feat: filter short audio segments (mic bumps) and add debug notebook

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>

pyproject.toml

@@ -27,3 +27,11 @@ build-backend = "hatchling.build"
 
 [tool.hatch.build.targets.wheel]
 packages = ["src/cohere_transcribe"]
+
+[dependency-groups]
+dev = [
+    "anywidget>=0.11.0",
+    "ipywidgets>=8.1.8",
+    "jupyterlab>=4.5.7",
+    "plotly>=6.7.0",
+]

src/cohere_transcribe/vad.py

@@ -10,6 +10,7 @@ PRE_ROLL_FRAMES = 6       # ~0.3s of audio before speech onset
 DEFAULT_SILENCE_FRAMES = 16  # ~0.8s of silence to end a segment
 SPEECH_ONSET_FRAMES = 3   # ~150ms of speech to trigger
 MAX_SPEECH_SECONDS = 30   # force chunk boundary
+MIN_SPEECH_SECONDS = 0.3  # discard segments shorter than this (mic bumps, clicks)
 
 
 def pause_seconds_to_frames(seconds: float) -> int:
@@ -67,7 +68,10 @@ class VADStateMachine:
 
         segment_duration = len(self.segment) * FRAME_SIZE / SAMPLE_RATE
         if self.silence_frames >= self.silence_limit or segment_duration >= MAX_SPEECH_SECONDS:
-            result = (self.segment_start_time, np.concatenate(self.segment))
+            speech_duration = segment_duration - self.silence_frames * FRAME_SIZE / SAMPLE_RATE
+            result = None
+            if speech_duration >= MIN_SPEECH_SECONDS:
+                result = (self.segment_start_time, np.concatenate(self.segment))
             self.speaking = False
             self.speech_frames = 0
             self.silence_frames = 0