Mel Frequency Cepstral Coefficients

A common pre-processing step in Machine Learning with audio signals is the application of a Mel Frequency Cepstral Coefficients (MFCC) transformation. They compress the signal to a very small number of coefficients (around 16 for every 10ms) and decorrelates the signal to express only the transmission function (e.g. only the formants of a utterance not the pitch). This makes them very popular in Automatic Speech Recognition (ASR), Room Classification, Speaker Recognition etc.

Usage

Add this to your Cargo.toml

[dependencies]
mfcc = "0.1"

The library can use two different FFT libraries. Either use rustfft (a pure rust FFT implementation) with the standard feature fftrust or use fftw (a popular FFT library) with

[dependencies.mfcc]
version = "0.1"
default-features = false
features = ["fftextern"]

A rough benchmark shows that their performance are comparable, for FFTW:

test tests::bench_mfcc ... bench:     123,959 ns/iter (+/- 22,979)

For rustfft:

test tests::bench_mfcc ... bench:     162,603 ns/iter (+/- 35,914)

How it works

First you need to segment you audio data in chunks of around 10ms-20ms (max 1024 samples for 48kHz). From these you can calculate the MFCC coefficients with

use mfcc::Transform;

let mut state = Transform::new(48000, 1024);
let mut output = vec![0.0; 16*3];

state.transform(&input, &mut output);

License

Licensed under either of

• Apache License, Version 2.0, (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

Contribution

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in the work by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.