Replace DeepFilterNet with an actively maintained alternative

[MaleicAcid]

Problem

DeepFilterNet (last release: v0.5.6, August 2023) has been unmaintained for nearly three years. This causes several concrete problems:

Breaks on modern PyTorch. DeepFilterNet's native dependency deepfilterlib imports torchaudio.backend.common, a module that was removed in torchaudio 2.6. Any user with torch>=2.6 and noise_suppress=True will hit an ImportError at runtime. This effectively pins the entire project to torch<=2.5, preventing users from benefiting from newer PyTorch performance improvements and bug fixes.

No pre-built wheels for Python 3.12+. deepfilterlib is a Rust native extension built with pyo3/maturin. The maintainer only published wheels for CPython 3.8–3.11. Users on Python 3.12 or later must have a working Rust toolchain (cargo, maturin) to compile from source — a significant barrier for most Python users. This is currently the sole blocker preventing openlrc from declaring Python 3.13/3.14 support.

Heavy and fragile dependency chain. Even though noise suppression is off by default, the [full] extra pulls in deepfilternet → deepfilterlib (Rust native) → torch → torchaudio. The Rust compilation step is slow, error-prone on Windows, and incompatible with common packaging tools like PyInstaller and Nuitka.

No sign of upstream activity. The DeepFilterNet repository has had no releases, no merged PRs addressing compatibility, and no maintainer responses on these issues since August 2023. There is no indication this will change.

Alternatives

SpeechBrain (MetricGAN+)

SpeechBrain (11k+ stars, latest release v1.1.0, March 2026) is an actively maintained speech toolkit. Its metricgan-plus-voicebank pretrained model provides speech enhancement quality comparable to DeepFilterNet.

• Requires torch>=2.1.0 with no upper bound — works with latest PyTorch
• Pure Python — no Rust toolchain needed
• Pretrained models hosted on HuggingFace, downloaded on first use
• Apache 2.0 license

from speechbrain.pretrained import SpectralMaskEnhancement
enhancer = SpectralMaskEnhancement.from_hparams(source="speechbrain/metricgan-plus-voicebank")
enhanced = enhancer.enhance_file("noisy.wav")

noisereduce

noisereduce (1.8k stars) uses classical spectral gating. No torch dependency at all (numpy/scipy only). Lower quality than neural approaches but adequate for mild noise and extremely lightweight (~100KB).

Comparison

	DeepFilterNet (current)	SpeechBrain	noisereduce
Last release	Aug 2023	Mar 2026	Oct 2024
Approach	Neural (deep filtering)	Neural (MetricGAN+)	Classical (spectral gating)
Torch compatibility	Pinned <=2.5	>=2.1, no upper bound	Not required
Rust toolchain	Required	No	No
Python 3.12+	No wheels	Yes	Yes
Enhancement quality	High	High	Moderate

Discussion

1. Should we run a quality comparison test before committing to SpeechBrain?
2. Should noisereduce be offered as a lightweight fallback (e.g. base noise_suppress=True uses noisereduce, [full] uses SpeechBrain)?

[MaleicAcid]

I'd like to nominate DPDFNet (ceva-ip/DPDFNet) as another viable replacement candidate worth considering.
Key points relevant to this issue:

• Actively maintained. Last release Jun 2026, by CEVA (a DSP company). Paper: "DPDFNet: Boosting DeepFilterNet2 via Dual-Path RNN".
• No PyTorch dependency. Inference runs on pure ONNX Runtime CPU — no torch, no torchaudio, no version pinning issues.
• No Rust compilation. Pure Python package (pip install dpdfnet), no cargo/maturin needed. Solves the Python 3.12+ wheel problem by design.
• Small footprint. ~50MB total (model files downloaded on demand). Compare to dragging in the full PyTorch stack (~2GB).
• Comparable quality. Same DCRN + deep filtering paradigm as DeepFilterNet2, enhanced with Dual-Path RNN blocks. Not identical output, but same class of neural enhancement.
  API is straightforward:

import dpdfnet
enhanced = dpdfnet.enhance(audio, sample_rate=48000, model="dpdfnet4")

@zh-plus We should probably make a decision on what to do about DeepFilterNet — replace it or just drop it. We can't avoid this forever.
In my opinion, noise suppression is a domain-specific problem best handled by dedicated tools. Bundling it into openlrc only bloats the dependency tree and drifts from the project's core purpose. Keeping openlrc focused on the translation pipeline, and letting users choose their own preprocessing tools, would be a cleaner design.

[zh-plus]

I’ve tested DPDFNet, and it looks like a viable replacement. We can drop the DeepFilterNet dependency and use DPDFNet instead.

I agree that noise suppression can be handled externally in some workflows, but I still see openlrc as an end-to-end toolchain that should work well out of the box. Requiring users to discover, install, and integrate a separate preprocessing tool would add friction and make the default experience less complete.

Since DPDFNet avoids the dependency and compatibility problems of DeepFilterNet while keeping the built-in denoising capability, I’d prefer replacing DeepFilterNet with DPDFNet rather than dropping noise suppression entirely.