Mic bumps produce transient spikes that pass VAD onset detection but
contain no real speech — the model hallucinates "thank you" from them.
Added MIN_SPEECH_SECONDS (0.3s) filter to discard segments where the
actual speech portion is too short.
Added a Jupyter notebook (notebooks/audio_debug.ipynb) for real-time
audio visualization: streams RMS + peak amplitude into a live Plotly
FigureWidget, then provides post-hoc waveform inspection, segment
playback, and side-by-side segment comparison.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
