deepfilter-stream

Real-time streaming noise cancellation with DeepFilterNet3 on ONNX Runtime.

deepfilter-stream wraps the DeepFilterNet3 model (arXiv:2110.05588) in a minimal, dependency-light Python API. It ships the model as an ONNX graph so that inference runs entirely on CPU via ONNX Runtime — no PyTorch required. The library is designed for low-latency, real-time pipelines: a single stream runs at roughly 7x faster than real-time on a modern CPU core, measured end-to-end algorithmic latency is ~32 ms, and a single machine can run 16+ concurrent real-time streams on a typical CPU.

Installation

pip install deepfilter-stream

The ONNX model (~13 MB) downloads automatically on first use and is cached in the user's platform cache directory. See Model download below.

Quickstart

Library API

from deepfilter_stream import DeepFilterModel, Denoiser

# Option A: shared model, one stream per thread
model = DeepFilterModel()
stream = model.new_stream(atten_lim_db=None)  # or atten_lim_db=20.0 to blend in dry signal

# Feed arbitrary-length chunks at any sample rate; returns enhanced float32 mono
import numpy as np
noisy = np.random.randn(9600).astype(np.float32)  # 0.2 s at 48 kHz
enhanced = stream.process(noisy, sr=48000)

# Flush any buffered tail at end of file
tail = stream.flush()

# Reset stream state without reloading the model (e.g. next utterance)
stream.reset()

# Option B: standalone Denoiser (loads its own model internally)
denoiser = Denoiser()
enhanced = denoiser.process(noisy, sr=48000)

Frame-by-frame API (lowest latency)

model = DeepFilterModel()
stream = model.new_stream()

frame_size = stream.frame_size   # 512 samples at 48 kHz
frame = np.zeros(frame_size, dtype=np.float32)
enhanced_frame = stream.process_frame(frame)

Live demo (mic -> denoise -> speakers)

Warning: wear headphones to avoid feedback when using the live demo.

deepfilter-stream                          # use default mic/speakers
deepfilter-stream --list-devices           # list audio devices
deepfilter-stream --input-device 2 --output-device 4
deepfilter-stream --atten-lim-db 20       # blend 20 dB noise reduction with dry signal

API reference

Latency

• Hop size: 512 samples @ 48 kHz = 10.67 ms per frame.
• latency_ms property: reports the per-frame STFT minimum (~10.7 ms), which is the minimum algorithmic offset introduced by framing alone.
• Measured end-to-end algorithmic latency: ~32 ms (≈ 3 hops), measured via impulse benchmark. This reflects the actual signal delay through the STFT/iSTFT and GRU pipeline.
• Real-world latency adds device/buffer round-trip time on top of the algorithmic offset.

Performance

Measured on a single CPU core (no GPU required):

Concurrency

The DeepFilterModel session is thread-safe and shareable. Each Denoiser stream holds its own GRU hidden state and is not thread-safe — use one stream per thread.

from deepfilter_stream import DeepFilterModel

# For servers with many concurrent streams, limit ONNX intra-op threads
# so that thread contention across streams does not reduce throughput:
model = DeepFilterModel(intra_op_num_threads=1)

# Each worker thread gets its own stream; model is shared
def worker():
    stream = model.new_stream()
    # ... process audio

Model download

The ONNX model is downloaded automatically on first use and cached under the platform cache directory (e.g. ~/.cache/deepfilter-stream/dfn3-512-v1/ on Linux).

To use a local copy, set the environment variable:

export DEEPFILTER_STREAM_MODEL_DIR=/path/to/folder
# folder must contain: denoiser_model.onnx, initial_states.npz, meta.json

Attribution

deepfilter-stream bundles the DeepFilterNet3 model weights. The original work:

• Repository: https://github.com/Rikorose/DeepFilterNet
• Paper: arXiv:2110.05588 — "DeepFilterNet: A Low Complexity Speech Enhancement Framework for Full-Band Audio based on Deep Filtering"

Dual-licensed under MIT and Apache 2.0 (your choice). See LICENSE and NOTICE.