Speech & Audio ML — Keyword Spotting

Overview

A keyword spotting (KWS) system that recognizes the spoken words 'yes', 'no', 'stop', and 'go' plus a silence/unknown class, built entirely with classical signal processing instead of deep learning. The goal was to show that hand-crafted time- and frequency-domain features paired with traditional classifiers can hit competitive accuracy while staying small and fast enough for embedded or System-on-Chip deployment. Completed for ASU's EEE 598 Speech and Audio Processing course with a project partner. **Highlight:** 84.67% test accuracy

Key Achievements

• Achieved 84.67% test accuracy on the 5-class task using an SVM (RBF kernel, ECOC) with a 30-dimensional feature vector, exceeding the project's 75% target
• Built a full audio-to-decision pipeline from scratch (resampling to 16 kHz, energy-based silence trimming at -45 dB, peak normalization, 1.0 s length standardization) in both MATLAB and Python
• Engineered a 30-D feature vector combining short-time energy, zero-crossing rate, and 13 MFCCs (mean and std per coefficient via STFT -> Mel filterbank -> log -> DCT), with delta/delta-delta features in the Python pipeline
• Ran a comparative study and feature ablation (k-NN vs SVM; MFCC-only vs MFCC+Energy+ZCR), finding SVM beat k-NN by ~5.2% and MFCCs alone already captured most discriminative power (+0.6% from time-domain features)
• Evaluated noise robustness across SNR levels with both white and real background noise, showing graceful degradation (84.67% clean down through ~57-69% at 20 dB) and >50% accuracy to roughly 10 dB SNR
• Demonstrated embedded feasibility: model under 1 MB, ~120 bytes per clip, and inference under 10 ms per 1-second clip, suitable for ARM Cortex-M4-class processors

Approach

Audio from the Google Speech Commands v2 dataset was preprocessed (16 kHz resample, silence trimming, normalization, fixed 1 s length) and split into train/validation/test sets. Features were extracted over 20-25 ms Hamming windows with a 10 ms hop: short-time energy, zero-crossing rate, and MFCCs, aggregated to a fixed-length per-clip vector. k-NN (k in {1,3,5,7}) and SVM (RBF kernel, grid-searched kernel scale and box constraint) classifiers were trained and tuned on the validation set, then evaluated on a held-out test set and under added noise. The system was implemented in MATLAB and re-implemented as a four-phase Python pipeline (librosa, scikit-learn).

Tools & Technologies

Gallery

Results

The best configuration (SVM with MFCC+Energy+ZCR features) reached 84.67% test accuracy on five classes, with the 'silence' class near-perfect (99.5%) and the main confusion between acoustically similar 'no' and 'go'. Under noise the system degraded gracefully, staying above 50% accuracy down to roughly 10 dB SNR. Full metrics, confusion matrices, and noise tables are in the final report.