Digit-Speech-Recognition

For technical details, please see the Report

File Description and Usage

Files

1. bag_of_frames.py: Code for N-fold CV for bag of frames method. Requires MFCC features to be calculated and stored in ./Extracted_Feats/ in the specified format.
2. DTW.py: Code for N-fold CV for DTW method. Requires MFCC features to be calculated and stored in ./Extracted_Feats/ in the specified format.
3. live_endpointing.py: Code for real time endpointing used for interactive demo
4. Live_Test.ipynb: IPython notebook to test the code in an interactive setting. Requires VQ codebook to be generated from each speakers data using k means. This was tested for 8 clusters, however accuracy may improve if the number of clusters are increased.
5. mfcc_feat.py: Code to get the MFCC feature vector for the wav file input, obtained after endpointing in ./Processed_Data. Stores the MFCC in ./Extracted_Feats/ directory
6. Report.pdf: Report made for the project, describing things in more detail and also includes observations
7. Speech_Endpointing.ipynb: Used to endpoint the speech signals. Warning ! Different thresholds maybe required for different speakers. Splits the input waveform into smaller waveforms which are free of noise, in ./Processed_Data directory.
8. VQ.py: Used to generate the files in ./VQ_codebooks/ directory. Requires MFCC features to be calculated and stored in ./Extracted_Feats/ in the specified format.
9. VQ_check.py: Code for n CV validation of VQ method. Requires ./VQ_codebooks to be populated appropriately.

Directories

1. Extracted_Feats: Stores the extracted MFCC feature vectors for each 64 utterance of each digit. See the directory structure to get more idea on storing part
2. Output_Logs: Has 3 text files storing the output logs of N-CV of Bag of Frames, VQ and DTW methods
3. Processed_Data: Has 64 wav files corresponding to every digit, obtained after end-pointing appropriately
4. Raw_Data: Has the zip files for the given data of 16 speakers
5. VQ_codebook: Has the VQ codebooks generated for each speaker by VQ.py