Crate aerosieve 

AeroSieve: a batch audio curation pipeline.

Audio files go in; WebDataset shards of clean speech clips come out. The stages follow Emilia-Pipe:

1. Decode (decode): any supported container to mono f32 at one sample rate, loudness-normalized.
2. Segment (segment): a hysteresis VAD cuts the audio into variable-length speech clips.
3. Score (acoustic): each clip is measured and judged against a FilterPolicy.
4. Shard (shard): accepted clips are written to tar shards with a manifest.

use aerosieve::{Pipeline, PipelineConfig};

let config = PipelineConfig::new("data/staging", "data/clean");
let mut pipeline = Pipeline::new(config)?;

pipeline.process_file("recording.wav".as_ref(), Some("hello world"))?;

let stats = pipeline.finish()?;
println!("retention {:.1}%", stats.retention_rate() * 100.0);

Re-exports

pub use aerosieve_acoustic as acoustic;

pub use aerosieve_decode as decode;

pub use aerosieve_lexical as lexical;

pub use aerosieve_segment as segment;

pub use aerosieve_shard as shard;

Structs

DecodeConfig

DecodedAudio

Mono audio at a known rate, with provenance about its source.

EnergyDetector

Level-based detector with no model and no dependencies.

FileOutcome

What happened to one input.

FilterPolicy

Thresholds a buffer must meet to be kept.

Inspector

Decodes, segments, and scores audio without writing anything.

Pipeline

The curation pipeline: an Inspector whose accepted clips are written to shards.

PipelineConfig

PipelineStats

Corpus-level totals.

QualityMetrics

Measurements for one buffer.

Segment

A half-open range of samples, [start_sample, end_sample).

SegmentConfig

Tuning for Segmenter.

SegmentOutcome

What happened to one segment.

ShardConfig

ShardStats

Enums

PipelineError

RejectReason

Why a buffer was rejected.

Traits

SpeechDetector

Scores consecutive windows of audio with a speech probability.