- 8.1%of words misrecognized in real-world speech (WER)
- ×4.7fewer errors than Whisper fine-tuned for Uzbek
- < 1 sto transcribe 10 seconds of audio on one warm NVIDIA T4
- 1,800+hours of Uzbek speech in the training data
Models
Two models from one family: maximum accuracy or a lower price.
uzasr-large
Our most accurate model and the default choice
- WER
- 8.1%
- parameters
- 600M
- File API
- Realtime mode
- Word timestamps
Accuracy on real-world speech
Measured on Clarion's open benchmark: 311 clips of spontaneous Uzbek speech from podcasts, interviews and multi-speaker conversations, with transcripts checked by hand.
WER is the share of misrecognized words. Lower is better. Punctuation is not counted.Open benchmark on Hugging FaceFull results and method →How we got here →
Your first request in three steps
The official OpenAI SDK. No new libraries needed.
- 1
Get a key
30 free minutes right after sign-up.
- 2
Install the SDK
The same package you use for OpenAI.
pip install openainpm install openai
- 3
Send audio
Set the Clarion base URL and the uzasr‑large model. Everything else works as with OpenAI.
import os from openai import OpenAI client = OpenAI( base_url="https://api.uzasr.uz/v1", api_key=os.environ["UZASR_API_KEY"],) result = client.audio.transcriptions.create( model="uzasr-large", file=open("call.mp3", "rb"), response_format="verbose_json", timestamp_granularities=["word"],)print(result.text)import fs from "node:fs"import OpenAI from "openai" const client = new OpenAI({ baseURL: "https://api.uzasr.uz/v1", apiKey: process.env.UZASR_API_KEY,}) const result = await client.audio.transcriptions.create({ model: "uzasr-large", file: fs.createReadStream("call.mp3"), response_format: "verbose_json", timestamp_granularities: ["word"],})console.log(result.text)What's in the response
One API response. Pick what to highlight.
How it works
OpenAI-compatible API: change one line in your existing code.
Who it's for
Typical customer use cases.
Test it on your own recordings
Upload a call and compare the result with your current solution.