Language Identification in Short Utterances Using Long Short-Term Memory (LSTM) Recurrent Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Natural Language Processing

Background:

• Automatic Language Identification (LID) is crucial for various applications.
• Previous state-of-the-art LID systems, like i-vector and Deep Neural Networks (DNNs), face challenges with short utterances.
• Long Short Term Memory (LSTM) Recurrent Neural Networks (RNNs) show promise in overcoming these limitations.

Purpose of the Study:

• To present an open-source, end-to-end LSTM RNN system for LID.
• To evaluate the system's performance on short utterances and limited computational resources.
• To assess the system's capability in modeling unseen languages (out-of-set modeling).

Main Methods:

• Development of an open-source, end-to-end LSTM RNN system.
• Evaluation on a subset of the NIST Language Recognition Evaluation dataset (8 target languages, 3s task).
• Testing with varying utterance lengths, down to 0.1s, and incorporating out-of-set language modeling.

Main Results:

• The LSTM RNN system outperformed a reference i-vector system by up to 26% on the 3s task.
• The system demonstrated robust performance in detecting and generalizing to unseen languages.
• An accuracy exceeding 50% was achieved with utterances as short as 0.5s.

Conclusions:

• Open-source LSTM RNNs offer a reproducible and resource-efficient approach to high-performance LID.
• LSTM RNNs with out-of-set modeling are effective for real-world LID applications requiring robustness to unknown languages.
• The developed system shows significant potential for accurate language identification even with extremely limited audio data.