Spoken language identification based on the enhanced self-adjusting extreme learning machine approach.

Area of Science:

• Speech processing
• Machine learning
• Artificial intelligence

Background:

• Spoken Language Identification (LID) relies on feature extraction, with established methods like MFCC, SDC, GMM, and i-vectors.
• Current LID models require improved learning optimization to fully utilize extracted features.
• Extreme Learning Machines (ELM) are effective for classification but suffer from suboptimal learning due to random weight initialization.

Purpose of the Study:

• To enhance the Extreme Learning Machine (ELM) for improved Spoken Language Identification (LID).
• To optimize the learning process of ELM by refining the weight selection phase.
• To introduce the Enhanced Self-Adjusting Extreme Learning Machine (ESA-ELM) for superior LID performance.

Main Methods:

• Utilized standard feature extraction techniques for LID.
• Employed the Self-Adjusting Extreme Learning Machine (SA-ELM) as a benchmark.
• Developed the Enhanced Self-Adjusting Extreme Learning Machine (ESA-ELM) by integrating Split-Ratio and K-Tournament methods into the SA-ELM optimization process.
• Evaluated LID performance on datasets comprising eight languages.

Main Results:

• The ESA-ELM demonstrated superior performance in Spoken Language Identification (LID) compared to the SA-ELM.
• ESA-ELM achieved an accuracy of 96.25% for LID.
• SA-ELM achieved an accuracy of 95.00% for LID.

Conclusions:

• The proposed ESA-ELM significantly improves LID accuracy over the SA-ELM.
• The enhanced optimization strategy in ESA-ELM effectively captures embedded knowledge from extracted features.
• ESA-ELM represents a more effective learning model for Spoken Language Identification.