Industry insight | is speech recognition really beyond human ears?

In recent years , With the development of artificial intelligence technology , The performance of speech recognition has been significantly improved . Many companies claim , The accuracy of speech recognition technology has reached 98% above , Is speech recognition really more effective than human ears ？

Of course, this conclusion cannot be drawn . After all, the human brain is the most accurate instrument in the world . There is a saying on the Internet that is very good ,“ Leaving the test set to say accuracy is like playing hooligans ”. When in a quiet environment , The recognition accuracy is about 98%, But when in a noisy environment , The accuracy will drop rapidly .

When at a party , It is difficult for speech recognition machines to pick up the speech of the target speaker from the overlapping speech , More difficult to identify accurately , This is a classic problem in the field of speech recognition —— Cocktail party problem （Cooktail Party Problem）. In the mix of various sounds , Hear the voice you want to pay attention to , It's human instinct . But for machines , This is it. “ explode ”, It must be through speech separation technology , First, separate the target speech , Then it can be identified .

Speech separation algorithm based on Neural Network

Speech separation is the solution in speech recognition “ cocktail lounge ” The first step of the problem . Add speech separation technology to the front end of speech recognition , Separating the voice of the target speaker from other interference can improve the robustness of the speech recognition system . Cocktail party problem refers to the collected audio signal except for the main speaker , There is also interference from other people's voices and noise . The goal of speech separation is to separate the main speaker's speech from these interferences .

At present, the mainstream speech separation algorithm is based on Neural Network , The main purpose of neural network is to learn an ideal binary masking （Ideal Binary Mask,IBM）, To determine which time-frequency units of the target signal in the spectrum （Time-frequency Units） Take the lead . If an auditory signal is divided into two dimensions: time domain and frequency domain （ Time-frequency two-dimensional ） To said , We can put the hour 、 The two dimensions of frequency are expressed as a two-dimensional matrix , Each element in this matrix is called a time-frequency unit . If you don't need to divide the target signal so carefully , Just once in a while —— It belongs to the target sound source , Or background noise , Then the time-frequency unit can be quantized as 2 It's worth , such as 0 and 1, This is binary . such , From the perspective of ideal binary masking , This problem becomes a supervised learning （Supervised Learning） The problem of classification .

Speech separation algorithm based on multimodal fusion

In addition to the above pure speech, do speech separation , Solve the cocktail party problem , Recently, there are many articles to solve the cocktail party problem with multimodal methods . Google from YouTube I searched for 10 10000 high-quality lectures and speech videos to generate training samples , Adoption of the covenant 2000 Hours of video clip analysis , Train a neural network based on multi stream convolution （CNN） Model of , Segment the synthetic cocktail party segment into separate audio streams for each speaker in the video . In the experiments , The input is one or more vocal objects , A video that is simultaneously disturbed by other objects or a noisy background . The output is to decompose the audio track of the input video into pure audio tracks , And correspond to the corresponding speaker .

Whether multimodal or monomodal Speech Separation Algorithm , Can not be separated from the support of voice data , The cost of voice data acquisition for multiple speakers is high 、 Difficulty in marking . and Magic Data As the world's leading AI Data service provider , It can provide many high-quality data for Algorithm Engineers , Provide experimental machine tools for solving cocktail party problems .

Edward Colin Cherry Published in 1957 Year of On Human Communication The book says ：“ up to now , No machine algorithm can solve ‘ cocktail lounge ’ problem .” I didn't expect that so far , This assertion is still not completely overturned .