Kitten TTS 2 on BestTTS: Ternary TTS with Voice Cloning

This BestTTS page is an independent reference for Kitten TTS 2, a 1.7-billion-parameter speech language model from KittenML. BestTTS is a free directory of speech models; it does not provide the model. The page explains how it works: text becomes S3 codec tokens that a vocoder converts to 24 kHz audio, with ternary transformer weights stored at about 1.58 bits. It covers reference voice cloning from a 5-30 second recording, 123 voices and 16 languages with named voices per the model card, CPU execution including a C++ runtime, and getting-started steps (pip install kittenml). The Stellon Labs Community License requires registration for commercial use and is not unrestricted open source. No benchmark is claimed.

posted in ai models | part of Week 45, 2026 digest