BirOvoz — Central Asian Voice Benchmark

Research and a local voice demo for Uzbek and Kazakh. The benchmark evaluates speech recognition, translation and dubbing; the demo runs an STT → LLM → TTS pipeline with streaming English-to-Uzbek/Kazakh dubbing.

• Research scope and methodology

• Frozen benchmark report

• Local setup and provider/model configuration

• Current handoff and release acceptance

• Architecture and data flows

Architecture at a glance

See the architecture notes for provider boundaries, the separate benchmark flow, and what still requires real-provider verification.

Quick start

From the repository root, with Python 3.11+ and ffmpeg on PATH:

python -m venv product/backend/.venv
# macOS/Linux
source product/backend/.venv/bin/activate
# Windows PowerShell instead: .\product\backend\.venv\Scripts\Activate.ps1
python -m pip install -r product/backend/requirements.txt

Copy product/backend/.env.example to product/backend/.env, then run:

python -m uvicorn product.backend.main:app --host 127.0.0.1 --port 8000

Open http://127.0.0.1:8000. Without an OpenAI key, the demo uses deterministic mock responses. MOCK results are not real accuracy measurements. Configure live providers in the local .env; keep credentials and voice/model assets outside version control. Local voice clone currently supports Uzbek only.

CPU verification

python -m pip install pytest pytest-asyncio httpx jiwer
python -m pytest tests product/backend/tests -q

These tests use provider/model substitutes. Real voice quality, Windows/NVIDIA compatibility, cold/warm latency and microphone playback acceptance remain manual checks listed in the handoff. The leaderboard serves frozen research results and does not recompute them from demo uploads.