NineNineSix releases KaniTTS2-en, a 400M English TTS with voice cloning
KaniTTS2-en is a 400M English TTS on LFM2 350M with cloning via a speaker embedder, about 3 GB VRAM; LFM 1.0 licence.
NineNineSix published KaniTTS2-en, an English text-to-speech model of 400M parameters on a LiquidAI LFM2 350M backbone. The card describes a two-stage pipeline pairing the language model with an FSQ audio codec (Nvidia NanoCodec), 22 kHz output, and voice cloning through a speaker embedder. It is aimed at real-time conversation.
The only figures are speed and training details. On an Nvidia RTX 5080 the card gives a real-time factor of about 0.2 with a custom executor, using about 3 GB of VRAM; it recommends batches of 8 to 16 and says it is optimised for NVIDIA Blackwell GPUs. Training used about 10k hours of data on 8 H100 GPUs for 6 hours. No quality benchmarks are given.
Limits on the card: quality degrades beyond about 40 seconds of input, it is primarily English, and vLLM integration and the tech report are marked "Coming soon". The licence is listed as lfm1.0. The card also lists prohibited uses, including impersonation without consent. The card gives no release date; the date used is the Hugging Face repo creation date, 12 February 2026.
A small, fast model with its evidence still to come.
Sources
Editor's first pass (AI-assisted)
- Real 4: Weights are public; vLLM integration and the tech report are marked coming soon.
- Useful 3: Real-time English with cloning at about 3 GB VRAM, tuned for one GPU family.
- Notable 2.5: A small-team model; no quality benchmarks given.
- Novel 3: A speaker embedder on an LFM2 backbone plus NanoCodec is a modest recombination.
Comments