ark-tech/filler-detector-coreml

Clarity Filler Detector (Core ML, FP16)

Finds filler words ("um", "uh") and stutters in speech for Clarity Sound Studio, a free, open-source macOS app. The app shows each detection for review and cuts the ones the user approves.

File	FillerDetector-v3.mlpackage, version 3.1.0 (ML Program, FP16, macOS 14+)
Model	A 6.6 M-parameter Conformer over log-mel frames, with a small tree scorer (stored in the model's fusion metadata) that combines its output with the transcript
Training	From scratch on the AMI training split and the ICSI Meeting Corpus (128.5 h), with reverberation augmentation from OpenSLR SLR28
Results	On the AMI test split (10.4 h, 2,805 fillers), counting a detection that overlaps a label: 85.1% precision, 83.0% recall

Licence

The model is released under the MIT licence; see LICENSE. Copyright (c) 2026 arkaydeus.

Training data attribution

• AMI Meeting Corpus (© the AMI Consortium and the University of Edinburgh), https://groups.inf.ed.ac.uk/ami/corpus/, used under the Creative Commons Attribution 4.0 International licence. The audio and annotations were cut, relabelled and augmented for training.
• ICSI Meeting Corpus (© the International Computer Science Institute, Berkeley; release distributed by the University of Edinburgh), https://groups.inf.ed.ac.uk/ami/icsi/, used under the Creative Commons Attribution 4.0 International licence. The audio and annotations were cut, relabelled and augmented for training.
• OpenSLR SLR28 simulated room impulse responses, https://www.openslr.org/28/ (T. Ko, V. Peddinti, D. Povey, M. L. Seltzer and S. Khudanpur, ICASSP 2017), under the Apache License 2.0, for augmentation.

No corpus audio is distributed in this repository. No warranty is given by the licensors.