English Speaker Accent Recognition

Using Transfer Learning with YAMNet

Approach

• Use YAMNet pre-trained model for feature extraction
• Train dense classifier on extracted features
• Classify UK & Ireland English accents

Dataset

17,877 audio files from 120 speakers across:

• Southern England
• Midlands
• Northern England
• Wales
• Scotland
• Ireland

Model Architecture

# YAMNet for features
yamnet_model = hub.load("https://tfhub.dev/google/yamnet/1")

# Dense classifier
model = Sequential([
    Dense(256, activation='relu'),
    Dropout(0.15),
    Dense(384, activation='relu'),  
    Dropout(0.2),
    Dense(7, activation='softmax')
])

Results

• Accuracy: ~51% validation
• AUC: ~0.89
• d-prime: 1.740

Transfer learning enables accent classification with limited data.