kernels in ONNX Runtime).

	Benchmark. VoiceBank-DEMAND test set (Valentini et al.: 824 utterances by two unseen speakers, 48 kHz, the set
	the DeepFilterNet papers report), enhanced in one call per file with the post-filter off (the package default).
	PESQ is wide-band (ITU-T P.862.2, at 16 kHz), STOI at 48 kHz, SI-SDR against the clean speech; "vs original" is
	the SI-SDR of each output against the PyTorch model's output (mean / worst file). Real-time factor with 4 CPU threads
	on an Intel laptop CPU (int8) or the RTX 3080 Laptop GPU (fp16), whole pipeline.

	\| \| PESQ \| STOI \| SI-SDR (dB) \| vs original (dB) \| Real-time factor \|
	\|---\|---\|---\|---\|---\|---\|
	\| noisy input \| 1.97 \| 0.921 \| 8.39 \| - \| - \|
	\| original (PyTorch fp32, CPU) \| 3.163 \| 0.9452 \| 18.91 \| - \| 0.043 \|
	\| official ONNX fp32 (CPU) \| 3.163 \| 0.9452 \| 18.91 \| 135 / 130 \| 0.017 \|
	\| `fp16/` (CUDA) \| 3.163 \| 0.9452 \| 18.91 \| 65.2 / 55.2 \| 0.047 \|
	\| `int8/` (CPU) \| 3.159 \| 0.9448 \| 18.89 \| 44.4 / 32.2 \| 0.018 \|

	## How these were made

	From [DeepFilterNet3](https://github.com/Rikorose/DeepFilterNet) at revision `7b14b895fe`, starting from an fp32 ONNX graph (where it came from is said above).
	fp16: onnxconverter-common with inputs and outputs kept fp32, or, where the network's activations leave fp16's
	range, fp16 weights with fp32 math (said above when so). int8: weight-only, every Conv / MatMul / GRU weight int8
	with a per-output-channel scale behind DequantizeLinear and activations in fp32, so it runs on the CPU provider
	without a calibration set. int4: ONNX Runtime's MatMulNBits (weight-only, block 32) on the MatMul weights, the rest
	int8 as above, where the graph has enough MatMul weights to matter. No bf16: ONNX Runtime has no bf16 convolution or
	GRU kernels. Every graph was benchmarked before upload; the `sha256` column above is what the files must hash to.

	## License

	MIT/Apache-2.0. These files are a derivative of the original model and carry its license: it is copied in this
	repository, and it governs your use of these files. The original model card follows unchanged.

	---

	# Original model card

	# DeepFilterNet
	A Low Complexity Speech Enhancement Framework for Full-Band Audio (48kHz) using on Deep Filtering.

	![deepfilternet3](https://user-images.githubusercontent.com/16517898/225623209-a54fea75-ca00-404c-a394-c91d2d1146d2.svg)

	For PipeWire integration as a virtual noise suppression microphone look [here](https://github.com/Rikorose/DeepFilterNet/blob/main/ladspa/README.md).

	### Demo

	https://github.com/Rikorose/DeepFilterNet/assets/16517898/79679fd7-de73-4c22-948c-891927c7d2ca

	To run the demo (linux only) use:
	```bash
	cargo +nightly run -p df-demo --features ui --bin df-demo --release
	```

	### News

	- New DeepFilterNet Demo: DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement
	- Paper: https://arxiv.org/abs/2305.08227
	- Video: https://youtu.be/EO7n96YwnyE

	- New Multi-Frame Filtering Paper: Deep Multi-Frame Filtering for Hearing Aids
	- Paper: https://arxiv.org/abs/2305.08225

	- Real-time version and a LADSPA plugin
	- [Pre-compiled binary](#deep-filter), no python dependencies. Usage: `deep-filter audio-file.wav`
	- [LADSPA plugin](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/ladspa/) with pipewire filter-chain integration for real-time noise reduction on your mic.

