All Models
H

Kokoro

by hexgradPublicapache-2.0
README.md

Kokoro-82M

Kokoro is an open-weight text-to-speech model published by hexgrad on Hugging Face. The upstream card presents it as an 82M-parameter model focused on lightweight, fast, cost-efficient speech generation while still being practical for local and production use.

LA Studio installs Kokoro as a virtual model under hexgrad/Kokoro-82M. The downloadable GGUF artifacts are hosted in conversion repositories:

  • cstr/kokoro-82m-GGUF for the Kokoro backbone GGUF.
  • cstr/kokoro-voices-GGUF for preset speaker voice packs.

Links

  • Upstream model card: https://huggingface.co/hexgrad/Kokoro-82M
  • Upstream GitHub: https://github.com/hexgrad/kokoro
  • Upstream demo: https://hf.co/spaces/hexgrad/Kokoro-TTS
  • Samples: https://huggingface.co/hexgrad/Kokoro-82M/blob/main/SAMPLES.md
  • Voices: https://huggingface.co/hexgrad/Kokoro-82M/blob/main/VOICES.md

Model Facts

  • Task: Text-to-Speech
  • Parameters: 82M
  • License: Apache-2.0
  • Current upstream release noted by the model card: v1.0, published 2025-01-27
  • Languages and voices noted for v1.0: 8 languages and 54 voices
  • Architecture: StyleTTS 2 with an iSTFTNet-derived, decoder-only TTS stack
  • Upstream SHA256 for v1.0: 496dba118d1a58f5f3db2efc88dbdc216e0483fc89fe6e47ee1f2c53f18ad1e4

Training Notes

The upstream card states that Kokoro was trained on permissive or non-copyrighted audio data with IPA phoneme labels. It lists public-domain audio, permissively licensed audio, and synthetic audio generated by closed-source TTS models as examples of training data sources. The card estimates the total training set at a few hundred hours and reports roughly 1000 A100 80GB GPU-hours across v0.19 and v1.0.

LA Studio Notes

Kokoro in LA Studio is a preset-voice TTS workflow. It needs one backbone model file and one installed voice pack:

  • Default backbone: kokoro-82m-f16.gguf
  • Default voice pack: kokoro-voice-af_heart.gguf

The CrispASR Kokoro runtime depends on espeak-ng for phonemizer support. Kokoro voice cloning from reference audio is not enabled in LA Studio; use the packaged voice GGUF files instead.

Configure Workflow Parameters

Configure these custom parameters to fine-tune the speech and voice workflows.

Temperaturefloat

Controls randomness of speech generation. Higher values increase variety.

1
Seedint

Random seed for reproducibility. Set to -1 for random.

-1
Max Codec Stepsint

Maximum decoding steps for the neural voice audio synthesizer.

100