Twine

Accent Classifier Project

LEE KIPNGETICH SANG
Published

0
  • Project creator

Accent Classifier – Multilingual Speech Pattern Recognition
Project Overview: The Accent Classifier is a machine learning system designed to identify and categorize speaker accents from audio input. It leverages deep learning models trained on multilingual speech datasets to support applications in voice personalization, accessibility, and forensic linguistics.

Key Components:

Audio Preprocessing:

Noise reduction, silence trimming, and MFCC (Mel-frequency cepstral coefficients) extraction

Spectrogram generation for CNN-based classification

Model Architecture:

Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) for temporal pattern recognition

Fine-tuned transformer-based models (e.g., Wav2Vec 2.0) for high-accuracy accent embeddings

Training Pipeline:

Labeled datasets across regional English accents (e.g., Kenyan, British, Indian, American)

Data augmentation via pitch shifting and time stretching to improve generalization

Evaluation using precision, recall, F1-score, and confusion matrix analysis

Deployment:

Real-time inference via serverless architecture (AWS Lambda + S3)

REST API integration for voice-enabled applications and chatbot personalization

Use Cases:

Accent-aware transcription and translation

Adaptive voice interfaces for global users

Forensic audio analysis in multilingual investigations #pitchdeck