Can Authorship Attribution Models Distinguish Speakers in Speech Transcripts?

Explore the challenges and possibilities of speaker attribution in transcribed speech through this 10-minute conference talk presented by Cristina Aggazzotti at TACL 2024. Delve into the research conducted by Aggazzotti, Nicholas Andrews, and Elizabeth Allyn Smith from the Center for Language & Speech Processing (CLSP) at Johns Hopkins University. Examine how authorship verification techniques, typically applied to written text, can be adapted for transcribed speech. Discover the unique challenges posed by speech transcripts, such as the absence of traditional stylistic markers and the presence of speech-specific patterns like filler words and backchannels. Learn about the new benchmark developed for speaker attribution in human-transcribed conversational speech, designed to control for topic-related biases. Analyze the performance of various neural and non-neural baseline models on this benchmark, and understand how the effectiveness of written text attribution models varies as conversational topics are increasingly controlled. Gain insights into the impact of transcription style on attribution performance and the potential benefits of fine-tuning models on speech transcripts. Access the full research paper for a comprehensive understanding of this innovative approach to speaker attribution in transcribed speech.

Can authorship attribution models distinguish speakers in speech transcripts?