Stutter Detection Using Prosody

This project implements a robust speech feature extraction pipeline for automated stutter detection, focusing on prosodic and acoustic features. It utilizes the SEP-28k dataset and extracts fine-grained frame-level and clip-level features to distinguish between fluent and stuttered speech.

Results

• The confusion matrix indicates balanced classification accuracy across both classes:

  • 84.7% of No Stutter samples were correctly identified.
  • 86.7% of Stutter samples were correctly identified.

• The ROC curve achieved an AUC of 0.915, showing excellent ability to distinguish between stutter and non-stutter speech.

• The prediction probability distribution shows:

  • Most No Stutter samples clustered near 0.
  • Most Stutter samples clustered near 1.

• This indicates the model makes confident and reliable predictions with limited ambiguity.

Quick Start: Data Acquisition

To set up the project environment, run the following commands to download the dataset and pre-extracted features.

1. Download SEP-28k Dataset (Kaggle)

Requires Kaggle CLI configured with your API key.

mkdir -p dataset
kaggle datasets download -d vudominhgiang/sep-28k-maintained -p dataset/
unzip dataset/sep-28k-maintained.zip -d dataset/
rm dataset/sep-28k-maintained.zip

2. Download Extracted Features (Hugging Face)

Requires huggingface-cli.

mkdir -p output
huggingface-cli download bropal/stutter_detection_prosody --local-dir output/ --repo-type space

Manual downloads are available at:

• Kaggle: SEP-28k Maintained
• Hugging Face: Stutter Detection Prosody (Output)

---

Features & Methodology

The pipeline extracts several layers of speech features based on prosodic dynamics and spectral characteristics:

1. Vowel Onset Point (VOP) Detection

Faithful implementation of Mary & Yegnanarayana (2008), using:

• LP Residual & Hilbert Envelope.
• Gabor filter convolution for evidence enhancement.
• Peak picking with dynamic thresholds and F0-based spurious reduction.

2. Syllable Prosody (7 Parameters)

Extracts prosodic dynamics between VOPs:

• Duration: Syllable duration and voiced duration.
• Intonation: Peak F0, Distance of peak from VOP ($D_p$), and F0 range ($\Delta F_0$).
• Tilt: Amplitude tilt and Duration tilt parameters.
• Stress: Delta Log Energy.

3. Acoustic & Spectral Features

• MFCCs: 13 coefficients + $\Delta$ + $\Delta\Delta$ (39 dims).
• Voice Quality: Jitter, Shimmer, CPP (Cepstral Peak Prominence).
• Prosody Contours: F0 (RAPT-inspired autocorrelation), RMS Energy, Zero-Crossing Rate.
• Pause Features: Silence duration, pause count, and max pause length (targeting 'Block' stutters).

---

This project is licensed under the GPL-3.0 License.