A comparison of automatic and human speech recognition in null grammar

Area of Science:

• Speech Recognition Technology
• Human-Computer Interaction
• Acoustic Modeling

Background:

• Current automatic speech recognition (ASR) evaluations integrate acoustic and language modeling, hindering isolated performance assessment.
• Existing ASR accuracy metrics do not isolate the impact of acoustic modeling due to inherent language model integration.
• Understanding ASR limitations requires separating acoustic performance from language context.

Purpose of the Study:

• To compare the accuracy of automatic speech recognition (ASR) and human speech recognition (HSR) under controlled conditions.
• To investigate the performance of ASR and HSR with null grammar sentences across varying signal-to-noise ratios (SNRs).
• To analyze the influence of vocabulary size on ASR and HSR accuracy in degraded acoustic environments.

Main Methods:

• Evaluated ASR and human speech recognition (HSR) accuracy using sentences with null grammar.
• Tested performance across diverse signal-to-noise ratios (SNRs) to simulate real-world acoustic challenges.
• Varied vocabulary sizes (1000, 2000, 4000, 8000) to assess system scalability and robustness.

Main Results:

• Significant differences were observed between ASR and HSR performance, particularly in noisy conditions.
• The study highlights the varying importance of bottom-up word recognition versus top-down context awareness for ASR and HSR.
• ASR systems showed distinct performance degradation patterns compared to HSR as noise levels increased and vocabulary expanded.

Conclusions:

• The research provides insights into the fundamental differences between machine and human speech processing capabilities.
• Findings suggest that ASR systems may rely more heavily on acoustic cues, while HSR leverages contextual information more effectively.
• This study aids in developing more robust ASR systems by identifying weaknesses in acoustic modeling and context integration.