A dynamic sound feature extraction method based on cosine similarity

Abstract

The invention discloses a method for extracting dynamic sound features based on cosine similarity, comprising the following steps: S1: using pre-emphasis, framing and windowing methods to preprocess the voice signal into a time-domain signal, and using a window function to decompose the voice signal Be the frame of certain length; S2: voice time domain signal is converted into frequency domain signal, obtain the discrete cosine inverse transform (IDCT-Cepstrum Coefficient) cepstral coefficient of 320 dimensions of every frame voice domain signal; S3: to the frequency domain signal of voice Domain signal is carried out the cosine similarity calculation between adjacent dimensions; S4: Find two columns of maximum adjacent dimensions of cosine similarity to merge; S5: Repeat operation S3 to S4 to reduce the dimensionality of 320-dimensional speech audio domain signal to 14 dimensions Speech domain feature; S6: represent the speech feature in the form of a histogram.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a dynamic sound feature extraction method based on cosine similarity. Background technique [0002] Speech recognition consists of three parts: speech feature extraction, speech recognition model building and speaker recognition. Among them, the speech feature extraction is very important in the whole speaker recognition process. Effectively extracting the speech features that represent the essential characteristics of the speaker will make the speech classification and recognition of the model more accurate and the recognition rate higher. Currently commonly used features include MFCC (Mel Frequency Cepstral Coefficient), Fbank (Filterbank feature), PLP (Linear Predictive Coding), etc. Currently, MFCC is used as a speech feature in this field. [0003] Although MFCC is the most popular sonic representation, it suffers from two major drawbacks. First, they are estima...

AI Technical Summary

Problems solved by technology

First, they are estimated with a window function from 10ms to 50ms and thus cannot capture long-term features in the signal

Second, MFCC believes that the adjacent frames of the speech signal are independent of each other, which leads to the loss of the dynamic characteristics of the speech signal during the feature extraction process.

Images