Training candidate selection for effective out-of-set rejection in robust open-set language identification

Area of Science:

• Speech processing
• Machine learning
• Computational linguistics

Background:

• Open-set language identification (LID) systems often struggle with effective out-of-set (OOS) language rejection.
• Accurate OOS language rejection is crucial for robust speech and language pre-processing pipelines.

Purpose of the Study:

• To develop and evaluate effective methods for selecting OOS language candidates for improved OOS rejection in LID systems.
• To enhance the performance of LID systems in open-set scenarios by optimizing the training data for OOS detection.

Main Methods:

• Three OOS candidate selection methods were developed: unsupervised k-means clustering, complementary candidate selection, and score-level language relationship-based selection.
• These methods utilized probe OOS data to ensure universal OOS language coverage for an i-vector LID system with a Gaussian backend.
• Evaluation was performed on the large-scale LRE-09 corpus, comprising 40 languages.

Main Results:

• The proposed selection methods reduced OOS training data diversity by 86%.
• Performance comparable to closed-set LID was achieved, even when using significantly less OOS training data.
• The methods demonstrated superior performance compared to random candidate selection, achieving sustained accuracy with fewer OOS candidates.

Conclusions:

• Effective OOS language selection is a viable strategy to enhance OOS rejection in open-set LID systems.
• The developed methods offer a significant improvement in data efficiency and performance for LID systems dealing with unknown languages.
• This research represents a key advancement in addressing the challenge of OOS language selection for improved LID performance.