As we have seen before, audio features can be roughly divided to time domain features and frequency domain features.Let's take a deeper look at several types of time domain features and the ways to calculate them using python codes. Amplitude EnvelopeThe envelope of a soundwave is a curve outlining it's extremes.Thus, it gives us rough ideas of loudness.Since it uses the maximum value(amplitude)..

Time-domain features First, we should implement ADC procedure, which converts analog signal to digital signal.Next, there is a framing procedure, which is needed in order to generate perceivable audio chunks(note that our ear's time resolution is about 10ms)It's noticeable that each frame overlaps with another. The reason will be revealed later.After framing, feature computation and it's aggrega..

The conversion of analog signals in nature to digital signals is one of the most important procedures in audio signal processing. Since sound waveforms can be understood in terms of two-dimensional graphs, there are two keywords when implementing ADC. SamplingSampling is a reduction of continuous-time signal to a discrete-time signal.It extracts values(especially air pressure) from original sign..

SoundSound is something that is caused by a vibrating object.A vibrating object causes nearby molecules to vibrate either, and they (normaly) change the air pressure.It is a type of mechanical waves, which can be represented as y = sin(2 * pi * f *t - k) The important thing is that every audial features such as intensity, timbre ... can be obtained by this equation. Every information is compress..

본 주제의 필요성에 대해 Audio Classification, Speech Recognition, Audio Denoising, Music Information Retrieval 과 같은 오디오 분야의 machine learning 은 image 분야의 그것과는 조금 다른 특징을 가지고 있는 것 같다. 바로 Preprocessing 과정의 비중이다.  우리 눈은 수억 개의 시각세포로 이루어져 있고, 이 시각세포는 deep learning에서의 unit cell로 비유될 수 있는 것처럼 보인다. deep convolutional network를 기본 바탕으로 현재 image processing 분야는 인간을 능가하는 수준의 성능을 보여주고 있는데, 대단한 이미지 전처리 없이도 이러한 성과를 이루고 있다...