Nonlinear Regularization Decoding Method for Speech Recognition

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Current end-to-end speech recognition relies on hybrid CTC and Transformer decoders.
• Error accumulation in these decoders limits accuracy improvements.
• Transformer models are often too complex for small datasets.

Purpose of the Study:

• To introduce a Nonlinear Regularization Decoding Method for Speech Recognition.
• To address limitations of existing Transformer-based speech recognition models.
• To enhance accuracy and efficiency, particularly for small datasets.

Main Methods:

• Implemented a nonlinear Transformer decoder allowing arbitrary character associations, overcoming left-to-right limitations.
• Introduced a regularization attention module to optimize attention scores and mitigate error propagation.
• Developed a tiny model to reduce parameter size and improve efficiency.

Main Results:

• The proposed model achieved significant Uyghur speech recognition improvements.
• Recognition accuracy increased by 0.12% on Aishell1, 0.54% on Primewords, 0.51% on Free ST Chinese Corpus, and 1.2% on Common Voice 16.1.
• The nonlinear approach demonstrated effectiveness on smaller datasets.

Conclusions:

• The Nonlinear Regularization Decoding Method offers a promising alternative to traditional speech recognition decoders.
• The method effectively reduces error accumulation and improves accuracy.
• The tiny model variant provides an efficient solution for resource-constrained environments.