Accent Recognition with Sequential MFCC Features

This project addresses the challenge of accent diversity in English speech recognition systems, aiming to enhance their accuracy through the implementation of a supervised machine learning model. The model utilizes Sequential Mel-frequency cepstral coefficients (MFCC) features to classify speakers as either having an Indian or American English accent.

Methodology

1. Proposed Method: Sequential MFCC features are extracted from audio signals to provide a unique perspective for accent identification.
2. Dataset: The dataset comprises 3-5 second audio clips from VCTK-corpus, categorized into training (80%) and testing (20%).
3. Preprocessing: To address data imbalances, oversampling is applied to the training set. Feature extraction involves calculating 20 MFCC coefficients for each audio file using the Librosa library.
4. Feature Extraction: Mel-frequency cepstral coefficients (MFCC) are calculated for each audio frame, and sequential concatenation of these features provides a robust input set for accent classification.
5. Supervised Learning: Various classifiers, including K-Nearest Neighbors, Support Vector Machines, Gaussian Mixture Models, Neural Networks, and Logistic Regression, are employed for accent classification.

Results and Discussion

• Evaluation Metrics: Precision, recall, reject rate, and overall accuracy are used to evaluate classifier performance.
• Top Performers: Neural Networks, K-Nearest Neighbors, and Logistic Regression demonstrate superior performance, with Neural Networks emerging as the top classifier.

Conclusion and Future Work

• Promising Results: Accent classification through Sequential MFCC features showcases promising results, with up to 95% accuracy.
• Future Work: Incorporating this classification into speech recognition systems, extending the model to diverse accents and dialects, and exploring Hidden Markov Models for classification are areas for future development.

Note

• This project was developed during the IIIT Hyderabad Hackathon as part of the Qualcomm problem statement on Accent Detection.
• The evaluation criteria include accuracy scores, novelty in methodology or ideation, and the clarity of presentation.