Ai I see and feature extraction MFCC algorithm understanding

First, artificial intelligence

From Lenex handwritten numeral recognition, alexnet image recognition, to driverless cars, to Alpha go, alpha go zero, Ai has undoubtedly become the hot technology of the moment. So what is AI? To be straightforward, AI is the intelligence that makes the machine owner. In order for the machine to have intelligence, the scientists have set a plan for the machine, from the perspective of how people identify, think and solve problems.

Neural networks are one of the best examples: in the early days, scientists invented airplanes from the wings of birds, and now scientists are starting with how people think, how the brain works, and then invented neural networks. Below we want to draw out the focus of this blog MFCC feature extraction algorithm, which is based on human behavior and invented.

Second, MFCC algorithm

MFCC is a set of algorithms developed to accomplish sound recognition, based on how people recognize sound. First clear Four points:

1. Most of the voice signal information is included in the low frequency component;

2. Most of the voice signal information is included in the low-amplitude section;

3. The sound level of the human ear is not linearly related to the sound frequency, but it is linearly proportional to the logarithm of the sound frequency;

4. People can not distinguish between all frequency components, only two frequency components of a certain bandwidth (below 1000hz, bandwidth constant 100hz;1000hz above, bandwidth and center frequency exponential relationship), human can distinguish, otherwise people will have two tones to listen to a, which is called Shielding effect , the bandwidth is called critical bandwidth ; (center frequency: The sound is mainly related to frequency, because audible audio is too wide (from 20Hz to 20000Hz), in order to facilitate the frequency analysis, it is divided into several segments, called the frequency range. The geometric mean of the upper and lower frequencies of each frequency range is called the center frequency of the frequency range)

　　MFCC, to some extent, simulates the processing characteristics of human ear to speech, and uses the research results of ear auditory perception, which improves the performance of the speech recognition system.

MFCC is a feature that is widely used in automatic speech and speaker recognition.

If you give us a voice now, we first get its spectral envelope (the smoothed curve that connects all the resonant peaks, the resonant peaks carry the recognizable attributes of the sound, like the human identity card), but for humans, the perception of human hearing is focused on specific areas rather than the entire spectrum envelope, and Mel Frequency analysis is based on the human auditory perception experiment. Experimental observations have found that the human ear, like a filter group, focuses only on certain frequency components. It has a lot of filters in the low frequency region, and less in the high frequency region.

The characteristics of human ear hearing are consistent with the growth of Mel frequency, and the Mel filter can extract the same characteristics as people. (GFCC is based on the GT filter)

Ai I see and feature extraction MFCC algorithm understanding