Overlapped Speech detection in multi-party conversations

---

Authors: Neeraj Sajjan, Shobhana Ganesh, Neeraj Sharma, Sriram Ganapathy, Neville Ryant

Reference: N. Sajjan, S. Ganesh, N. Sharma, S. Ganapathy, N.Ryant, "Leveraging LSTM models for overlap detection in multi-party meetings", IEEE International Conference on Acoustics, Speech and Signal Processing, April 2018.

---

This is a project on overlapped speech detection in multi-party conversation meetings. We explore the effectivness of various features such as Mel Spectrogram, kurtosis etc using a neural network approach on two datasets: TIMIT and AMI. We make use of force alignment to rectify the errors inherent in human annotations of the AMI dataset.

Dependencies

1. Python 2.7
2. HTK Toolkit
3. Keras

Folder Structure and Description

Codes

1. Feature Handling Codes:

• htkmfc.py : python interface to reading and writing htk files
• htk_dataprep.sh : shell script to generate htk feature files
• mfcc_config.cfg : config file for mel spectorgram[fbank] feature generation
• get_gammatone_feats.py : code for extraction of gammatone spectorgram features
• gammatonegram_package.py : necessary functions for get_gammatone_feats.py
• generator.py : code to generate artifical overlapped speech wav file from two single speaker files[for TIMIT]
• wrapper_for_gen.py : wrapper for generator.py, takes in a list of single speaker wav files to generate overlapped wav files
• PAR_make_context_feats.py : Code to generate context features for multiple files parallely
• PAR_do_cmvn_feats.py : Code for cepstral mean variance normalisation
• concatenator.py : Final train/val/test data generation code using htk file format
• kurtosis_extractor.py : Code to extract kurtosis
• sfm.py : Code to extract spectral flatness measure features

2. Model_Train_Test_Codes:

• rnn.py : Code for final lstm model in Keras
• cnn.py : CNN model
• dnn.py : Three layered dnn
• clstm.py : CNN followed by LSTM
• test_rnn.py : Code for testing final lstm model
• confusion_matrix_gen.py : Code to generate confusion matrix for three classes[Single/Overlap/Filler]

Labels

1. Original_Labels : Labels of AMI train, dev[val], eval[test] sets before force alignment.
2. Force_Aligned_Labels : After Force alignment