Choosing an ML model (RNNoise / Speex / DTLN / DeepFilterNet) turned the
browser's auto gain control off along with its noise suppressor, so the
call carried the mic's raw level: voices came out 10-20 dB quieter than
in the browser tier and DeepFilterNet "made people quiet". Measured with
speech through each model exactly as the call runs it, the models keep
the voice within ~1.5 dB (DTLN -0.1, DFN3 -1.7, RNNoise -1.3, Speex -1.5)
while cutting background noise 6-29 dB, so the loss was AGC, not the
models. Keep AGC on in every tier.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017PCb59uY3UKqDqt7B6hzd5