🎙️ Uzbek Voice AI

I built a live, streaming voice assistant that lets Uzbek speakers ask questions and get spoken answers — then ran a full research track to find (or build) the best Uzbek text-to-speech voice, since Uzbek is a low-resource language with very few options.

🔊 Listen — Uzbek TTS demos

Same sentences synthesized by the best off-the-shelf model vs. my fine-tuned voice. All audio is generated from text — no human recording.

🧭 The approach

• Built the voice agent first — a streaming pipeline (speech-in → recognition → LLM → speech-out) with voice-activity detection, instant filler words, and barge-in (the user can interrupt), targeting under ~2 s to first audio.
• Benchmarked Uzbek TTS — Uzbek has almost no good open voices, so I tested 5 options and A/B-compared the audio.
• Built a 60-hour training corpus — cloned and cleaned the FeruzaSpeech studio dataset, automated end-to-end training on Kaggle GPUs via the API.
• Fine-tuned a VITS/MMS model on 5.2 h of clean single-speaker Uzbek (7,080 steps, ~4 h on a Tesla P100), debugging a deep stack of environment issues (CUDA/GPU-arch mismatches, library version conflicts, GAN training quirks).
• Measured honestly — the strong base model actually beat my fine-tune on limited data, a realistic and useful negative result.

📊 Models evaluated

🛠️ Tech stack

Python · PyTorch · HuggingFace Transformers · VITS / Meta MMS · Chatterbox · FastAPI + WebSockets · gRPC (Yandex SpeechKit STT/TTS v3) · Silero/energy VAD · Kaggle GPU automation (Kaggle API) · soundfile / datasets.

💡 Key takeaways

• Right tool > biggest tool. A small purpose-built VITS beat a 0.5 B voice-cloning model for a single-speaker language task.
• Start from a model that already knows the language. Teaching Uzbek from zero failed; adapting an Uzbek-capable base was the workable path.
• Licensing is a first-class constraint. The best open weights are non-commercial (CC-BY-NC) — production needs a commercial voice or owned data.