Audio editing system and audio editing method

Abstract

The audio editing system includes: a plurality of initial segmentation devices, which respectively initially segment the audio streams from multiple channels into a plurality of different paragraphs; Synthesize, and select the audio stream of the optimal channel from every two adjacent segmentation points, so as to obtain multiple initial segmentation segments, and fuse the obtained multiple initial segmentation segments to form a unified audio data file ; The audio clustering device, based on the algorithm of hierarchical clustering, carries out supervised clustering to a plurality of initial segmentation segments, and the initial segmentation segments belonging to the same nature are gathered into a category; the re-segmentation device utilizes the audio clustering device As a result of the clustering, train the hidden Markov model corresponding to each category, and perform Viterbi alignment segmentation on the unified audio file to obtain the re-segmented audio stream. Through the above high-precision speaker segmentation system, the accuracy of final speaker clustering can be improved.

Description

technical field [0001] The invention relates to the technical field of audio clustering, in particular to an audio editing system and an audio editing method. Background technique [0002] Speaker clustering is a specific application of clustering technology in speech signal processing. Its purpose is to classify speech segments so that each class only contains the same speaker data, and the data of the same speaker are merged into in the same class to obtain speaker-specific information. From an application point of view, speaker clustering technology can be applied to audio information management, retrieval and other fields. It facilitates speaker tracking in audio streams of conferences, voicemails, lectures, and news broadcasts, enabling structured analysis, understanding, and management of audio data. In particular, clustering algorithms are also of great practical value to speech recognition systems. Almost all automatic speech recognition systems today use adaptive ...

AI Technical Summary

Problems solved by technology

These two frameworks have their own advantages and disadvantages. The former inherits the error of the segmentation step without correction in the clustering process. Due to the limitations of the distance measurement method, errors will accumulate; Markov model, the initialization of the model is carried out by directly dividing the audio data into equal parts, the initial error introduced is relatively large, which brings certain problems to the convergence speed of the model, and because the hidden Markov model is based on the characteristics of frame classification , there will be a certain amount of error introduced without restrictions when performing segmentation. The general approach is to add a certain duration limit to the dwell time of each hidden Markov model. This approach brings a lot of flexibility to the system. big limitations

Images