alvayliu / language-identification-with-cnn Goto Github PK
View Code? Open in Web Editor NEWSpoken language identification systems (LID) allow for automatic language detection given speech data. Among the many available methods that can be applied to this classification task, modern machine learning and deep learning approaches have been reported as effective. A previous study approached the problem of spoken language identification in the image domain by transforming speech samples to spectrograms and classifying them using convolutional neural networks (CNN). We have implemented two similar types of CNNs and trained them on data for five languages from the SpeechDat database. Then, we investigated how well their performance generalised on speech samples from another source then SpeechDat. The results indicated that even though the models could achieve over 80 % in test accuracy on SpeechDat data, they did not perform well on speech samples not originating from the SpeechDat database, with the best model achieving 37.5 % accuracy.
License: MIT License