I was the Audio and Features lead. I built the signal-processing pipeline: audio decoding and validation (16 kHz mono, clear error codes for short, silent or mismatched input), word-level alignment with a deterministic fallback aligner and an optional WhisperX backend, and feature extraction (energy in dB, F0 via Praat autocorrelation, pitch in semitones relative to the speaker median, per-word and pause features). I made the normalization speaker-relative so comparisons don't rely on raw cross-speaker pitch or loudness, and I checked that the same input and config give the same output.
Log in or sign up for Devpost to join the conversation.