Voiced/Unvoiced Classification in Speech (TER Report)

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 or unvoiced.

For my TER, I set myself a practical question:

With modest data and a small CNN, is it smarter to feed the model compact MFCCs or full Mel spectrograms?

I had three speakers recorded at 16 kHz. I chopped audio into 25 ms frames (10 ms hop), and built input windows of ~0.5 s (50 frames). To avoid any train/val bleed, I used 7-fold cross-validation by file (no file ever split).

Why I Thought Mel Might Win (…and Why It Didn’t)Copy link to section

Mel spectrograms are like a high-res photo of time-frequency structure; they keep fine harmonic patterns that seem perfect for detecting voicing. MFCCs, meanwhile, compress the spectrum down to a handful of coefficients—a tidy spectral envelope.

I tried both in the same lightweight CNN (two conv blocks → max-pool → dropout → dense). I tuned learning rate and dropout a bit (Adam, early stopping). Nothing fancy on purpose: I wanted to see which representation helps most when the model and data are small.

Results (validation, mean over 7 file-level folds):

MFCCs won by ~3–4 points. Increasing Mel bands raised training accuracy but didn’t lift validation—a neat example of overfitting when you add dimensionality without adding data or model capacity.

What I LearnedCopy link to section

• Small data loves compact features. MFCCs funnel the cues that matter for voicing (periodicity/spectral envelope) into a size the CNN can actually digest.
• Boundaries are hard. Frames at voiced↔unvoiced transitions confused the model most; soft labels or ignoring ±1–2 frames could help.
• Typical mistakes made sense. Strong high-frequency /s/ sometimes looked “voiced”; breathy/low-HNR vowels sometimes looked “unvoiced.” Adding HNR/CPPS features would likely reduce both errors.
• Protocol matters. File-level CV avoids leakage across segments of the same file, but it’s still within-speaker. A future leave-one-speaker-out pass would test cross-speaker generalization properly.

If I Had Another WeekCopy link to section

• Add Δ/ΔΔ to MFCCs and/or combine MFCC + HNR/CPPS.
• Use a slightly richer head (BatchNorm, global average pooling, tiny L2).
• Try threshold tuning (0.5 isn’t sacred) for better F1 or balanced accuracy.
• Do speaker-disjoint evaluation and light augmentation (time masking, mild noise).

TakeawayCopy link to section

If you’re building a frame-level voiced/unvoiced switch with limited data and a small CNN, start with MFCCs. They’re the right abstraction at the right scale. Mel spectrograms can surpass them—but usually when you also scale data, model, and regularization.

Tech stack: Python, TensorFlow/Keras, scikit-learn, librosa · 25 ms / 10 ms · 16 kHz
People: TER @ Université Paris-Saclay · Mentor: Marc Evrard · With input from Albert Rillard & Saulo Santos