Keyword spotting for trigger words in speech recognition systems.
Used the Google Speech Commands Dataset. The task is to classify the words into 12 classes: yes, no, up, down, left, right, on, off, stop, go, silence, unknown
Accuracy achieved: 85.6% (Placed in top 30% - Kaggle Rank: 383/1315)
The kaggle competition winning model achieved an accuracy of 91%.
- Collecting files and organizing data, making noise samples.ipynb
- Creating noised data.ipynb
- Data augmentation - Time shifting noised and unnoised data.ipynb
- Augmenting classes 0 to 9.ipynb
- MFCC EXTRACTION.ipynb
- Neural network.ipynb
- Kaggle testing set performance.ipynb
-
MFCC parameters: Number of MFCC features: 128, Hop length: 320, FFT window: 2048
-
Convolutional neural network:
Note: The data was augmented by random time shifting between 1 and 200 ms, and adding noise as described in the data augmentation jupyter notebooks.