Real-time detection method of initial position of human sound in song

Abstract

The invention belongs to the technical field of digital audio processing, and relates to human sound detection problems, in particular to an estimating method of the initial position of the human sound in a song. The method comprises the steps that an audio is framed by using a high overlapping long window, the dynamic characteristics of two dimensions including inter-frequency (a frequency domain) and inter-frame (a time domain) are extracted, and the audio characteristics in the initial sound production stage are effectively captured. Through learning of a starting-point fragment of the song, the song is divided into musical instrument sound and human sound or musical instrument and human mixed sound, the initial position of the human sound can be accurately estimated, and the method ishigh in human sound / musical instrument sound determining fault tolerance. The method is simple in algorithm, high in processing speed, capable of being widely used in transmission of programs in broadcasting station, automatic digital media management and the like.

Description

Technical field [0001] The invention belongs to the technical field of digital audio processing, and relates to the human voice detection problem, and specifically is a method for estimating the starting position of the human voice in a song. The method can be applied to real-time human voice position marking of broadcast audio. Background technique [0002] A song is usually composed of pure accompaniment and singing. The pure accompaniment part is purely produced by the accompaniment instrument and does not contain the vocal part, while the singing part is the superposition of the vocal and accompaniment music. In the current management of digital media data, it is often necessary to mark the starting position (starting point) of the human voice in a song. The starting point of the vocal information has many uses. For example, in the live program of a radio station, the starting position of the vocal can help the host control the speaking time, set the cross-fade of adjacent so...

AI Technical Summary

Problems solved by technology

According to the previous analysis, due to the influence of musical instruments in the song on the human voice, many common vocal features become blurred or even invalid, and the multi-feature combination has little effect, which is not enough to make up for the calculation cost brought by the introduction of multiple features. ; In terms of classifiers, the difference in the effect of each classifier is not very obvious; in addition, the ANN method with relatively good effect still has disadvantages such as long training time and large number of samples required

In short, in the absence of effective feature expression for the instrument-human voice mixture, the accuracy of human voice detection is currently lower than 90%, which makes it difficult to estimate the accuracy of the starting point of the human voice to meet practical requirements

Images