kale-tts-1-nano

kale-tts-1-nano has only 750 thousand parameters, with an int8 file of 14.8 MB for CPU and an fp32 file of 30.6 MB for browser WebGPU. The 10 voice styles are designed by Maya-1 based on a single text description and do not correspond to real people; first, 204.7 hours of training audio were recorded using Maya-1, then the prosody planning component was distilled from kale-tts-1-base, which is fine-tuned from Kokoro-82M. The vocalization part uses a small harmonic plus noise generator, without diffusion or flow matching. The authors claim it runs at 44 times real-time on the CPU of a Samsung Galaxy Tab S10 Lite, with the first sentence spoken in 0.09 seconds, and that words uncertain to the recognizer are only 1.6 percentage points more frequent than in Maya-1's original voice.

kale-tts-1-nano has only 750 thousand parameters, with an int8 file of 14.8 MB for CPU and an fp32 file of 30.6 MB for browser WebGPU. The 10 voice styles are designed by Maya-1 based on a single text description and do not correspond to real people; first, 204.7 hours of training audio were recorded using Maya-1, then the prosody planning component was distilled from kale-tts-1-base, which is fine-tuned from Kokoro-82M. The vocalization part uses a small harmonic plus noise generator, without diffusion or flow matching. The authors claim it runs at 44 times real-time on the CPU of a Samsung Galaxy Tab S10 Lite, with the first sentence spoken in 0.09 seconds, and that words uncertain to the recognizer are only 1.6 percentage points more frequent than in Maya-1's original voice.

Why it matters Trained for fully local speech on children's devices, it uses a single file and generates one sentence per forward pass; the audio data for the 10 training voices is also released under CC BY 4.0.

Hugging Face likes 14 · Hugging Face trending score 14 · Hugging Face downloads 256

License: Commercial use allowed · ComfyUI: Not available · VRAM: VRAM not specified

Reasons

  • There is an online demo on Hugging Face where you can listen to all voice styles directly; the training dataset kale-tts-1-data is public
  • Apache-2.0, commercial use allowed; for better audio quality, consider the slightly larger kale-tts-1-tiny from the same series

Usage

  • First listen in the Hugging Face Space KaleComputer/kale-tts-1, then download kale-tts-1-nano.onnx to run locally using ONNX Runtime (use the fp32 file with WebGPU in the browser)

Risks

  • Supports English only; pitch still differs from Maya-1's original voice (the authors report a melody difference of 7.46 percentage points)
  • The model card still contains a 'draft, unpublished' note at the beginning, so content may change; all performance figures are self-tested by the authors

Open

抱脸 text-to-speech 上升榜

Explore the latest AI Radar updates