Automatic speech recognition in cocktail-party situations: a specific training for separated speech.

Area of Science:

• Acoustic Signal Processing
• Speech Recognition Technology
• Artificial Intelligence

Background:

• Automatic speech recognition (ASR) systems struggle with accuracy, particularly in adverse acoustic conditions like cocktail-party environments with simultaneous speech.
• Existing source separation methods for simultaneous speech produce artifacts that further degrade ASR performance.
• Human speech recognition capabilities far exceed current ASR system performance in complex acoustic scenarios.

Purpose of the Study:

• To propose a specific training methodology to enhance ASR performance in real-world simultaneous speech scenarios.
• To investigate the combined effectiveness of source separation techniques and the proposed specialized training.
• To evaluate the improvements in ASR accuracy under various acoustic conditions.

Main Methods:

• Development of a specialized training regimen tailored for ASR systems handling simultaneous speech.
• Integration of source separation algorithms to pre-process acoustic signals containing overlapping speech.
• Experimental evaluation of the combined approach across diverse adverse acoustic conditions.

Main Results:

• The proposed specific training significantly improves the percentage of recognized words in simultaneous speech.
• Combining source separation with the specialized training leads to substantial gains in ASR performance.
• Improvements of up to 35% in ASR accuracy were observed under tested acoustical conditions.

Conclusions:

• The developed training strategy effectively addresses the challenges of ASR in simultaneous speech environments.
• The synergistic application of source separation and targeted training offers a viable solution for improving ASR robustness.
• This research presents a significant advancement in achieving more accurate speech recognition in complex acoustic settings.