Module audio 

Audio intelligence + call recording (PROCESSOR-DESIGN §6.1).

This module is the home for the audio-intelligence processors that sit on the hot media path inside the cascaded pipeline — VAD, end-of-turn, noise filters, and mixers — alongside the existing call recorder. The analyzer traits (VadAnalyzer/TurnAnalyzer) are frozen in crate::service; the ONNX (ort) impls compile only under the vad-ort feature so flowcat-core stays dependency-light by default.

Module map:

• vad — VadAnalyzer impls (Silero behind vad-ort, a mock always).
• turn — TurnAnalyzer impls (Smart-Turn behind vad-ort, a mock always).
• strategy — turn start/stop/mute + interruption strategies (pure logic).
• filter — RNNoise/Koala audio-filter processors (Krisp deferred).
• mixer — audio mixers.

Call recording (existing)

Tap both legs (caller + bot), mono-mix, render to WAV bytes.

The recorder accumulates audio from each leg and produces a single mono WAV (via hound) suitable for upload to the object store (see DESIGN.md “Audio path”).

The two legs typically arrive at different sample rates (caller ≈ 8 kHz μ-law decoded, bot ≈ 24 kHz from Gemini Live), so each leg is buffered with its own rate and resampled to the recorder’s common sample_rate only at render time. Mixing is a plain per-sample sum clamped back into i16.

Re-exports

pub use filter::AudioFilter;

pub use filter::AudioFilterProcessor;

pub use filter::IdentityFilter;

pub use mixer::AudioMixer;

pub use mixer::MixerProcessor;

pub use mixer::SilenceMixer;

pub use mixer::SoundfileMixer;

pub use strategy::AlwaysInterrupt;

pub use strategy::AlwaysMute;

pub use strategy::ExternalTurnStart;

pub use strategy::FirstSpeechMute;

pub use strategy::FunctionCallMute;

pub use strategy::InterruptionStrategy;

pub use strategy::MinWordsInterrupt;

pub use strategy::MinWordsTurnStart;

pub use strategy::MuteStrategy;

pub use strategy::MuteUntilFirstBotComplete;

pub use strategy::SpeechTimeoutStop;

pub use strategy::TurnAnalyzerStop;

pub use strategy::TurnStartDecision;

pub use strategy::TurnStartStrategy;

pub use strategy::TurnStopDecision;

pub use strategy::TurnStopStrategy;

pub use strategy::VadTurnStart;

pub use turn::MockTurn;

pub use turn::TurnProcessor;

pub use turn::TurnSilenceTracker;

pub use vad::window_volume;

pub use vad::MockVad;

pub use vad::VadProcessor;

pub use vad::VadStateMachine;

Modules

filter

Input-leg audio-filter processors (noise suppression).

mixer

Output-leg audio mixers.

strategy

Turn start/stop/mute + interruption strategies (pure framework logic).

turn

End-of-turn / semantic-completion analyzer impls + the TurnProcessor.

vad

Voice-activity-detector impls + the VadProcessor.

Structs

AudioRecorder

Accumulates inbound (caller) and outbound (bot) audio on a shared wall-clock timeline and mixes them down to a single mono WAV byte buffer.