Lossy Coding Impact on Speech Recognition with Convolutional Neural Networks

This paper presents research of lossy coding impact on speech recognition with convolutional neural networks. For this purpose, google speech commands dataset containing utterances of 30 words was encoded using four most common all-purpose codecs: mp3, aac, wma and ogg. A convolutional neural network was taught using part of the original files and later tested with the rest of the files, as well as their counterparts encoded with different codecs and bitrates. The same network model was also taught using mp3 encoded data showing the biggest loss in effectiveness of the previous network. Results show that lossy coding does have an effect on speech recognition, especially for low bitrates.