sam-castro/mow-disfluency-classifier

Mów Disfluency Classifier (ONNX INT8)

Quantized ONNX export of arielcerdap/modernbert-base-multiclass-disfluency-v2 for on-device disfluency removal in Mów, a macOS voice-to-text app.

What it does

Tags each word in a speech transcript as one of:

Label	Meaning	Action
O	Fluent	Keep
FP	Filled pause (um, uh, er)	Remove
RP	Repetition (the the)	Remove
RV	Revision / self-correction	Remove
PW	Partial word	Remove

Model details

• Base model: ModernBERT-base (150M parameters)
• Task: Token classification (5 classes)
• Training data: FluencyBank corpus
• Accuracy: 93.2%, F1 0.99 on filled pauses, 0.90 on repetitions
• Format: ONNX, INT8 dynamic quantization
• Size: ~143 MB
• Inference: ~5-50ms per sentence on Apple Silicon via ONNX Runtime

Files

• DisfluencyClassifier.onnx — quantized ONNX model
• tokenizer.json — HuggingFace tokenizer configuration
• tokenizer_config.json — tokenizer metadata
• label_map.json — class ID to label mapping

How to regenerate

From the Mów repo root:

./scripts/export-disfluency-model.sh

This downloads the original PyTorch model from HuggingFace, exports to ONNX, and quantizes to INT8.

License

CC BY 4.0 (same as the base model). Attribution: arielcerdap/modernbert-base-multiclass-disfluency-v2.