timdietereberhardt / bachelor_thesis_speech-recognition-system-for-technical-domains

Bachelor Thesis - Conception of a Recurrent Neural Network based Speech Recognition System for Technical Domains with Comparison to Existing Systems

Abstract

The aim of the present bachelor thesis was to provide a proof of concept for the own development of speech recognition systems in the industrial sector. For this purpose nine keywords were defined which a recurrent neural network should recognize in an environment full of noises. The neural network was trained, validated and tested with collected data. To achieve best possible results of the neural network methods like: Data Augmentation, Dropout, L1/ L2 Regularization, Transfer Learning and Hyperparameter Optimization methods were used and their influence of the predictions was analyzed. By these methods a best network architecture could be chosen. In addition to that, a comparison between systems from providers such as Google and IBM and the developed network was done. For this purpose, 18 recordings without noise and 18 recordings with noise were used for this comparison. The result is that the own network performs better than the other systems in both variants of the recordings. To round off this comparison, these recordings were played to a test person to compare the results with the results of the technical systems.

Unfortunately, the data cannot be published according to German data protection law.

data

comparison_data: The test data used for comparing the systems and testing the best model converted_Warden_data: The Train, Val and Test MFCC Warden data V. 2 in npy format labeled_own_data: The labeled own data in WAV format, divided into: Train, Val & Test orginal_data: The original collection of data without labeling Testdaten_Python_Skripts: For the test of the data augmentation etc. these data were used

models

final_model: The best model in h5 format test_Warden_model: The model trained on the Warden for the Transfer Learning Method test.

notebooks

1. development NN, LSTM and GRU with the own data
2. hyperparameter optimization variant 1 & 2
3. development GRU with Warden data
4. test of the Transfer Learning Method
5. test of the best model

python_scripts

Prior to development, the methods were tested on smaller datasets. The py data includes these tests.

system_comparison_results

Excel file with the results of the tests of the other systems and the test person.

test_results

Output of the test of the best model in csv format

thesis

Digital pdf form of the thesis

Required Python libraries:

Keras, TensorFlow, Librosa, Scikit Learn, Numpy, MatPlotLib, os, pylon, json, pandas, wave, context lib