Spoken Language Identification Using Deep Learning.

Area of Science:

• Speech processing and machine learning
• Computational linguistics
• Digital signal processing

Background:

• Spoken Language Identification (SLID) is crucial for multilingual applications.
• Accurate language detection requires robust feature extraction from audio.
• Existing methods face challenges with speaker variability and diverse languages.

Purpose of the Study:

• To develop and evaluate a CNN-based model for accurate SLID.
• To identify key acoustic features for distinguishing between multiple languages.
• To assess the model's performance on a diverse dataset of spoken utterances.

Main Methods:

• Audio files were converted into spectrogram images.
• A Convolutional Neural Network (CNN) was employed for feature extraction.
• The model was trained and tested on the Kaggle spoken language identification dataset.
• Utterances were fixed at a 10-second duration and split into training and testing sets.

Main Results:

• Preliminary results indicated an overall accuracy of 98%.
• Extensive testing demonstrated a final overall accuracy of 88%.
• The CNN model effectively extracted discriminative features for language identification.

Conclusions:

• CNNs are effective for spoken language identification, achieving high accuracy.
• The model shows promise for real-world multilingual audio processing applications.
• Further research can explore optimizations for improved performance and broader language coverage.