Kani TTS 2 is a compact open‑source text to speech model that aims to bring realistic voice synthesis and cloning to regular consumer hardware rather than datacenter‑class machines. With roughly 400 million parameters and a 22 kHz output, it runs in about 3 GB of VRAM while still offering zero‑shot voice cloning from a short reference clip and support for English and early multilingual variants. The system is built around a token based audio representation and a two stage pipeline that separates semantic/acoustic token generation from final waveform decoding, which helps it keep latency low enough for near real time conversations on modern GPUs. In practice, users report a real time factor around 0.2 on high end cards, so generating 10 seconds of speech takes roughly 2 seconds, and typical VRAM usage stays slightly under the advertised 3 GB limit.
The model also focuses on capturing accents and local speech styles, offering region flavored English voices and the option to mimic a custom speaker without per speaker fine tuning, which makes it attractive for tools, personal assistants and content workflows that need distinctive but reproducible voices.
At the same time, its relatively small size means it will not match very large proprietary systems in extreme expressiveness or multilingual coverage, and its strong voice cloning raises the usual ethical concerns around consent, impersonation and deepfake style misuse, so it is better seen as a practical, well engineered option in the open ecosystem rather than a magic solution to every TTS problem.
