denoise-audio 0.1.0

Denoise WAV audio using RNNoise, DeepFilterNet, and FacebookResearch Denoiser.

denoise-audio

A CLI tool for denoising WAV audio with 3 leading backends:

• rnnoise (fast, CPU-only, low-latency)
• deepfilternet (high-quality full-band denoising)
• fbdenoiser (FacebookResearch Denoiser / causal Demucs; strong enhancement)

Inputs/outputs are WAV files. If your audio is not WAV, convert it first.

---

Install uv

Install uv using Astral’s standalone installer:

https://docs.astral.sh/uv/getting-started/installation/#standalone-installer

Verify:

uv --version

---

Install dependencies

git clone https://github.com/Surya-Rayala/denoise-audio.git
cd denoise-audio
uv sync

---

CLI usage

Run everything with uv run to ensure you’re using the project environment:

Global help

uv run python -m denoise --help

List available backends

uv run python -m denoise --list-models

Model-specific help

Pass --model <name> and --help to see only that backend’s options:

uv run python -m denoise --model rnnoise --help
uv run python -m denoise --model deepfilternet --help
uv run python -m denoise --model fbdenoiser --help

---

Basic command (all models)

uv run python -m denoise --model <model> --input <in.wav> --output <out.wav>

• <model> is one of: rnnoise, deepfilternet, fbdenoiser
• --input and --output must be WAV files

---

Models

RNNoise (--model rnnoise)

Basic command

uv run python -m denoise --model rnnoise --input <in.wav> --output <out.wav>

Arguments

• --rnnoise-sample-rate <int>: Force the RNNoise wrapper sample rate. If omitted, the tool infers the sample rate from the input WAV (recommended).

---

DeepFilterNet (--model deepfilternet)

Basic command

uv run python -m denoise --model deepfilternet --input <in.wav> --output <out.wav>

Arguments

• --df-model <name>: Select which pretrained DeepFilterNet model to load. Common options: DeepFilterNet, DeepFilterNet2, DeepFilterNet3.
  • To see available models, visit: https://github.com/Rikorose/DeepFilterNet/tree/main/models
• --df-pf: Enable the post-filter (can reduce residual noise; may sound more aggressive in very noisy sections).
• --df-compensate-delay: Add padding to compensate processing delay (useful when you need better alignment with the original audio).

---

FacebookResearch Denoiser (--model fbdenoiser)

Basic command

uv run python -m denoise --model fbdenoiser --input <in.wav> --output <out.wav>

Arguments

• --fb-model <name>: Choose the pretrained model:
  • dns48: pre-trained real time H=48 model trained on DNS
  • dns64: pre-trained real time H=64 model trained on DNS
  • master64: pre-trained real time H=64 model trained on DNS and Valentini
• --fb-device <device>: Inference device, e.g. cpu or cuda. If omitted, it automatically uses CUDA if available, otherwise CPU.
• --fb-dry <float>: Dry/wet mix (0.0 = original input only, 1.0 = denoised output only). Values outside [0.0, 1.0] are clamped.
• --fb-streaming: Enable streaming mode (flag is accepted by the CLI).
• --fb-batch-size <int>: Batch size (flag is accepted by the CLI).
• --fb-num-workers <int>: Number of workers (flag is accepted by the CLI).
• --fb-verbose: Enable verbose logging (flag is accepted by the CLI).

---

Updating from Git

If you’re using the repo locally with uv sync + uv run, updating to new code from Git is:

cd denoise-audio
git pull
uv sync

If you have local changes and git pull refuses

Use one of these (pick what you intend):

• Keep your local changes and rebase on top:

git pull --rebase
uv sync

• Discard local changes and reset to remote (⚠️ destructive):

git fetch origin
git reset --hard origin/main
uv sync

License

This project's source code is licensed under the MIT License.

Note on Dependencies: This tool relies on the denoiser library (Facebook Research), which is licensed under CC-BY-NC 4.0 (Non-Commercial). Consequently, this tool as a whole is suitable for research and personal use only, unless you obtain a commercial license for the denoiser dependency.