Abstract
Oral History Multi-Speaker Interpretation-Kit is a tiered workflow for creating oral history transcriptions that implements various Whisper models for speech-to-text recognition and SpeechBrain for diarization, or the process of sorting an audio recording into segments that indicate who is speaking when. The kit is organized to batch process collections of recordings, outputting a CSV file with timestamps and dialogue separated by speaker. Both Whisper and SpeechBrain are open source, do not require login or tokens to access and run locally once their pre-trained models are downloaded.
The Python scripts are designed to batch process the greatest number of recordings first, then apply more advanced scripts to more difficult recordings. Elements such as low audio fidelity, suboptimal recording environments, crosstalk, and vocal similarity between speakers may introduce errors into Whisper's pattern recognition, and result in dialogue clusters being under parsed or punctuation being dropped, which the kit's more advanced scripts can help mediate.
This kit was developed over time to facilitate the transcription of the Latah County Oral History Collection, an initiative conducted in the 1970's by the Latah County Historical Society and later digitized by the University of Idaho's Center for Digital Inquiry and Learning (CDIL) in 2015. The author developed this kit to transcribe the over 550 hour collection during the spring and summer of 2026 to make the material more discoverable for researchers and providing the Latah County community with easier access to its history. This kit was developed for implementation in the CDIL's Oral History as Data framework developed by Devin Becker, as well as the author's oral history transcript mining method outlined in Distant Listening: Using Python and Apps Scripts to Text Mine and Tag Oral History Collections.
Future iterations of ohmsi-kit may include an editing workspace where users can open processed transcriptions locally to aid and streamline the copyediting process, leveraging Oral History as Data’s playback interface and Visual Studio Code’s user dictionary capabilities. Other advancements may include automated sequential processing of audio files based on a programmatic survey that evaluates transcripts for accurate dialogue clustering.