README.md

metadata

title: Uzbek STT
emoji: 🎙️
colorFrom: green
colorTo: indigo
sdk: docker
app_port: 7860
pinned: false
license: apache-2.0

Uzbek Speech-to-Text — a deployable service

An end-to-end Uzbek speech-to-text service: a documented REST API, a demo UI, and an honest multi-engine benchmark — deployed to a live public URL on a $0 budget (no paid server, no paid API, no GPU). Built to demonstrate shipping and operating software, not just writing it.

Live demo: coming with the M1 deploy → https://huggingface.co/spaces/bintuulugbek/Uzbek-STT API docs: /docs on the same URL.

---

Why this project

Uzbek is an underserved language for speech tech. This project takes verified open Uzbek ASR models and wraps them in production-grade engineering: a clean FastAPI service, a swappable engine seam, problem+json errors, and a reproducible benchmark with real WER/CER numbers measured by us (not quoted from model cards).

Architecture

flowchart TD
    U[User: mic / upload / API call] -->|audio| FE[Gradio UI '/']
    U -->|POST audio| API[FastAPI '/api/*']
    FE --> SVC[TranscriptionService]
    API --> SVC
    SVC --> SEAM[Engine seam - Protocol]
    SEAM --> E1[whisper-small-uz  FAST]
    SEAM --> E2[rubaistt_v2_medium  ACCURATE]
    API --> DOCS[OpenAPI '/docs']
    subgraph Offline [Benchmark - offline]
        BM[harness] --> SEAM
        BM --> E3[xls-r-uzbek  benchmark-only]
        BM --> M[(metrics.csv + plots)]
    end
    subgraph Deploy [HF Spaces - Docker SDK - FREE]
        API
        FE
        DOCS
    end

The $0 stack

Need	Free tool
API + demo hosting	Hugging Face Spaces (Docker SDK, free CPU, 16 GB)
Models	Hugging Face Hub (Apache-2.0 Uzbek ASR)
Benchmark data	FLEURS uz_uz test split
Code	GitHub

Engines

Role	Model	Deployed
Fast (default)	BlueRaccoon/whisper-small-uz	✅
Accurate	islomov/rubaistt_v2_medium	✅
Benchmark baseline	lucio/xls-r-uzbek-cv8	offline only

API

Endpoint	Purpose
GET /api/health	readiness (engines loaded)
POST /api/transcribe	transcribe audio → text + word timings
GET /api/examples	bundled FLEURS clips
POST /api/transcribe_example	transcribe a bundled clip + live WER
GET /api/benchmark	latest WER/CER per engine

Errors follow RFC 9457 (application/problem+json). All times are integer milliseconds.

Why there's no login

Deliberate: this is a public evaluation demo with no user data to protect. Auth would only add friction. Abuse is bounded by per-IP rate limiting + upload-size and duration caps instead.

Roadmap

• M1 — deployable skeleton: live /health + /docs on HF Spaces
• M2 — fast engine wired: real Uzbek transcription via /api/transcribe
• M3 — Gradio UI (4 tabs) + one-click FLEURS examples with live WER
• M4 — 3-engine benchmark on FLEURS + error dashboard
• M5 — docs polish
• v2 — long-audio chunking

Local development

pip install -e ".[dev]"
uvicorn app.main:app --reload --port 7860
# open http://localhost:7860/docs

License

Apache-2.0. All deployed models are Apache-2.0. This is original work; it follows common FastAPI architectural conventions but contains no third-party application code.