Read Project_Report.pdf for more details.
To recognize a single spoken number from zero to nine. To be specific, as one speaks a number (0 - 9), the program will recognize the correct number.
Thanks to the dataset from Pannous, http://pannous.net/files/spoken_numbers_pcm.tar.
The dataset includes 2850 .wav files of 15 different people (male and female) speaking number 0 - 9. Besides, 400 .wav files recorded by me and my roommate are added to the dataset.
tensorflow-gpu, keras, librosa, numpy, matplotlib, pyaudio, h5py
Draw the spectrogram of each .wav file, and save as an image. In this way, the speech recognition problem is transfered into an image recognition problem.
Use CNN to build a classifier for the dataset. The CNN model includes 2 Dense (fully connected) layers and 5 Convolution layers, with Max-Pooling and BatchNormalization layers in it.
At the training stage, the acc reaches 100% after 20 epochs, and val_acc reaches 98%. Very nice!
At the real-time test stage, the error rate keeps very low. Nice!
[1] https://github.com/libphy/which_animal
[2] https://github.com/pannous/tensorflow-speech-recognition
[3] https://yerevann.github.io/2016/06/26/combining-cnn-and-rnn-for-spoken-language-identification/