aerosieve-decode

Standardization: any audio file in, mono f32 at one sample rate and one loudness out.

1. Demux and decode with symphonia. The container is identified by probing its bytes; the file extension is only a hint.
2. Downmix by averaging channels.
3. Resample with rubato's SincFixedIn. Skipped when the source is already at the target rate.
4. Normalize RMS level toward the target, with the applied gain clamped.

use std::path::Path;
use aerosieve_decode::{decode_file, DecodeConfig};

let audio = decode_file(Path::new("interview.mp3"), &DecodeConfig::default())?;
println!(
    "{:.2}s from {} Hz / {} ch, gain {:+.2} dB",
    audio.duration_s(),
    audio.source_sample_rate,
    audio.source_channels,
    audio.applied_gain_db,
);

decode_reader does the same for any Read + Seek source, such as an in-memory buffer.

Configuration

The loudness defaults follow stage 1 of Emilia-Pipe (arXiv 2407.05361). The clamp matters: an unbounded normalizer amplifies a quiet recording's noise floor along with its speech. Level is plain RMS, not LUFS.

Formats

Whatever symphonia's all feature enables: WAV, MP3, FLAC, Ogg Vorbis, MP4/M4A (AAC, ALAC), MKV/WebM, and CAF. All decoders are pure Rust.

Duration is preserved exactly

Resampled output is round(input_len × target_rate / source_rate) samples long. That takes two deliberate steps, both covered by tests:

• The trailing partial chunk is flushed through process_partial_into_buffer; otherwise the end of the file is dropped.
• The filter's output_delay() frames of ramp-up are discarded from the front and the same amount is flushed from the back; otherwise the audio is offset and its tail is lost.

Errors

Malformed packets are skipped, as symphonia recommends. A file that cannot be opened, is not audio, or has no usable track returns a DecodeError.