Multilingual Transcription Engine
A first draft for every track. An LLM corrects it. Editors verify what's uncertain.
Five tracks — or upload your own.
Raw ASR draft
unpunctuated · uneditedAfter LLM correction
awaiting runA signal chain, not a black box.
1 · Vocal isolation
Vocals isolated from the mix before transcription.
Runs server-side2 · ASR draft
A first-pass transcript, tuned for singing, not speech.
Runs server-side3 · LLM correction
Homophones, word breaks, and mis-hears resolved in context.
● Live in this demo4 · Confidence routing
Uncertain lines go to an editor. Certain lines move on.
● Live in this demo5 · Feedback loop
Every correction retrains the next draft.
Runs server-sideFrom catalog gap to published lyric.
1 · Detect
New ingestion checked against the licensed-lyrics feed. Gaps outside Tier 1 auto-enqueue.
2 · Prioritize
Queued by release recency, streaming signal, and market coverage targets — not ingestion order.
3 · Route by confidence
High confidence goes to spot-check. Medium goes to standard review. Low goes to a specialist or holds.
4 · Editor acts
Approve confirms the draft. Edit corrects it. Reject escalates it — nothing publishes on a guess.
5 · Publish + feedback
Approved and edited lines sync to TTML and ship. Edits feed the next fine-tune.
Human transcription doesn't scale past the top of the catalog.
The bottleneck was never talent. It was time.