Audio ML Training, Evaluation & Ablations

What I implemented

• Implemented acoustic feature extraction using RMS, zero crossing rate, spectral centroid and spectral flatness.
• Trained a logistic speech-activity classifier with standardized inputs, gradient descent and regularization.
• Ran RMS-only, multi-feature and noise-augmented experiments on fixed synthetic holdouts, recording confusion counts, precision, recall and F1.

Data and execution path

1. 01

   Deterministic voiced/noise frame generation

2. 02

   Held-out clean and noisy splits

3. 03

   Feature extraction and train-set normalization

4. 04

   CPU gradient-descent classifier training

5. 05

   Controlled feature/augmentation ablations

6. 06

   Reproducible metrics and error analysis

Experiment and results

In the seeded synthetic noisy holdout, the clean-trained multi-feature model scored F1 0.361, while the noise-augmented model scored F1 0.867; the RMS-only baseline scored F1 0.689. These are outcomes on generated waveforms only, not general speech benchmarks. The large drop before augmentation is precisely why controlled evaluation is useful.

These results are from the specified local dataset or offline rendering pipeline. I keep them separate from real speech-corpus and live-call metrics because the populations and failure modes differ.

System-level constraints

• Dataset licensing, consent, speaker-disjoint splits and class balance determine what a classifier score can legitimately support.
• Corpus manifests, preprocessing, seeds, model checkpoints and evaluation artifacts need stable identities for reproducible comparisons.
• Architecture comparisons need controlled ablations and uncertainty estimates; synthetic F1 does not substitute for live speech-quality measurements.
• Distributed GPU training introduces sharding, synchronization, mixed precision and checkpoint recovery; the local single-device baseline says nothing about their throughput or correctness.