audiosronnx 0.9.0a1

Pure-ONNX multi-engine audio super-resolution / bandwidth extension (upscale speech to 48 kHz)

audiosronnx

Speech restoration, denoising and bandwidth extension in pure ONNX. Clean up noisy recordings, and upscale narrowband speech — 8 kHz telephony, muffled captures, low-fidelity TTS — to a full 48 kHz signal.

Inference runs entirely on onnxruntime. There is no torch at runtime: every STFT, filterbank and resampler lives in numpy/scipy, so the ONNX graphs stay portable and the front-ends stay inspectable.

pip install audiosronnx

from audiosronnx import load_denoise, load_sr

clean, rate = load_denoise("dpdfnet").denoise("noisy_call.wav")   # remove noise
wide,  _    = load_sr("lavasr").upscale(clean, rate)              # extend to 48 kHz

audiosronnx denoise noisy.wav clean.wav
audiosronnx upscale clean.wav wide_48k.wav

Runtime dependencies are numpy, onnxruntime, scipy, soundfile and huggingface_hub. Weights download on first use and cache under ~/.local/share/audiosronnx, pinned by revision.

Two jobs

Denoising removes background noise and keeps the sample rate. Bandwidth extension invents high-frequency content that was never in the input and always outputs 48 kHz. They have separate entry points — load_denoise() and load_sr() — which refuse each other's engines.

They compose, in that order: denoise first, then extend. Run them the other way round and the extender reconstructs a high band out of the noise.

Denoisers

Engine	Rate	Size	License	Reach for it when
dpdfnet (default)	8 / 16 / 48 kHz	8.7–14.9 MB	Apache-2.0	you have no other reason to choose
mossformer2	48 kHz	229 MB	Apache-2.0	the source is fullband and quality wins
frcrn	16 kHz	57.5 MB	Apache-2.0	reproducing published benchmarks
mpsenet	16 kHz	9.7 MB	MIT	you want a spectral model that stays small
gtcrn	16 kHz	0.54 MB	MIT	footprint is the binding constraint
cmgan	16 kHz	7.8 MB	MIT	you want a conformer metric-GAN specifically
metadenoiser	16 kHz	19–34 MB	CC-BY-NC-4.0	you want a time-domain model and NC is acceptable
mossformergan	16 kHz	17.7 MB	Apache-2.0	highest published PESQ (3.47)
voicefixer	44.1 kHz	415 MB	MIT	restoration — noise, reverb, clipping and band loss at once
deepfilternet	48 kHz	~2 MB	MIT	you already depend on libdf

SNR recovered on one clip at 19 / 11 / 5 dB input SNR: dpdfnet +4.9 / +10.3 / +13.7 dB, mossformer2 +5.9 / +10.4 / +13.4, mpsenet +4.8 / +8.9 / +11.7, frcrn +4.6 / +8.8 / +11.7, gtcrn +3.5 / +7.3 / +7.5. That is broadband Gaussian noise — a hostile synthetic case that ranks engines consistently but predicts little about babble or codec artefacts.

Recommendations: start with dpdfnet — no extra dependencies, 8/16/48 kHz, top-tier measured gain. Take mossformer2 for the best fullband quality, mossformergan for the best benchmark scores, gtcrn when footprint is the constraint, metadenoiser for a time-domain model. Engines are kept even when something else beats them, so published results stay reproducible — cmgan is dominated and stays for that reason.

Full detail in docs/denoising.md.

Bandwidth extension

Engine	Input	Size	CPU speed	License
lavasr (default)	8–48 kHz	~52 MB	~50× realtime	Apache-2.0
novasr	16 kHz	~0.2 MB	~1000× realtime	Apache-2.0
flowhigh	any	~200 MB	slow (4 ODE steps)	MIT
hifiganbwe	any	~4 MB	fast	MIT
apbwe	any (12 kHz band)	~120 MB	moderate	MIT
sidon	16 kHz	~410 MB	~0.6× realtime	MIT
callenhancer	8–16 kHz	~3 GB / ~1.3 GB int8	~0.3–0.4× realtime	CC-BY-NC-4.0

lavasr is the general-purpose choice. sidon and callenhancer do something different: they resynthesise speech rather than extend its band, repairing codec damage a bandwidth extender cannot, at a much higher cost. callenhancer is trained on telephony specifically, and its weights are non-commercial.

Full detail in docs/engines.md.

Documentation

Page	Contents
docs/engines.md	Every bandwidth-extension engine, options, precision
docs/denoising.md	Every denoiser, and how to choose
docs/api.md	load_sr, load_denoise, the model contracts, the registry
docs/cli.md	Command-line reference
docs/custom-engines.md	Registering your own engine
docs/conversion.md	How the ONNX graphs are produced and validated
docs/not-shipped.md	Models evaluated and rejected, with reasons

Runnable scripts are in examples/.

API

sr = load_sr("lavasr")
out, rate = sr.upscale("telephone_8k.wav")   # path, int16 bytes, or numpy array
sr.upscale_file("in.wav", "out_48k.wav")
sr.upscale_dir("clips_in/", "clips_out/")

dn = load_denoise("dpdfnet")
clean, rate = dn.denoise(audio, in_rate)     # returns the ENGINE's rate
dn.denoise_file("in.wav", "clean.wav")
dn.denoise_dir("clips_in/", "clips_out/")

available_models(); available_denoisers(); get_engine(alias); register_engine(entry)

Output is always mono float32. Empty, silent and non-finite input produce finite output rather than raising. See docs/api.md.

Which engines ship

An engine ships when it exports to a single static ONNX graph, runs on CPU through onnxruntime, carries a permissive license, and is validated end-to-end against the upstream pipeline — not merely against the model, since reproducing a network but not its surrounding normalisation yields audio that sounds plausible and is quietly wrong.

That rules out diffusion and flow-matching models, location-variable convolutions, and unlicensed checkpoints. docs/not-shipped.md names every candidate considered and which condition it failed.

License

Apache-2.0.

Model weights carry their upstream licenses, listed per engine in the docs and reported by audiosronnx list. Most are MIT or Apache-2.0; callenhancer is CC-BY-NC-4.0, which covers the weights rather than audio processed with them.