speech-prep

Speech-focused audio preprocessing for Rust.

Features

• Voice Activity Detection — dual-metric (energy + spectral flux) with adaptive thresholds
• Multi-format decoding — WAV, MP3, FLAC, OGG, M4A, Opus → 16kHz mono PCM
• Preprocessing — DC removal, high-pass filter, spectral noise reduction, normalization
• Chunking — speech-aligned segmentation with configurable duration and overlap
• Quality assessment — signal metrics for downstream processing gates

Quick start

cargo run --example vad_detect

Detected 1 speech segment(s):
  Segment 1: 0.290s — 1.540s  (confidence: 1.00, energy: 0.0362)

Usage

use std::sync::Arc;
use speech_prep::vad::{NoopVadMetricsCollector, VadConfig, VadDetector, VadMetricsCollector};

let config = VadConfig::default();
let metrics: Arc<dyn VadMetricsCollector> = Arc::new(NoopVadMetricsCollector);
let detector = VadDetector::new(config, metrics)?;

let segments = detector.detect(&audio_samples)?;
for seg in &segments {
    println!("{:.3}s — {:.3}s", seg.start_time.as_secs(), seg.end_time.as_secs());
}

Pipeline

Raw audio bytes
    │
    ▼
Format detection ─→ Decoding ─→ Resampling ─→ Channel mixing
  (format.rs)      (decoder/)    (16kHz)       (mono)
    │
    ▼
Preprocessing ─→ VAD ─→ Chunking
  (preprocessing/)  (vad/)  (chunker/)
    │
    ▼
Processed audio chunks with speech metadata

Modules

Configuration

use speech_prep::VadConfig;

let config = VadConfig {
    base_threshold: 0.02,      // energy threshold for speech detection
    energy_weight: 0.6,        // weight of energy vs spectral flux
    ..VadConfig::default()
};

use speech_prep::ChunkerConfig;

let config = ChunkerConfig::default(); // 500ms target chunks

License

MIT OR Apache-2.0