	- DeepFilterNet2 Paper: DeepFilterNet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio
	- Paper: https://arxiv.org/abs/2205.05474
	- Samples: https://rikorose.github.io/DeepFilterNet2-Samples/
	- Demo: https://huggingface.co/spaces/hshr/DeepFilterNet2

	- Original DeepFilterNet Paper: DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering
	- Paper: https://arxiv.org/abs/2110.05588
	- Samples: https://rikorose.github.io/DeepFilterNet-Samples/
	- Demo: https://huggingface.co/spaces/hshr/DeepFilterNet
	- Video Lecture: https://youtu.be/it90gBqkY6k

	## Usage

	### deep-filter

	Download a pre-compiled deep-filter binary from the [release page](https://github.com/Rikorose/DeepFilterNet/releases/).
	You can use `deep-filter` to suppress noise in noisy .wav audio files. Currently, only wav files with a sampling rate of 48kHz are supported.

	```bash
	USAGE:
	deep-filter [OPTIONS] [FILES]...

	ARGS:
	<FILES>...

	OPTIONS:
	-D, --compensate-delay
	Compensate delay of STFT and model lookahead
	-h, --help
	Print help information
	-m, --model <MODEL>
	Path to model tar.gz. Defaults to DeepFilterNet2.
	-o, --out-dir <OUT_DIR>
	[default: out]
	--pf
	Enable postfilter
	-v, --verbose
	Logging verbosity
	-V, --version
	Print version information
	```

	If you want to use the pytorch backend e.g. for GPU processing, see further below for the Python usage.

	### DeepFilterNet Framework

	This framework supports Linux, MacOS and Windows. Training is only tested under Linux. The framework is structured as follows:

	* `libDF` contains Rust code used for data loading and augmentation.
	* `DeepFilterNet` contains DeepFilterNet code training, evaluation and visualization as well as pretrained model weights.
	* `pyDF` contains a Python wrapper of libDF STFT/ISTFT processing loop.
	* `pyDF-data` contains a Python wrapper of libDF dataset functionality and provides a pytorch data loader.
	* `ladspa` contains a LADSPA plugin for real-time noise suppression.
	* `models` contains pretrained for usage in DeepFilterNet (Python) or libDF/deep-filter (Rust)

	### DeepFilterNet Python: PyPI

	Install the DeepFilterNet Python wheel via pip:
	```bash
	# Install cpu/cuda pytorch (>=1.9) dependency from pytorch.org, e.g.:
	pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
	# Install DeepFilterNet
	pip install deepfilternet
	# Or install DeepFilterNet including data loading functionality for training (Linux only)
	pip install deepfilternet[train]
	```

	To enhance noisy audio files using DeepFilterNet run
	```bash
	# Specify an output directory with --output-dir [OUTPUT_DIR]
	deepFilter path/to/noisy_audio.wav
	```

	### Manual Installation

	Install cargo via [rustup](https://rustup.rs/). Usage of a `conda` or `virtualenv` recommended.
	Please read the comments and only execute the commands that you need.

	Installation of python dependencies and libDF:
	```bash
	cd path/to/DeepFilterNet/ # cd into repository
	# Recommended: Install or activate a python env
	# Mandatory: Install cpu/cuda pytorch (>=1.8) dependency from pytorch.org, e.g.:
	pip install torch torchaudio -f https://download.pytorch.org/whl/cpu/torch_stable.html
	# Install build dependencies used to compile libdf and DeepFilterNet python wheels
	pip install maturin poetry

	# Install remaining DeepFilterNet python dependencies
	# Option A: Install DeepFilterNet python wheel globally within your environment. Do this if you want use
	# this repos as is, and don't want to develop within this repository.
	poetry -C DeepFilterNet install -E train -E eval
	# Option B: If you want to develop within this repo, install only dependencies and work with the repository version
	poetry -C DeepFilterNet install -E train -E eval --no-root
	export PYTHONPATH=$PWD/DeepFilterNet # And set the python path correctly

	# Build and install libdf python package required for enhance.py
	maturin develop --release -m pyDF/Cargo.toml
	# Optional: Install libdfdata python package with dataset and dataloading functionality for training
	# Required build dependency: HDF5 headers (e.g. ubuntu: libhdf5-dev)
	maturin develop --release -m pyDF-data/Cargo.toml
	# If you have troubles with hdf5 you may try to build and link hdf5 statically:
	maturin develop --release --features hdf5-static -m pyDF-data/Cargo.toml
	```

