mcelp-codec

A Rust implementation of the Mitsubishi CELP speech codec (三菱CELP方式音声コーデック).

The crate ships both halves of the codec. Each stage is written as the algorithm it implements, and both the encoder and the decoder reproduce the reference bit for bit — the regression suite replays the reference's own output, stage by stage.

Frame format

• Audio — raw 8 kHz mono μ-law PCM, 320 bytes (40 ms) per frame.
• Bit stream — 18 bytes per frame, i.e. 3.6 kbit/s.
• Container — .mcelp text file, one frame per line, each line a 36-character hex string encoding one 18-byte frame.

A trailing partial frame is dropped, as the reference implementation drops it.

Library

[dependencies]
mcelp = "1.0"

Encoder and Decoder are the entry points. Both carry state between frames, so one instance has to see a stream in order.

use mcelp::{Decoder, Encoder, FRAME};

let speech: Vec<u8> = read_ulaw();          // 8 kHz mono μ-law

let (mut encoder, mut decoder) = (Encoder::new(), Decoder::new());
let mut out = Vec::new();
for block in speech.chunks_exact(FRAME) {
    let frame = encoder.encode(block.try_into().unwrap());   // [u8; 18]
    if let Some(pcm) = decoder.decode(&frame) {              // [u8; 320]
        out.extend_from_slice(&pcm);
    }
}

Decoder::decode returns None only for an in-band reset frame, which resets the decoder instead of producing audio; an erased frame is concealed and still yields samples. Decoder::decode_linear gives the same samples as 16-bit linear PCM instead of μ-law.

mcelp::bitstream exposes the frame's fourteen parameter fields (unpack/pack) and the hex container (parse_hex_line/to_hex_line) for callers that need to work below the frame.

The remaining modules are the stages the two entry points are built from. They are public so that each can be replayed against the reference on its own, which is what the test suite does; they are not a stable interface.

Command line

cargo build --release

mcelp_encode reads raw μ-law on stdin and writes hex frames on stdout; mcelp_decode does the reverse.

cargo run --release --bin mcelp_encode < speech.ulaw > speech.mcelp
cargo run --release --bin mcelp_decode < examples/female.mcelp > female.ulaw

Or pipe straight into a player — with sox:

cargo run --release --bin mcelp_decode < examples/female.mcelp | play -t ul -r 8000 -c 1 -

…or with ffplay:

cargo run --release --bin mcelp_decode < examples/female.mcelp | ffplay -f mulaw -ar 8000 -ac 1 -

Tests

cargo test --release

Each stage is checked against data captured from the reference, and tests/api.rs pins the byte-level round trip. tests/endtoend.rs and tests/encode_endtoend.rs carry the whole of both bundled examples, so a change that moves a single bit fails the suite.

Continuous integration

Every push and pull request runs, in .github/workflows/ci.yml:

Each test job also runs the two command line tools over the bundled examples and compares the bytes, so the codec is shown to be bit exact on every platform, not merely to compile there.

Tagging a commit v<version> runs .github/workflows/release.yml, which checks the tag against Cargo.toml, publishes to crates.io and attaches prebuilt tools to a GitHub release.

Documentation

docs/ holds a technical description of the codec, stage by stage — frame structure, the two pipelines, each coding stage, and the fixed-point model everything runs on.

• English
• 日本語

License

MIT — see LICENSE.