audio_processing 0.2.0 
audio_processing: ^0.2.0 copied to clipboard
Published 39 days ago • 
ksh.dev
ksh.devStateful provider-neutral PCM transforms, meters, and WAV utilities.
Changelog #
0.2.0 #
- Add
AecDelayEstimator: envelope cross-correlation delay estimation between a loopback reference and a microphone capture, with recency-weighted median smoothing and two lock tiers. Amplitude thresholds are float32-native; seeAecDelayEstimator.defaultMinNearStd. - Add
AecBlockAccumulatorplusfloatToPcm16/pcm16ToFloatfor chopping float32 captures into the exact 10 ms int16 blocks an echo canceller requires. - Add
mixTracksToMonoandpeakBucketsfor offline mixdown and waveform rendering of completed recordings of differing lengths, which the streamingAudioMixercannot express. - Adapted from Control Center (MIT © 2026 Samuel Alev); see
NOTICE. - Add
AudioRing: a fixed-capacity float32 ring addressed by a monotonic absolute write index, so a position reported by a VAD — counted from the start of the stream — indexes the audio directly no matter how often the buffer has wrapped. Ranges older than the capacity are clipped rather than wrapped, because returning samples from the wrong moment is indistinguishable from success. Defaults to 160,000 samples, ten seconds at 16 kHz. - Add
AdaptiveGainandAdaptiveGainProcessor: slow automatic gain control that amplifies quiet capture toward a 0.3 target peak (30x ceiling, 0.995 per-chunk running-peak decay, 1.05 deadband) and never attenuates. Speech models trained on normalized audio collapse on the 0.05-peak recordings real microphones produce; the slow decay is what keeps the noise floor from being pumped up between utterances. The processor keeps one running peak per stream and drops it on a discontinuity.
0.1.0 #
- Initial public release.
Metadata
0
likes
150
points
43
downloads
Documentation
Publisher
Weekly Downloads
Metadata
Stateful provider-neutral PCM transforms, meters, and WAV utilities.
Repository (GitHub)
View/report issues
Topics
License
MIT (license)