Products · Transcription
Shruti — Long-form transcription and translation, priced on what was actually said.
Shruti transcribes and translates long recordings — interviews, lectures, hearings, archives — including multilingual material that switches language mid-sentence. It runs on AI servers we own and operate, and it bills on detected speech rather than file duration, so silence costs nothing.
Shruti
Transcription · Early access
What it does
You hand Shruti hours of audio or video and get back a transcript, and if you want it, a translation. It is built for the material that defeats a naive transcription API: several hours long, more than one language, speakers who switch language mid-sentence, and recordings made in rooms rather than studios.
Why it is built this way
Transcription is the last pass, not the first. Shruti first maps where speech actually is, identifies the language across the recording, and smooths that into blocks — each block then gets the decode strategy that suits it, and blocks that are genuinely mixed are decoded without forcing a single language. A post-pass removes the failure modes long-form transcription is prone to: repetition loops, hallucinated text over silence, and duplicated seams where segments overlap.
What you pay for
Billing counts detected speech, not the length of the file. A three-hour recording with forty minutes of talking is charged as forty minutes. Processing runs on hardware we operate ourselves, so the media is never handed to a third-party API.
- Multilingual
- Self-hosted models
- Billed on speech, not duration
Where it lives