Pay per minute
No subscription fee. Per-second billing with a 15-second minimum per request.
Early-stage startups get a free license to run the model on their own server.Program terms
Estimate your month
Estimate at current rates. You pay only for seconds of transcribed audio.
On your own server
The model and API run in your environment. Audio never leaves your network.
Get a quote
Custom license
One NVIDIA T4, uzasr-large
- 1 minute of audio
- 0.75–3 s
- full transcription
- 5 minutes of audio
- 2.7–5.5 s
- with timestamps and events
- 1 hour of audio
- ≈1 min
- async job
- WER on real-world speech
- 8.1%
- the same model as in the cloud
- GPU
- NVIDIA GeForce RTX 2060 or newer (RTX 3060, 4060, 5060) or data center T4, L4
- VRAM
- 4.6 GB for uzasr-large, 2.7 GB for uzasr-base
- CPU and RAM
- 2 cores and 8 GB per worker
- Driver
- NVIDIA with CUDA 12.8+ and NVIDIA Container Toolkit
- Software
- Docker 29+, Docker Compose 2.24+
By the seconds of actual audio, with a 15-second minimum per request. Failed requests aren't billed.
From a prepaid balance. Payme and Click are being connected; for now, companies pay by invoice.
uzasr-large is more accurate (WER 8.1%) and supports realtime mode. uzasr-base is lighter (WER 9.4%) and transcribes files only.
Synchronous up to 10 minutes of audio, up to 3 hours as a job. Up to 8 concurrent requests per key. Realtime sessions up to 5 minutes, up to 4 sessions at once.
The GPU sleeps after 30 seconds idle, and the first request wakes it up: 5–25 seconds, usually about 12. On your own server the GPU runs all the time.
Yes. The model, API and realtime mode run on your GPU, and audio never leaves your network. Requirements and components are in the “On your own server” section.
Audio is deleted right after transcription, and the backup copy is kept no longer than 24 h. Synchronous responses aren't stored; async job results are kept for 7 days.