kani-tts

Hi everyone! Thanks for the awesome feedback on our first KaniTTS release! We’ve been hard at work, and released kani-tts-370m. It’s still built for speed and quality on consumer hardware, but now with expanded language support and more English voice options. What’s New: Multilingual Support: German, Korean, Chinese, Arabic, and Spanish (with fine-tuning support). Prosody and naturalness improved across these languages. More English Voices: Added a variety of new English voices. Architecture: Same two-stage pipeline (LiquidAI LFM2-370M backbone + NVIDIA NanoCodec). Trained on ~80k hours of diverse data. Performance: Generates 15s of audio in ~0.9s on an RTX 5080, using 2GB VRAM. Use Cases: Conversational AI, edge devices, accessibility, or research. It’s still Apache 2.0 licensed, so dive in and experiment. Repo: https://github.com/nineni

Hey HN, We've been tinkering with TTS models for a while, and I'm excited to share KaniTTS – an open-source text-to-speech system we built at NineNineSix.ai. It's designed for speed and quality, hitting real-time generation on consumer GPUs while sounding natural and expressive. Quick overview: Architecture: Two-stage pipeline – a LiquidAI LFM2-350M backbone generates compact semantic/acoustic tokens from text (handling prosody, punctuation, etc.), then NVIDIA's NanoCodec synthesizes them into 22kHz waveforms. Trained on ~50k hours of data. Performance: On an RTX 5080, it generates 15s of audio in ~1s with only 2GB VRAM. Languages: English-focused, but tokenizer supports Arabic, Chinese, French, German, Japanese, Korean, Spanish (fine-tune for better non-English prosody). Use cases: Conversational AI, edge devices, accessibility, or resea