Root causes of disconnection and slow transcription:
- AudioWorklet was firing every 128 native samples (~48 kHz), sending
~375 tiny WebSocket messages/sec. Server flooded with tiny frames
during silence → keepalive ping timed out → connection dropped.
- JS resampling 48 kHz → 16 kHz added CPU overhead on every chunk.
- Audio started on ws.onopen before server sent SERVER_READY, so early
frames were dropped.
Fixes:
- audioWorklet.js: accumulate 4096 samples before posting (256 ms/chunk
at 16 kHz, ~4 messages/sec), transfer ArrayBuffer zero-copy.
- transcriptionService: AudioContext({ sampleRate: 16000 }) — browser
handles native resampling, no JS resampling needed. Remove
resampleTo16kHZ entirely.
- Wait for SERVER_READY message before calling setupAudioProcessing().
- Send 'END_OF_AUDIO' string on stop so server can finalise last segment.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
kcar
4139eb8fd3