MFCCs vs. Mel spectrograms for frame-level voiced/unvoiced detection with a small CNN—compact cepstra outperform higher-dimensional inputs on limited data.
My TER Report (PDF): When you say “zebra,” the /z/ rides on vibrating vocal folds (voiced) while the /s/ in “snake” is all hiss (unvoiced). Lots of tools—pitch trackers, ASR, TTS—work better if we can flip a tiny frame level switch: voiced...