This repository contains implementations of audio feature extractions including spectrogram, mel-scale spectroram, mel-frequency cepstrum coefficients (MFCC) with numpy. Although these audio data preprocessing implementation are provided by many libraries, I wrote a code from scratch to enhance the theoritical understanding.
In this work, three audio samples taken from YouTube are used. The audio is a speech from the U.S. president Joe Biden, EU commision president Ursula von der Leyen, and Swedish young activist Greta Thunberg.
Each audio signal looks like below.
The spectrogram is a time-frequency domain speech features. The spectorgram contains time, frequency and amplitude information of an audio signal.
Extracted spectrogram looks like below.
The mel-scale spectrogram is a spectrogram in mel-scale (imitation of human auditory perception scale). Each spectrum is wrapped with mel-scale filter banks. As a result, the dimension is significantly reduced.
MFCC is the features containing the spectrum contor over time. As it is onle spectrum envelope, the dimension is more reduced compared to mel-scale spectrogram.
You need to have numpy and scipy installed at your environment.
python feature_extraction --audio_dir audio_samples --feature_type "spectrogram"
You can specify the direcotry cotaining audio files with audio_dir. feawture_type is a type offeature to be extracted from audio files.
The image of signals and spectrograms are saved under the plots direcotry. If you run the above command, you will get the dimension of the specified feature.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent