Back to Case Studies

Private Uzbek Text-to-Speech API Deployed for Regional Voice AI Application

Executive Summary

GenAI Protos deployed Piper Uzbek TTS as a private, CPU-based text-to-speech service with custom FastAPI routes and gateway pass-through access. The service supports 22 kHz Uzbek speech synthesis, 16 speaker profiles, low memory usage, and zero GPU dependency for regional voice applications.

01Client Requirement

Most multilingual AI systems still underperform for lower-resource, region-specific languages like Uzbek. The requirement was a native Uzbek speech capability that could run privately, stay lightweight, and plug into wider AI assistant workflows. Specifically, the service needed to: Provide a private Uzbek TTS endpoint for regional language applications Run on CPU to avoid competing for GPU resources with larger AI workloads Expose a custom API shape through governed gateway routes, rather than forcing it into a standard speech interface Support multiple speaker profiles and adjustable speech speed Be documented well enough for repeatable deployment

Section image

02Our Approach

Our Approach We packaged Piper Uzbek into a custom private TTS service with a simple, product-ready API surface: Custom FastAPI service wrapped around Piper VITS ONNX inference, since the model's API shape differs from standard chat or speech endpoints Dedicated routes for text-to-speech, streaming, voice listing, model metadata, and health checks Gateway pass-through integration so requests are authenticated and governed by the same pattern used across the wider GenAI Protos model fleet CPU-only deployment that loads quickly and runs on demand per request Operational documentation covering voice IDs, speed controls, response headers, rate limits, and use cases

03Outcome

  • Native Uzbek speech at 22,050 Hz delivered through a private API
  • 16 speaker profiles, including 2 hero Tashkent voices and 14 community speakers
  • 74 MB model size and ~150 to 200 MB RAM footprint
  • Zero GPU required, reducing contention for larger AI workloads
  • Consistent, governed access through the same gateway used by the rest of the model fleet
  • Real-time factor of 0.1 to 0.3, so audio generates much faster than playback
  • 60 requests per minute and 30,000 tokens per minute rate limits configured
  • A reusable pattern that can pair with regional speech-to-text models for full voice-to-voice assistants

Frequently asked questions

It is a self-hosted service that converts Uzbek text into natural-sounding speech without sending data to a public third-party API. GenAI Protos built it using the Piper VITS Uzbek model served through FastAPI.

No. This deployment runs entirely on CPU using ONNX inference, with about 150 to 200 MB of RAM and no GPU dependency.

The service supports 16 speaker profiles: 2 hero voices from Tashkent and 14 community speakers. Speech speed is also adjustable.

The real-time factor is typically 0.1 to 0.3, which means audio is generated three to ten times faster than it takes to play back.

Common use cases include voice assistants, IVR systems, accessibility audio, education platforms, and localised product or content experiences for Uzbek-speaking users.

Build regional voice AI that works on private infrastructure.

Tell us what you want to build, test, or improve. We'll help you define a practical first step.

Explore our Custom Private AI and Edge Solutions