Obtaining phonetic transcriptions: a comparison between expert listeners and a continuous speech recognizer.

Area of Science:

• Phonetics and Phonology
• Speech Technology
• Computational Linguistics

Background:

• Accurate phonetic and phonological representations are crucial for speech analysis.
• Expert listeners are traditionally used for phonetic transcription, a time-consuming process.
• The utility of automated tools like continuous speech recognition (CSR) for phonetic analysis requires evaluation.

Purpose of the Study:

• To compare the performance of a CSR tool against expert listeners in identifying specific phonetic segments.
• To assess the feasibility of using CSR for large-scale phonetic and phonological analysis.
• To investigate the impact of phonological processes like schwa-deletion and schwa-insertion on CSR accuracy.

Main Methods:

• Two experiments were conducted comparing CSR performance with nine expert listeners.
• Participants judged the presence or absence of prespecified phones in 467 speech segments.
• The second experiment specifically focused on analyzing schwa-deletion and schwa-insertion processes.

Main Results:

• Significant performance differences were observed between the CSR tool and human listeners.
• Variability in performance was also noted among individual expert listeners.
• Despite statistical significance, the magnitude of differences may be acceptable depending on the application's requirements.

Conclusions:

• Continuous speech recognition (CSR) tools can serve as a viable alternative to human listeners for identifying the presence of phones.
• The ability of CSR to process large datasets may outweigh the errors it introduces.
• CSR offers a scalable solution for phonetic tasks, enabling exploration of extensive speech data.