kelvinyelyen / asr-ug

Automatic Speech Recognition (ASR) Model for Ghanaian Languages

Objective

The objective of this project is to develop an automatic speech recognition (ASR) model for Ghanaian languages using the SpeechData dataset. The task involves pre-processing the dataset, designing and training a deep learning model for speech recognition in Akan, Ewe, Dagbani, Dagaaare, or Ikposo languages, and evaluating its performance using appropriate metrics.

Task Description

Dataset Description

• The SpeechData dataset comprises audio recordings of spoken sentences in five Ghanaian languages and accents.
• The dataset consists of 1000 hours of audio for each language, accompanied by 100 hours of transcripts.
• The dataset is structured, with separate folders for audio files and their corresponding transcripts.

Pre-processing

• Divided the dataset into training, validation, and testing sets.
• Applied audio pre-processing techniques including normalization, feature extraction (MFCC), and data augmentation.
• Converted text transcripts into numerical representations suitable for training the ASR model.

Model Development

• Designed a deep learning architecture for speech recognition using TensorFlow.
• Experimented with various architectures including CNNs, RNNs, LSTMs, and Transformer models.
• Implemented attention mechanisms to enhance model performance.

Training

• Trained the ASR model using the training set.
• Fine-tuned hyperparameters such as learning rate, batch size, and regularization techniques.
• Monitored training progress and visualized training/validation loss and accuracy.

Evaluation

• Evaluated the trained model on the validation set using metrics like Word Error Rate (WER), Character Error Rate (CER), and Accuracy.
• Fine-tuned the model based on validation performance.

Testing

• Assessed the final trained model's generalization performance on the testing set.
• Calculated and reported final performance metrics on the testing set.

Documentation and Reporting

• Prepared a comprehensive report documenting the entire process, including dataset pre-processing, model development, training, evaluation, and testing.
• Described the chosen model architecture, hyperparameters, training methodology, evaluation metrics, and results analysis.
• Presented visualizations such as loss curves and confusion matrices to aid in understanding the model's behavior.
• Discussed insights gained from the project, challenges encountered, and potential avenues for future improvements.

For detailed instructions on how to replicate the project and insights gained, please refer to the documentation provided in this repository.