Kitten-TTS 2
An open-source low-bit speech synthesis model from KittenML that employs ternary quantization (1.58 bits). It supports real-time inference on CPU, zero-shot voice cloning, multilingual output and emotional control.
Timeline (last 90 days)
KittenML open-sourced the code and weights of Kitten-TTS 2, hosted on GitHub (KittenML/KittenTTS) and Hugging Face (KittenML/kitten-tts-2).
Kitten-TTS 2 is a 1.58-bit TTS model whose linear layer weights are ternary {-1, 0, +1}, paired with scaling factors with a group size of 128.
Kitten-TTS 2 can generate speech in real time on CPU alone, with no GPU required.
Kitten-TTS 2 supports zero-shot voice cloning, requiring less than 5 seconds of reference audio to clone a target timbre without fine-tuning.
Kitten-TTS 2 supports voice cloning in more than 20 languages and supports cross-lingual cloning (e.g., synthesizing English speech with a Chinese reference audio to produce an accent effect).
Kitten-TTS 2 supports emotion and stress control, allowing guidance of the overall expressive style and emphasis on specific words.
Kitten-TTS 2 is a completely new architectural rewrite from V0.8 to 2, with changes including improved speech quality, introduction of ternary weights, zero-shot voice cloning, multilingual support, and emotion/stress control.
The ternary kernel degrades multiplication into addition, negation, or zeroing, avoiding floating-point multiply-add operations and reducing CPU computation overhead, which is the underlying reason Kitten-TTS 2 can perform real-time inference on consumer-grade CPUs.
While 1.58-bit quantization improves efficiency, there is still room for trade-offs in the upper limit of audio quality compared with full-precision large models.
The publisher of Kitten-TTS 2 positions it as "an open-source alternative for those short on GPU resources," emphasizing ease of self-deployment and local operation.
All Facts (10)
The publisher of Kitten-TTS 2 positions it as "an open-source alternative for those short on GPU resources," emphasizing ease of self-deployment and local operation.
50%UnverifiedKitten-TTS 2 is a 1.58-bit TTS model whose linear layer weights are ternary {-1, 0, +1}, paired with scaling factors with a group size of 128.
50%UnverifiedKitten-TTS 2 can generate speech in real time on CPU alone, with no GPU required.
50%UnverifiedKitten-TTS 2 supports zero-shot voice cloning, requiring less than 5 seconds of reference audio to clone a target timbre without fine-tuning.
50%UnverifiedKitten-TTS 2 supports voice cloning in more than 20 languages and supports cross-lingual cloning (e.g., synthesizing English speech with a Chinese reference audio to produce an accent effect).
50%UnverifiedKitten-TTS 2 supports emotion and stress control, allowing guidance of the overall expressive style and emphasis on specific words.
50%UnverifiedKitten-TTS 2 is a completely new architectural rewrite from V0.8 to 2, with changes including improved speech quality, introduction of ternary weights, zero-shot voice cloning, multilingual support, and emotion/stress control.
50%UnverifiedThe ternary kernel degrades multiplication into addition, negation, or zeroing, avoiding floating-point multiply-add operations and reducing CPU computation overhead, which is the underlying reason Kitten-TTS 2 can perform real-time inference on consumer-grade CPUs.
50%UnverifiedWhile 1.58-bit quantization improves efficiency, there is still room for trade-offs in the upper limit of audio quality compared with full-precision large models.
50%UnverifiedKittenML open-sourced the code and weights of Kitten-TTS 2, hosted on GitHub (KittenML/KittenTTS) and Hugging Face (KittenML/kitten-tts-2).
50%