Adaptive context biasing in transformer-based ASR systems.

Area of Science:

• Artificial Intelligence
• Computer Science
• Computational Linguistics

Background:

• End-to-end neural automatic speech recognition (ASR) systems excel at identifying contextually biased words.
• Current ASR systems face challenges with bias layers, including increased computational complexity, resource demands, and redundant biases.

Purpose of the Study:

• To propose a novel Context Bias Adaptive Model for ASR systems.
• To dynamically assess and apply context biasing only when biased words are present in the input audio.
• To mitigate the negative effects of contextual bias while reducing computational overhead.

Main Methods:

• Development of a Context Bias Adaptive Model.
• Dynamic assessment of biased words in input audio.
• Conditional activation of the bias layer based on detected biased words.

Main Results:

• The Context Bias Adaptive Model effectively mitigates adverse effects of contextual bias.
• Substantial reduction in computational costs compared to traditional ASR systems.
• Bias layer activation is optimized, occurring only when necessary.

Conclusions:

• The proposed adaptive model offers an efficient approach to handling contextual bias in ASR.
• This method balances the need for contextual accuracy with computational resource management.
• The adaptive strategy presents a significant advancement in optimizing ASR system performance and efficiency.