Full articles are free to read through December 10, 2026 (IST). Subscriptions remain available.
By S Vijay Kumar, International Editor | 18th February 2026
In a notable development from Central Asia's emerging tech scene, the Kyrgyz AI startup NineNineSix backed by the High Technology Park of the Kyrgyz Republic has launched Kani TTS 2, an open-source text-to-speech (TTS) model that extends continuous speech generation to up to 40 seconds, supports zero-shot voice cloning, and emphasizes accessibility for low-resource languages.
Announced this week with strong early traction on Hugging Face (where related checkpoints rank among top TTS models), the release comes amid growing global demand for efficient, non-proprietary voice AI alternatives to closed systems from companies like OpenAI, Google, and ElevenLabs. Kani TTS 2 builds directly on its predecessor, addressing key limitations in generation length and stability while preserving a lightweight footprint suitable for edge and server deployments.
The model, with approximately 400 million parameters, delivers 22kHz audio at a real-time factor of roughly 0.2 on consumer GPUs like the NVIDIA RTX 5080, requiring only about 3 GB of VRAM. Its two-stage architecture integrates LiquidAI's LFM2-350M language model backbone with NVIDIA's NanoCodec (FSQ-based) for efficient audio tokenization. A key innovation frame-level position encodings allows audio tokens within the same frame to share positional IDs, improving coherence over longer sequences without heavy masking or computational overhead.




