- Updated: February 15, 2026
- 2 min read
Kani‑TTS‑2: Open‑Source 400M‑Parameter Text‑to‑Speech Model with Voice Cloning Runs on 3 GB VRAM
UBOS Tech brings you the latest breakthrough in AI‑driven speech synthesis – Kani‑TTS‑2, a 400‑million‑parameter open‑source text‑to‑speech (TTS) model that delivers high‑fidelity voice generation and zero‑shot voice cloning while fitting comfortably into just 3 GB of VRAM. This new model, released by the research team behind the original Kani‑TTS, pushes the boundaries of what developers can achieve on consumer‑grade GPUs.
Key Highlights
- Compact yet powerful: 400 M parameters, comparable to many commercial TTS services, but optimized to run on a single 3 GB GPU.
- Zero‑shot voice cloning: Generates realistic speech in a target voice after hearing only a few seconds of reference audio.
- Fast training speed: Achieves full‑model convergence in under 12 hours on an RTX 3060.
- Open‑source license: Released under the Apache 2.0 license, encouraging community contributions and commercial use.
Technical Overview
Kani‑TTS‑2 builds on a diffusion‑based architecture, combining a lightweight encoder with a high‑resolution mel‑spectrogram decoder. The model leverages advanced diffusion techniques to produce natural prosody and expressive intonation. Training was performed on the LibriTTS and VCTK corpora, with data augmentation strategies that improve robustness across accents and speaking styles.
Why It Matters for Developers
For startups and hobbyists constrained by hardware budgets, Kani‑TTS‑2 offers a cost‑effective alternative to cloud‑based TTS APIs. Its small memory footprint means it can be deployed on edge devices, enabling offline voice assistants, interactive audiobooks, and real‑time dubbing applications without sacrificing quality.
Getting Started
The model, checkpoints, and inference scripts are available on GitHub. Detailed documentation walks you through installation, fine‑tuning on custom datasets, and integrating the API into your existing pipelines. For a quick demo, check out our live showcase.
Looking Ahead
The release of Kani‑TTS‑2 signals a shift toward more accessible, high‑quality speech synthesis. Combined with UBOS Tech’s suite of AI tools, developers can now build end‑to‑end voice solutions that were previously limited to large enterprises.
Read the original announcement on MarkTechPost for deeper technical details.
Andrii Bidochko
CTO UBOS
Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.