Speech AI for Research Recordings
Transcription and speaker-separation benchmark (in progress)
2026 · Python, open-source speech models, Claude
Overview: Study sessions produce hours of recordings. I'm building a pipeline that turns them into transcripts researchers can trust, with the right words attributed to the right speaker. That way quotes, behaviors and findings can be pulled out accurately and at scale.
My role: I designed and built a fully local pipeline that transcribes, diarizes (separates speakers) and aligns multi-speaker session audio into speaker-labeled transcripts and clips. I also built the benchmark for choosing the speech models behind it.
Approach:
• Scored open-source transcription and diarization models against hand-coded ground truth on public meeting audio.
• Used three measures: diarization error rate, a facilitator/participant role-swap metric, and signal-detection (d′) scoring. The d′ score separates missed speech from speech credited to the wrong speaker.
• Compared accuracy and speed across different GPUs.
• Worked with Claude as a coding and analysis partner, from writing and debugging pipeline code to running benchmarks and synthesizing results.
Outcome: The benchmark on public audio is done. The best model depends on how many people are talking (see the chart below). Next, I'll run the pipeline on study recordings and automate scoring new models against earlier ones. A full write-up is coming.