Kani-TTS-2 Overview

summarize in a paragraph with no dashes: https://www.marktechpost.com/2026/02/15/meet-kani-tts-2-a-400m-param-open-source-text-to-speech-model-that-runs-in-3gb-vram-with-voice-cloning-support/

Kani-TTS-2 is a newly released open-source text-to-speech model designed for efficiency and accessibility, developed by nineninesix.ai and available on Hugging Face. It uses a compact 400 million-parameter architecture that treats audio as discrete tokens, enabling high-quality, human-like speech synthesis with voice cloning support while running on just about 3 GB of GPU VRAM. The model was trained on roughly 10,000 hours of speech in just six hours on 8 NVIDIA H100 GPUs and can perform zero-shot voice cloning from short audio clips without fine-tuning, with real-time generation speeds that make it practical for consumer-grade GPUs under an Apache 2.0 license. citeturn0view0

now do the same with : https://huggingface.co/nineninesix/kani-tts-2-en

nineninesix’s **KaniTTS2-en** is an open-source English text-to-speech model on Hugging Face that uses a two-stage pipeline combining a language model with an FSQ audio codec to generate human-like speech at a 22 kHz sample rate from input text with just 400 million parameters. It can run efficiently on consumer GPUs with about 3 GB of VRAM and supports voice cloning by embedding a reference voice for personalized output, making it suitable for real-time conversational AI and research use cases, though performance can drop on very long inputs and its English optimization means other languages may need extra training. citeturn0view0