Automatic recognition of second language speech-in-noise.

Area of Science:

• Speech Communication
• Acoustic Phonetics
• Computational Linguistics

Background:

• Speech intelligibility measurement is crucial for speech communication research and applications.
• Current methods like human transcription are labor-intensive and costly.
• Advancements in automatic speech recognition (ASR) offer potential for automated intelligibility assessment.

Purpose of the Study:

• To evaluate the performance of state-of-the-art ASR systems in measuring speech intelligibility.
• To compare ASR system accuracy against human listener performance.
• To identify the opportunities and limitations of ASR for speech intelligibility modeling.

Main Methods:

• Tested four advanced ASR systems using second language speech presented with background noise.
• Analyzed system performance against human transcription accuracy.
• Examined response divergence between ASR and human listeners, particularly at low signal-to-noise ratios.

Main Results:

• One ASR system, whisper, achieved performance comparable to or exceeding human listener accuracy.
• Significant differences were observed between whisper's responses and human responses, especially in noisy conditions.
• The findings indicate ASR systems can approach human accuracy but may not fully replicate human perception.

Conclusions:

• ASR systems show promise for automating speech intelligibility measurements.
• Divergence in response content highlights the need for further refinement in ASR-based intelligibility models.
• Future research should explore how ASR outputs can be better leveraged for theoretical, technological, and clinical speech communication.