	### Use DeepFilterNet from command line

	To enhance noisy audio files using DeepFilterNet run
	```bash
	$ python DeepFilterNet/df/enhance.py --help
	usage: enhance.py [-h] [--model-base-dir MODEL_BASE_DIR] [--pf] [--output-dir OUTPUT_DIR] [--log-level LOG_LEVEL] [--compensate-delay]
	noisy_audio_files [noisy_audio_files ...]

	positional arguments:
	noisy_audio_files List of noise files to mix with the clean speech file.

	optional arguments:
	-h, --help show this help message and exit
	--model-base-dir MODEL_BASE_DIR, -m MODEL_BASE_DIR
	Model directory containing checkpoints and config.
	To load a pretrained model, you may just provide the model name, e.g. `DeepFilterNet`.
	By default, the pretrained DeepFilterNet2 model is loaded.
	--pf Post-filter that slightly over-attenuates very noisy sections.
	--output-dir OUTPUT_DIR, -o OUTPUT_DIR
	Directory in which the enhanced audio files will be stored.
	--log-level LOG_LEVEL
	Logger verbosity. Can be one of (debug, info, error, none)
	--compensate-delay, -D
	Add some paddig to compensate the delay introduced by the real-time STFT/ISTFT implementation.

	# Enhance audio with original DeepFilterNet
	python DeepFilterNet/df/enhance.py -m DeepFilterNet path/to/noisy_audio.wav

	# Enhance audio with DeepFilterNet2
	python DeepFilterNet/df/enhance.py -m DeepFilterNet2 path/to/noisy_audio.wav
	```

	### Use DeepFilterNet within your Python script

	```py
	from df import enhance, init_df

	model, df_state, _ = init_df() # Load default model
	enhanced_audio = enhance(model, df_state, noisy_audio)
	```

	See [here](https://github.com/Rikorose/DeepFilterNet/blob/main/scripts/external_usage.py) for a full example.

	### Training

	The entry point is `DeepFilterNet/df/train.py`. It expects a data directory containing HDF5 dataset
	as well as a dataset configuration json file.

	So, you first need to create your datasets in HDF5 format. Each dataset typically only
	holds training, validation, or test set of noise, speech or RIRs.
	```py
	# Install additional dependencies for dataset creation
	pip install h5py librosa soundfile
	# Go to DeepFilterNet python package
	cd path/to/DeepFilterNet/DeepFilterNet
	# Prepare text file (e.g. called training_set.txt) containing paths to .wav files
	#
	# usage: prepare_data.py [-h] [--num_workers NUM_WORKERS] [--max_freq MAX_FREQ] [--sr SR] [--dtype DTYPE]
	# [--codec CODEC] [--mono] [--compression COMPRESSION]
	# type audio_files hdf5_db
	#
	# where:
	# type: One of `speech`, `noise`, `rir`
	# audio_files: Text file containing paths to audio files to include in the dataset
	# hdf5_db: Output HDF5 dataset.
	python df/scripts/prepare_data.py --sr 48000 speech training_set.txt TRAIN_SET_SPEECH.hdf5
	```
	All datasets should be made available in one dataset folder for the train script.

	The dataset configuration file should contain 3 entries: "train", "valid", "test". Each of those
	contains a list of datasets (e.g. a speech, noise and a RIR dataset). You can use multiple speech
	or noise dataset. Optionally, a sampling factor may be specified that can be used to over/under-sample
	the dataset. Say, you have a specific dataset with transient noises and want to increase the amount
	of non-stationary noises by oversampling. In most cases you want to set this factor to 1.

