Filled pause refinement based on the pronunciation probability for lecture speech.

Area of Science:

• Computational Linguistics
• Speech Processing
• Artificial Intelligence

Background:

• Automatic speech recognition (ASR) struggles with disfluent speech, particularly lecture and conversational content.
• Filled pauses (FPs) are common disfluencies that degrade ASR accuracy due to poor annotation and acoustic similarity to words.
• Existing ASR systems lack robust methods for handling FPs in training data.

Purpose of the Study:

• To propose and evaluate a novel automatic refinement approach for detecting filled pauses (FPs) in British English lecture speech.
• To enhance the performance of ASR systems by addressing the challenge of disfluencies in spontaneous speech.
• To improve the accuracy of speech transcription for lecture-based content.

Main Methods:

• Developed a modified forced-alignment framework integrating pronunciation probabilities and acoustic language model scores.
• Implemented an automatic refinement approach specifically designed for filled pause detection.
• Evaluated the method on the Reith Lectures speech transcription task using imperfect training data.

Main Results:

• Achieved successful results in filled pause detection for both development and evaluation datasets.
• Demonstrated the effectiveness of the proposed FP refinement approach on ASR performance.
• Investigated the impact of acoustic models trained on different speech genres on FP refinement.

Conclusions:

• The proposed automatic refinement approach effectively detects filled pauses in British English lecture speech.
• Enhancing ASR systems with FP detection improves transcription accuracy for spontaneous speech.
• The method shows promise for improving ASR performance even with imperfect training transcriptions.