YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
Speechless ZipEnhancer noise suppression
Speechless uses the unchanged ONNX export from Navi's listening comparison.
The upstream model is iic/speech_zipenhancer_ans_multiloss_16k_base (Apache-2.0).
ZipformerAns is the runtime class name; the model is ZipEnhancer.
- File:
zipenhancer-16k-200ms-v1.onnx, 10,708,641 bytes. - SHA-256:
92ffea0323d4342ff5d0a0782e3e9adaf70afd67559e3375955a631a25b9eebf. - Public repository:
trevorlink/speechless-zipenhancer-onnx, revisionv1.0.0. - Input/output: float32 magnitude and phase,
[1, 201, 33]. - 16 kHz, 200 ms windows, 100 ms overlap; FFT/window 400, hop 100, periodic Hann window, reflection padding, magnitude compression 0.3.
zipformer.rs ports Navi's STFT/model/ISTFT and overlap/add math. CPU inference
uses one thread, matching the implementation auditioned before integration.
File processing scans disk for per-channel recording-level normalization, then
streams fixed windows with bounded memory. Every original channel is processed
independently. FFmpeg restores the source sample rate/channel count and copies
video while taking the enhanced audio from the second input. The model's
16 kHz bandwidth applies even when exported at a higher sample rate.
Both Agent audio_cleanup(operation=denoise) and Audio Polish preview/export
use denoise.rs. Balanced/strong use the full model output; light mixes 65%
enhanced and 35% original audio. Audio Polish keeps its additional high-pass,
compression, optional declick and loudness processing.
The normal model downloader installs the pinned size/hash-verified artifact
in the application data models/zipenhancer-ans-v1 folder at first use.
The ORT session is reused across requests; downloads, waiting for the session,
decoding, each inference window and encoding support cancellation.
Publish using scripts/publish-zipenhancer.py --model /path/to/model.onnx --env-file ~/AI/.env.local. Weights are hosted externally and never committed.
The export script, upstream model card, numerical validation and Apache license
are retained here for reproducibility and attribution.