How should a speech recognizer work?

Area of Science:

• Cognitive Science
• Computer Science
• Computational Linguistics

Background:

• Human speech recognition (HSR) and automatic speech recognition (ASR) share goals but lack interdisciplinary communication.
• Research in both fields has historically focused on algorithmic and implementational levels, neglecting the computational level of speech recognition.

Purpose of the Study:

• To provide a computational-level analysis of speech recognition, highlighting parallels between HSR and ASR.
• To introduce a novel computational model for human spoken-word recognition.
• To demonstrate the efficacy of ASR techniques in advancing HSR models.

Main Methods:

• Computational-level analysis of the speech recognition task.
• Development of a new computational model for human spoken-word recognition.
• Application of automatic speech recognition techniques to human speech recognition.

Main Results:

• Revealed significant parallels between HSR and ASR research at the computational level.
• The new model successfully recognizes words from real speech input.
• The model, built with ASR techniques, outperforms existing HSR models in recognizing real speech.

Conclusions:

• A computational-level approach fosters greater integration between HSR and ASR research.
• ASR techniques offer a powerful framework for developing more effective HSR models.
• The developed model represents a significant advancement in understanding and replicating human spoken-word recognition.