Kokoro 82M v1.0

Kokoro-82M-v1.0-ONNX is a text-to-speech (TTS) model that converts text input into audio output, designed for efficient, high-quality speech synthesis with 82 million parameters. It is based on the StyleTTS 2 architecture (arXiv 2306.07691) combined with ISTFTNet (arXiv 2203.02395), operating as a decoder-only model without diffusion or encoder components. The model supports 8 languages and offers 54 voices, including American and British English male and female variants. It is resilient to quantization, enabling efficient deployment at reduced sizes while maintaining audio quality; for example, the 8-bit quantized version (model_quantized.onnx) is 92.4 MB, compared to 326 MB for the full fp32 model. Training cost approximately $1000 for 1000 A100 80GB GPU hours, using a few hundred hours of permissive and non-copyrighted audio data, including public domain, Apache/MIT licensed, and synthetic sources. The model is licensed under Apache-2.0. For text preprocessing, the companion G2P engine misaki supports English, Japanese, Korean, Chinese, and Vietnamese. Audio samples for each voice and quantization variant are available below.

Voices and Samples

All samples use the sentence: "Life is like a box of chocolates. You never know what you're gonna get."

Voice Name	Nationality	Gender	Sample
af_heart	American	Female
af_alloy	American	Female
af_aoede	American	Female
af_bella	American	Female
af_jessica	American	Female
af_kore	American	Female
af_nicole	American	Female
af_nova	American	Female
af_river	American	Female
af_sarah	American	Female
af_sky	American	Female
am_adam	American	Male
am_echo	American	Male