Product
POST /v1/stt
Speech to text
Transcribe support calls, meetings and voice notes where half the sentence is English and the other half isn’t. Word timings and speaker labels come back in the same response.
503 model_not_servable rather than a transcript. Everything below describes the contract you can build against today with a vx_test_ key — the request shape, the response shape and the error codes are settled and will not move when the model lands. The sample transcript on this page is an illustration, not a live capture.Models
| Version | Aliases | Rate / min |
|---|---|---|
orphix-pro-1.0.0 Registered, not yet served. Batch transcription. | scribe-1 | — |
orphix-flash-1.0.0 Registered, not yet served. No alias points at it. | — | — |
Billed per second of audio, rounded up per request. A 12-second call costs 12 seconds.
Output
What comes back
Diarisation
Who said what
Stable spk_1 / spk_2 labels across the whole recording — speaker 1 at minute 2 is speaker 1 at minute 40.
Timings
Word-level timestamps
Every word carries a start, end and confidence. Enough to build a scrubbing transcript without a second pass.
Formats
json · verbose_json · text · srt · vtt
Subtitles come out ready to attach.
Hints
Vocabulary biasing
Pass a prompt of order IDs, product names or jargon to pull rare words into range.
Languages
Auto-detect or pin
Set language=auto, or force one when you already know.
Sample
A real support call.
| Start | Speaker | Text | Conf |
|---|---|---|---|
| 0:00.2 | spk_1 | हाँ जी, मेरा order अभी तक deliver नहीं हुआ है। | 96% |
| 0:03.4 | spk_1 | Tracking page पर कुछ update नहीं है। | 88% |
| 0:06.1 | spk_2 | I can check that right away. | 97% |
| 0:08.0 | spk_2 | Your parcel left the Ahmedabad hub this morning. | 91% |
Confidence is returned per word and per segment, so you can route anything below your threshold to a human instead of acting on it.