Real-time role separation transcription method, device and system

Abstract

The invention discloses a real-time role separation transcription method, which comprises the following steps of: detecting whether an acquired left channel sound signal and an acquired right channelsound signal contain a voice segment or not, and if the voice segment is detected, extracting a left channel voice segment and a right channel voice segment corresponding to the voice segment; judgingwhether the voice segment is a single-side speech or a double-side speech based on the phase difference, the amplitude difference of the left channel voice segment and the right channel voice segmentand fundamental frequency detection if the voice segment is the single-side speech, judging whether a speaker is located at the left side or the right side; if the speaker is located at the left side, clustering the left channel voice segments to form a left clustering center; if the speaker is located on the right side, clustering the right channel voice segments to form a right clustering center; separating left and right side voice signals contained in the left channel voice segment and the right channel voice segment if the voice segment is the double-side speech; and sending the separated left side voice signal and right side voice signal to a voice recognition engine. According to the method, roles can be accurately separated. The invention further discloses a real-time role separation transcription device and system.

Description

technical field [0001] The invention relates to the technical field of speech recognition, in particular to a real-time role-based transcription method, device and system. Background technique [0002] Nowadays, in the process of one-to-one window service, the service quality requirements for customer service are getting higher and higher. In such occasions (such as telecom business halls and automobile 4S stores), customer service and customers are often located on the inside and outside of the service counter or window. , to carry out one-on-one dialogue service activities; now there are various assessments on the service quality of customer service, including service words, professional terminology expressions, service attitudes, emotions, and whether to guide customers correctly. The content is clearly collected, and the back-end is used for speech recognition. The back-end processing generally uses a real-time transcription system, and then organizes and analyzes the tr...

AI Technical Summary

Problems solved by technology

[0003] The current front-end voice collection is often based on deep learning solutions or requires the use of front-end recognition equipment; deep learning-based solutions require a local or cloud voice recognition server to perform voice-to-text transcription on the voice streams collected in real time. In addition to identification, roles often need to be processed in the cloud. It is an end-to-end one-stop solution. This solution is not suitable for simple conversations between the two parties. It is generally suitable for meetings where multiple people speak mode; and this solution has high requirements for the hardware configuration of the cloud or local speech recognition engine, the accuracy of role classification is related to the frequency of the speaker, if in some scenarios, a speaker simply said "yes" , "Okay", and there is no speech in the whole speaking process. At this time, the accuracy of role classification is very low. Therefore, this scheme adopts a large and comprehensive scheme to solve some specific problems, and its existence cannot be analyzed locally. The specific situation to do very specific corresponding defects

[0004] With the help of front-end recognition equipment, microphone arrays are generally used to form two directions of 180 degrees to each other. When transcribing, the two speakers need to speak face to face. The device is placed in the center of the connection between the two parties, which makes people feel unfriendly. The voices collected at both ends are compared in real time. Generally, a solution with a high real-time volume is adopted to determine the voice stream of a certain character; There is often no way to deal with the scene of speaking. This solution is to judge the energy of the speakers at both ends. Which side has the loudest volume will classify the voice as that side. If one side is not a voice but a loud volume If there is no noise, the "speaking right" will be taken to this side by mistake; if the speakers at both ends speak at the same time, the strategy adopted in this scheme is to always select the loud voice on one side and keep it until the end. This is a simple and rude behavior. This solution does not have a good estimate of the problem that this kind of "simultaneous speaking" should be treated specially, and there is a defect that it cannot perform special treatment on "simultaneous speaking", and cannot obtain a better recognition effect

Images