Autoscribe: An automated tool for creating transcribed TextGrids from audio-recorded conversations

Area of Science:

• Linguistics
• Computational Linguistics
• Speech Processing

Background:

• Conversational research requires time-consuming segmentation and transcription of audio recordings.
• Existing automatic speech recognition tools often perform poorly on conversational speech compared to monologues.

Purpose of the Study:

• To develop and validate a user-friendly, automated tool for transcribing dyadic conversations.
• To create a tool that outputs Praat TextGrids with accurate turn boundaries and transcripts.

Main Methods:

• Developed Autoscribe, a tool converting dyadic conversational audio into Praat TextGrids.
• Validated Autoscribe on two distinct conversational corpora.
• Evaluated transcription accuracy and utterance boundary placement.

Main Results:

• Autoscribe reduced active working time for TextGrid creation by over 70%.
• Achieved an average transcription accuracy of 92%.
• Demonstrated an average utterance boundary placement accuracy of 95%.

Conclusions:

• Autoscribe significantly decreases the time and resources needed for conversational segmentation and transcription.
• The tool provides a practical solution for researchers in conversational analysis.
• Autoscribe enhances the efficiency of processing conversational audio data.