Developed offline speech recognition system that works with an accuracy of 97%

As a rule, the various speech recognition systems, persons, interpreters and others use large server capacity for their work. And in order to make them accessible to everyone, the developers transmit data over the Internet, which makes it impossible to use them offline. However, current algorithms of neural networks to help achieve truly amazing results. Not long ago, Microsoft and Google have already made their translators on the basis of neural networks is completely independent of the network, and now it is time to voice recognition algorithms.

Developed offline speech recognition system that works with an accuracy of 97%

is responsible for the development team of researchers from the University of Waterloo and a startup called DarwinAI. Their technology is called EdgeSpeechNets.

"In this study we use a strategy of creating architecture with a low level of load on the device, but with all the pluses approach with powerful neural network with deep machine learning."

For a start, experts have created a prototype of the future system that performs speech recognition, but had a limited vocabulary. At the same time he was able to identify him known keywords, even from the very rapid flow of speech. Thereafter, the resulting data were used for converting the audio signal into a mathematical formula. This formula has been used in the future for the design of neural networks, which would have high performance, but would not be demanding on the hardware. They then decided to try to get the program. For this purpose we used Google Speech Commands store that contains 65000 1-second audio samples. As a result, one of the versions of the system, namely EdgeSpeechNet-D, showed a great result, reaching 97% accuracy on a fairly weak smartphone Motorola Moto E c processor 1, 4 GHz.

"EdgeSpeechNet has a higher recognition accuracy at a much lower cost of computing. These results demonstrate that EdgeSpeechNet able to achieve the most advanced performance, requiring considerably less computing power, which makes them very suitable for use in mobile devices and applications. "

This and other news you can discuss in our in the chat Telegram.