Specialized audio models for real production work.
General-purpose speech AI is trained on clean, single-microphone audio. Real productions have several mics, overlapping voices, room noise and long recordings. We build small models trained for those conditions, sized to run on an editor’s own computer.
Who said what, on-device.
of words attributed to the correct speaker in two-person conversations
words across the evaluation set of full-length recorded conversations
bytes uploaded. The model runs entirely on the user’s machine
Early internal benchmark of our first in-house speaker model, November 2026. We’ll publish updated results as the evaluation set grows.
What we’re working on
Speaker diarization
Identifying who is speaking, word by word, in long multi-person recordings. It powers speaker labels in EasyScribe and camera decisions in PodDraft.
Conversation audio cleanup
Reducing noise, crosstalk and microphone bleed in multi-mic podcast and interview recordings while keeping voices natural.
Word-accurate timing
Placing every word to the frame, so edits made in a transcript land exactly right in the video timeline.
Editorial selection
Choosing the strongest material from hours of footage against a target length, and explaining each choice in a way an editor can verify.
Post-production is still mostly manual.
Editors spend hours on mechanical work before any creative decision: syncing, labeling speakers, cleaning audio, logging footage. Cloud tools help, but many productions can’t upload client media, and general models stumble on real-world audio.
Small specialized models that run locally solve both problems. Training them needs large amounts of real conversation audio and significant GPU time, and that is where our research investment goes.
Get in touch about research partnerships or evaluation data.