	<details>
	<summary>Dataset config example:</summary>
	<p>

	`dataset.cfg`

	```json
	{
	"train": [
	[
	"TRAIN_SET_SPEECH.hdf5",
	1.0
	],
	[
	"TRAIN_SET_NOISE.hdf5",
	1.0
	],
	[
	"TRAIN_SET_RIR.hdf5",
	1.0
	]
	],
	"valid": [
	[
	"VALID_SET_SPEECH.hdf5",
	1.0
	],
	[
	"VALID_SET_NOISE.hdf5",
	1.0
	],
	[
	"VALID_SET_RIR.hdf5",
	1.0
	]
	],
	"test": [
	[
	"TEST_SET_SPEECH.hdf5",
	1.0
	],
	[
	"TEST_SET_NOISE.hdf5",
	1.0
	],
	[
	"TEST_SET_RIR.hdf5",
	1.0
	]
	]
	}
	```

	</p>
	</details>

	Finally, start the training script. The training script may create a model `base_dir` if not
	existing used for logging, some audio samples, model checkpoints, and config. If no config file is
	found, it will create a default config. See
	[DeepFilterNet/pretrained_models/DeepFilterNet](https://github.com/Rikorose/DeepFilterNet/blob/main/DeepFilterNet/pretrained_models/DeepFilterNet/config.ini)
	for a config file.
	```py
	# usage: train.py [-h] [--debug] data_config_file data_dir base_dir
	python df/train.py path/to/dataset.cfg path/to/data_dir/ path/to/base_dir/
	```

	## Citation Guide

	To reproduce any metrics, we recomend to use the python implementation via `pip install deepfilternet`.

	If you use this framework, please cite: DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering
	```bibtex
	@inproceedings{schroeter2022deepfilternet,
	title={{DeepFilterNet}: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering},
	author = {Schröter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
	booktitle={ICASSP 2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
	year={2022},
	organization={IEEE}
	}
	```

	If you use the DeepFilterNet2 model, please cite: DeepFilterNet2: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio

	```bibtex
	@inproceedings{schroeter2022deepfilternet2,
	title = {{DeepFilterNet2}: Towards Real-Time Speech Enhancement on Embedded Devices for Full-Band Audio},
	author = {Schröter, Hendrik and Escalante-B., Alberto N. and Rosenkranz, Tobias and Maier, Andreas},
	booktitle={17th International Workshop on Acoustic Signal Enhancement (IWAENC 2022)},
	year = {2022},
	}
	```

	If you use the DeepFilterNet3 model, please cite: DeepFilterNet: Perceptually Motivated Real-Time Speech Enhancement

	```bibtex
	@inproceedings{schroeter2023deepfilternet3,
	title = {{DeepFilterNet}: Perceptually Motivated Real-Time Speech Enhancement},
	author = {Schröter, Hendrik and Rosenkranz, Tobias and Escalante-B., Alberto N. and Maier, Andreas},
	booktitle={INTERSPEECH},
	year = {2023},
	}
	```

	If you use the multi-frame beamforming algorithms. please cite Deep Multi-Frame Filtering for Hearing Aids

	```bibtex
	@inproceedings{schroeter2023deep_mf,
	title = {Deep Multi-Frame Filtering for Hearing Aids},
	author = {Schröter, Hendrik and Rosenkranz, Tobias and Escalante-B., Alberto N. and Maier, Andreas},
	booktitle={INTERSPEECH},
	year = {2023},
	}
	```

	## License

	DeepFilterNet is free and open source! All code and model weights in this repository are dual-licensed under either:

	* MIT License ([LICENSE-MIT](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/LICENSE-MIT) or [http://opensource.org/licenses/MIT](http://opensource.org/licenses/MIT))
	* Apache License, Version 2.0 ([LICENSE-APACHE](https://raw.githubusercontent.com/Rikorose/DeepFilterNet/7b14b895fe7fc42cab72b4a4c1e49c982bd6404b/LICENSE-APACHE) or [http://www.apache.org/licenses/LICENSE-2.0](http://www.apache.org/licenses/LICENSE-2.0))

	at your option. This means you can select the license you prefer!

	Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.