Uzbek Speech-to-Text API
Clarion is an API that converts Uzbek speech to text with punctuation and word-level timestamps. It gets 8.1% of words wrong on real conversations, costs $0.01 per minute of audio and works with the official OpenAI SDK.
Accuracy
Measured on Clarion's open benchmark: 311 clips of spontaneous Uzbek speech from podcasts, interviews and multi-speaker conversations, with transcripts checked by hand.
uzasr-large gets 8.1% of words wrong. Untuned GigaAM-large gets 14.4%, and Whisper fine-tuned for Uzbek gets 38.6%.
| Model | Parameters | WER | CER |
|---|---|---|---|
| uzasr-largeClarion | 600M | 8.1% | 2.5% |
| uzasr-baseClarion | 220M | 9.4% | 3.0% |
| GigaAM-large, not fine-tuned | 600M | 14.4% | 5.3% |
| whisper-large-v3-turbo-uzbek | 809M | 38.6% | 18.5% |
WER is the share of misrecognized words. Lower is better. Punctuation is not counted.
Full results and method →Price
No subscription fee. Per-second billing with a 15-second minimum per request.
File transcription costs $0.01 per minute, realtime mode $0.025. First 30 minutes free. That is 37.5% cheaper than Google Speech-to-Text V2 ($0.016 per minute). Whisper is cheaper through the OpenAI API, but Whisper fine-tuned for Uzbek gets 38.6% of words wrong on the same benchmark.
| Service | Per file minute | WER |
|---|---|---|
| Clarion | $0.01 | 8.1% |
| Google Speech-to-Text V2 | $0.016 | — |
| Whisper | $0.006 | 38.6% |
Prices are public per-minute rates. Google's WER has not been measured on our benchmark; the Whisper row shows whisper-large-v3-turbo-uzbek.
All plans and the cost calculator →What the API does
Files over HTTP, live audio over WebSocket. The response carries text, timestamps and sound events.
- Synchronous transcription
- Files up to 10 minutes. The response comes back in the same request.
- Async jobs
- Files up to 3 hours. Create a job and poll its status.
- Realtime mode
- WebSocket: stream PCM in, receive text events out.
- Punctuation
- Periods, commas and capitals from intonation, with no post-processing.
- Timestamps
- Start and end of every word: highlighting, subtitles, jumping to a moment in the recording.
- Sounds
- Music, laughter and applause arrive in the events field with start and end times.
- Code-switching
- English and Russian words inside an Uzbek sentence.
- Audio is deleted after transcription
- We use recordings only to transcribe them. Logs and metrics contain neither audio nor text.
Request example
OpenAI-compatible API: change one line in your existing code.
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)Self-hosting and startups
The same model that runs in the API can run inside your own network.
On your own server
The model and API run in your environment. Audio never leaves your network.
More about self-hosting →Clarion for startups
A free license to run our Uzbek speech recognition model on your own server: 12 months or until your first paying customer.
Program terms →
Frequently asked questions
Short answers on accuracy, price and integration.
- How accurate is Clarion on Uzbek speech?
- uzasr-large gets 8.1% of words wrong on an open benchmark of 311 clips of spontaneous Uzbek speech: podcasts, interviews and multi-speaker conversations. Punctuation is not counted.
- How much does Uzbek speech-to-text cost?
- $0.01 per file minute and $0.025 per realtime minute, billed by the second. First 30 minutes free.
- Can it transcribe speech in real time?
- Yes, over WebSocket with uzasr-large. Text appears while the person is speaking, and a phrase is finalized after a 0.8 s pause.
- Does it recognize Russian and English words?
- Yes. The model recognizes Russian and English words inside an Uzbek sentence. Text comes back in Latin script, with punctuation.
- How do I connect to the API?
- The API is OpenAI-compatible: in the official OpenAI SDK for Python or Node.js, set the Clarion base URL and the uzasr-large model. You get a key after signing up.