Speech Recognition of Accented Mandarin Based on Improved Conformer.

Area of Science:

• Speech Recognition
• Artificial Intelligence
• Machine Learning

Background:

• Conformer models use convolution for translationally invariant features in speech recognition.
• Convolutional networks excel at local features, but dialect recognition needs long-range contextual information.
• Existing models struggle with the diversity of speech signals in Mandarin accent recognition.

Purpose of the Study:

• To propose the SE-Conformer-TCN model for improved Mandarin accent recognition.
• To enhance the Conformer architecture by integrating squeeze-excitation and temporal convolutional networks.
• To better model long-range contextual dependencies in speech spectrograms.

Main Methods:

• Embedded a squeeze-excitation (SE) block into the Conformer to model channel interdependencies.
• Integrated multi-head self-attention with a temporal convolutional network (TCN) in parallel.
• Utilized dilated causal convolutions within the TCN to capture sequential location information.

Main Results:

• The SE-Conformer-TCN demonstrated superior performance on four public Mandarin accent recognition datasets.
• Achieved a 2.1% reduction in sentence error rate compared to the baseline Conformer.
• Attained a character error rate of 4.9%.

Conclusions:

• The SE-Conformer-TCN effectively models both local and long-range features for Mandarin accent recognition.
• Integrating SE blocks and TCNs enhances the Conformer's ability to process diverse speech signals.
• The proposed model offers a significant improvement for accent-specific speech recognition tasks.