The python notebook contains the code used for the manual transcription of SEP-28k Clips- inference runs of Whisper 2 and 3 on this audio- the data analysis- and the graphing of results for this work.
There is also an audio folder with all the audio clips from SEP-28k used in this research.
The CSV file contains the results of the project- and for each clip- it shows
- stutter_present: Binary label for whether the audio clip contains a stuttering event
- manual_prolongation: Binary label for whether the audio clip contains a Prolongation
- manual_block: Binary label for whether the audio clip contains a Block
- manual_soundRep: Binary label for whether the audio clip contains a Sound Repetition
- manual_wordRep: Binary label for whether the audio clip contains a Word Repetition
- manual_interject: Binary label for whether the audio clip contains an Interjection
- hallucination_binary: Binary label for whether the audio clip contains a Hallucination
- manual_transcription_semantic: Transcription of audio clip with stuttering events removed
- manual_transcription_literal: Transcription of audio clip with stuttering events included
- Whisper2Annotation: Transcription of audio clip created by Whisper 2
- Whisper3Annotation: Transcription of audio clip created by Whisper 3
- audio_clip_name: (show name)_(episode id)_(clip id)
J-j-j-just Stutter: Benchmarking Whisper’s Performance Disparities on Different Stuttering Patterns. Charan Sridhar, Shaomei Wu. Accepted to appear in Proceedings of the InterSpeech Conference. 2025. (Preprint)
