feat: Sound Event Detection module (Goal 1) with YAMNet and India-specific labels

[bhuvan-somisetty]

Adds a focused, modular implementation of Goal 1 (Sound Event Detection) from issue #2.

The core of this module is SoundEventDetector in src/audio/detector.py. It loads YAMNet from TF-Hub, runs it frame-by-frame on a 16 kHz mono WAV file, suppresses speech classes, maps surviving events to CC labels, and merges adjacent same-label detections within a 0.5 s gap. All heavy imports (tensorflow_hub, soundfile) are deferred to method bodies so the module is importable without any ML stack — this keeps the test suite fast and dependency-free.

The label layer in src/audio/labels.py does a few things worth calling out. The SPEECH_LABELS frozenset covers all 26 YAMNet classes that represent human speech or breath sounds, which is the main source of over-captioning. The LABEL_MAP goes beyond generic AudioSet names and includes India-specific sounds that come up in regional-language content: Fireworks maps to [firecrackers] (Diwali crackers), Tabla and Dhol map to their own labels, and Temple bells gets its own entry. Unmapped non-speech events fall back to [<lowercased class name>] rather than being silently dropped.

The test suite covers all of this without requiring TensorFlow or soundfile installed. The detector tests inject a fake model directly and mock soundfile at the sys.modules level, so the 26 tests run in under 0.3 s.

Tested with pytest tests/ -v — 26/26 passed.

Closes #2