Tutorial 05 — Audio lane (synthetic spectrograms)

Master guide: ../GUIDE_E2E.md · Prev: 04 · Next: 06

Goal: Prove STE + binary CNN pattern on audio features, offline-friendly.

Honest scope

Classic BNNs are not production ASR/TTS. For real speech use INT8 Whisper / ONNX Runtime / vendor NPU INT8. This demo shows packing/QAT patterns on spectrogram-like inputs.

Quick run (always works offline)

bnn train-audio --epochs 5 --n-train 800 --n-test 200 --seed 0

Or:

python scripts/train_audio.py --epochs 5 --seed 0

Committed golden: results/audio_synth.json (verify: bnn repro).

Outputs: results/audio_synth.json (+ .md).

Pipeline

1. Synthetic musical tones (8 pitch classes) — reproducible, no download required.
2. NumPy STFT → cheap mel-like pool (bnn/audio/features.py).
3. FP CNN vs Bi-Real-style binary CNN (bnn/audio/models.py).

Optional NPZ cache under data/audio_cache/ after first run.

Smoke test

pytest tests\test_audio_smoke.py -q

No network. Safe for CI.