Speech Recognition on Spoken Digit Dataset
This github repository contains the code that I have used to perform Speech Recognition on Spoken Digit Dataset. In other words this github repository contains the code for isolated word recognition on the Spoken Digit Dataset consisting of trimmed audio signals of spoken digits.
Dataset Description
The dataset used here is Free Spoken Digit Dataset (FSDD), which is an open source dataset consisting of audio recordings of the spoken digits 0-9 in English. The dataset is essentially a spoken version of the popular MNIST dataset which contains images of handwritten digits.
Hand-written and spoken digit 7 from the MNIST and FSDD datasets, respectively.
Problem Statement
The dataset contains audio recordings of the spoken digits 0-9 in English. Each audio recording is consisting of the utterance of a single word. The task is to determine the word that was uttered in each of the recordings, given the audio signal of the recording.
Approch
There are many different ways to represent audio, including raw digital signals, spectrograms, chromagrams, scalograms and Mel-spectrograms. However, one of the most effective and frequently used representations in the domain of automatic speech recognition is Mel-Frequency Cepstral Coefficients (MFCCs). MFCCs are a collection of features that can be used to represent a section (or frame) of audio. The number of MFCCs to use is an adjustable hyper-parameter, but this is generally fixed to specific values such as 13 or 40. For these experiments, I choose to use 39 MFCCs (40, but with the constant offset coefficient remove as it does contain relevant information).
Standardization is performed on MFCCs of a audio signal. After Standardization each frame of MFCCs is passed one by one sequentially on every iteration of a Bidirectional LSTM RNN.
Bidirectional Recurrent Neural Network with Long Short-Term Memory (BiLSTM-RNN)
RNN - The Recurrent Neural Network (RNN) is a type of neural network that is designed to handle sequential data of various length. RNNs work by feeding the input sequence into an RNN cell one time step at a time.
RNN cell unfolded to accept an input sequence with 
time steps.
For sequence classification, we typically only take the last hidden state vector 
and use this for classification, effectively using it to represent the context of the entire sequence.
LSTM RNN - However, the vanilla RNN suffers from issues with vanishing gradients, effectively resulting in earlier parts of the input sequence being "forgotten". The Recurrent Neural Network with Long Short-Term Memory (LSTM-RNN) is a special way of defining a RNN cell by using gates to allow the network to decide what information to remember, and for how long to remember it.
Bidirectional LSTM RNN - We can further improve the performance of LSTM RNN by introducing the concept of bidirectionality. A Bidirectional Recurrent Neural Network with Long Short-Term Memory (BiLSTM-RNN) uses two cells: one to process the input sequence in the normal forward direction, and one to process the sequence backwards.
The last hidden state from Bidirectional LSTM RNN is then passed to a Dropout Layer. Then the result from the Dropout layer is passed to a Linear Feed Forward Neural Netowork followed by ReLU Activation and a Dropout Layer. The resultant is then passed again through a Linear Feed Forward Neural Network and then LogSoftmax function is applied to get classification result.
Model Diagram
Training Results
| Dataset | Accuracy (%) |
|---|---|
| Training | 96.50 |
| Validation | 98.00 |
| Testing | 97.33 |
Training History
Confusion Matrix
Following are the hyperparameters (The below hyperparamter settings were found to be optimal):
- Batch size (unchanged) : 64
- Number of epochs (unchanged) : 15
- Learning rate (unchanged) : 0.002 with Adam Optimizer
- Number of MFCCs : 39
- Number of LSTM layer : 1
- Number of Hidden state dimensions : 50
- Number of units in Linear Feed forward Neural Network : 50
The dataset was splited as 80% train dataset, 10% validation dataset and 10% test_dataset during this experiment.
Requirements
pip version used is 22.0.2 and python3 version used is 3.10.6
The following packages are used in this project:
matplotlib==3.5.3
numpy==1.23.3
pandas==1.4.3
PyYAML==6.0
scikit_learn==1.1.2
seaborn==0.12.1
torch==1.12.1
torchaudio==0.12.1
torchvision==0.13.1
tqdm==4.64.0