Characterizing correlations in partial credit speech recognition scoring with beta-binomial distributions.

Area of Science:

• Auditory neuroscience
• Speech processing
• Statistical modeling

Background:

• Partial credit scoring can improve measurement precision in speech recognition.
• Assessing this improvement is complex due to contextual cues creating correlations in token identification probabilities.

Purpose of the Study:

• To estimate speech recognition accuracy and intraclass correlation using beta-binomial distributions.
• To investigate these correlations in phonemes within words and words within sentences for cochlear implant listeners.

Main Methods:

• Utilized beta-binomial distributions to model recognition accuracy.
• Analyzed phoneme and word recognition in 20 cochlear implant users.
• Calculated intraclass correlation coefficients for recognition accuracy within stimuli.

Main Results:

• Demonstrated substantial intraclass correlation in recognition accuracy within stimuli.
• Found these correlations to be invariant across individual participants.
• Highlighted the impact of contextual cues on speech recognition data.

Conclusions:

• Intraclass correlations in speech recognition accuracy are significant and consistent across individuals.
• These correlations necessitate consideration in power analyses for partial credit scoring.
• Findings are crucial for accurate measurement in cochlear implant research.