Long short-term memory for speaker generalization in supervised speech separation.

Area of Science:

• Signal Processing
• Machine Learning
• Acoustics

Background:

• Speech separation is crucial for enhancing audio quality in noisy environments.
• Supervised speech separation models struggle with generalization to new noises and speakers.
• Deep Neural Networks (DNNs) show promise but have limitations in modeling diverse speaker characteristics.

Purpose of the Study:

• To develop a speech separation model that improves generalization to unseen speakers and noises.
• To leverage the temporal modeling capabilities of Long Short-Term Memory (LSTM) networks for enhanced speaker generalization.
• To evaluate the proposed LSTM-based model against DNN-based approaches for speech intelligibility and efficiency.

Main Methods:

• Formulating speech separation as estimating a time-frequency mask from acoustic features.
• Developing a novel speech separation model utilizing Long Short-Term Memory (LSTM) architecture.
• Conducting systematic evaluations comparing the LSTM model with a DNN-based model on objective speech intelligibility metrics.

Main Results:

• The proposed LSTM model significantly outperforms the DNN-based model on unseen speakers and noises.
• LSTM's internal representations demonstrate effective capture of long-term speech contexts.
• The LSTM model shows advantages for low-latency speech separation, even without future frame information.

Conclusions:

• The LSTM-based model offers an effective solution for speaker- and noise-independent speech separation.
• LSTM networks provide superior generalization capabilities compared to traditional DNNs for this task.
• The proposed approach is promising for real-time and robust speech enhancement applications.