Deep bottleneck features for spoken language identification.

Area of Science:

• Speech processing
• Machine learning
• Computational linguistics

Background:

• Effective language-specific representations are crucial for spoken language identification (LID).
• Current LID methods, while improved by machine learning, struggle with short utterances and are sensitive to speaker, content, and noise variations.
• Existing representations may be insufficient due to latent and statistically dependent language information in speech.

Purpose of the Study:

• To propose Deep Bottleneck Features (DBF) as a novel representation for spoken language identification.
• To develop and evaluate acoustic models (DBF-TV and PDBF-TV) utilizing DBF-based i-vectors.
• To enhance LID performance, particularly for short-duration speech, by addressing limitations of existing methods.

Main Methods:

• Utilized Deep Neural Networks (DNNs) to extract Deep Bottleneck Features (DBFs).
• Developed DBF-based i-vector representations for speech utterances.
• Designed and implemented two acoustic models: DBF-TV and parallel DBF-TV (PDBF-TV).

Main Results:

• Achieved significant improvements in LID performance on the NIST language recognition evaluation 2009 (LRE09) dataset.
• Demonstrated superior performance over state-of-the-art systems, especially for short utterances (3s, 10s, 30s).
• Obtained Equal Error Rates (EER) of 7.01% (3s), 1.89% (10s), and 1.08% (30s) by fusing phonotactic and acoustic approaches.

Conclusions:

• Deep Bottleneck Features (DBF) provide a powerful, low-dimensional, and discriminative representation for spoken language identification.
• The proposed DBF-based models (DBF-TV, PDBF-TV) significantly advance the state-of-the-art in LID.
• DBFs effectively capture latent language information, offering robustness against variations and improving performance on challenging short utterances.