PyRnNoise

0.4.3 verified Fri May 01 auth: no python

Python wrapper for the RNN noise suppression library (RNNoise), enabling real-time speech denoising. Current version 0.4.3, maintained on GitHub.

pip install pyrnnoise

Common errors

error AttributeError: module 'pyrnnoise' has no attribute 'RNNoise' ↓

cause Incorrect import path; often due to an older version or typo.

fix

Use 'from pyrnnoise import RNNoise'. Ensure you have version 0.4.0 or later.

error RuntimeError: Input audio must be 1D array of float32 ↓

cause Audio data is int16, int32, or multi-channel (2D array).

fix

Convert to mono float32: audio = audio.astype(np.float32) if audio.dtype != np.float32 else audio; if audio.ndim > 1: audio = audio.mean(axis=1).

error OSError: No such file or directory: 'librnnoise.so' ↓

cause Missing native dependency (RNNoise shared library). This can happen if the platform is not supported or the C library is not compiled.

fix

Install from source: on Linux, apt install build-essential cmake; on Windows, use precompiled wheels or build manually.

Warnings

breaking The RNNoise class expects audio data as a numpy array of shape (N,) with dtype float32. Providing int16 or multichannel data will cause silent errors or poor results. ↓

fix Convert to float32 and ensure mono: audio = audio.astype(np.float32) / 32768.0 for int16 input.

gotcha The filter() method expects the entire audio at once; it does not support streaming or chunked processing out of the box. Attempting to call filter() on small chunks may produce artifacts at boundaries. ↓

fix Process the whole audio array in one call, or implement overlap-add for streaming.

deprecated The 'sample_rate' parameter in RNNoise constructor is deprecated and ignored; the model always operates at 48 kHz. Passing it may raise a warning in future versions. ↓

fix Remove the sample_rate argument; resample your audio to 48 kHz before passing to denoiser.

Imports

RNNoise
wrong
```
import rnnoise
```
correct
```
from pyrnnoise import RNNoise
```
The package is named pyrnnoise, not rnnoise.

Quickstart

Load a noisy audio file, filter it with RNNoise, and save the denoised result.

from pyrnnoise import RNNoise
import soundfile as sf

denoiser = RNNoise()
data, samplerate = sf.read('noisy_audio.wav')
output = denoiser.filter(data)
sf.write('denoised_audio.wav', output, samplerate)