metadata
license: mit
tags:
- coreml
- audio
- speech-enhancement
Clarity transient network (Core ML, FP16)
Finds clicks, knocks and other short noises in speech for Clarity Sound Studio, a free, open-source macOS app. It runs after noise reduction and gives a keep mask per STFT bin and an event probability per frame.
| Files | TransientNet-v1-fp16.mlmodelc (mask and events) and TransientNet-v1-fp16-event.mlmodelc (the same checkpoint, event output only); compiled ML Programs, FP16, macOS 14+ |
| Model | 1.1 M parameters over log power spectrum frames |
| Training | Speech from the CSTR VCTK Corpus mixed with sound events and backgrounds from FSD50K (speakers p232 and p257 held out) |
Licence
The model is released under the MIT licence; see LICENSE. Copyright (c) 2026 arkaydeus.
Training data attribution
- CSTR VCTK Corpus version 0.92 (J. Yamagishi, C. Veaux and K. MacDonald, Centre for Speech Technology Research, University of Edinburgh), https://doi.org/10.7488/ds/2645, used under the Creative Commons Attribution 4.0 International licence. The recordings were cut, mixed with other sounds and passed through noise reduction for training.
- FSD50K (E. Fonseca, X. Favory, J. Pons, F. Font and X. Serra, "FSD50K: an Open Dataset of Human-Labeled Sound Events", IEEE/ACM TASLP 2022), https://zenodo.org/records/4060432, used under the Creative Commons Attribution 4.0 International licence. Only clips under CC0 1.0 or CC BY 3.0 were used; every CC BY 3.0 clip and its author is credited in
FSD50K-CREDITS.md. The clips were cut and mixed with speech for training.
No corpus audio or clip is distributed in this repository. No warranty is given by the licensors.