coke123/go-tts-models

Go TTS Models

Pre-converted ONNX model files for chuccp/go-tts. This repository removes the need to download and quantize Kokoro before running the app.

Kokoro-82M

File	Size	Purpose
kokoro/kokoro-v1.0-quantized.onnx	88.08 MiB	Default Kokoro v1.0 quantized model
kokoro/kokoro-v0_19.onnx	310.45 MiB	Original FP32 model
kokoro/kokoro-v0_19-uint8.onnx	169.25 MiB	Balanced UINT8 model
kokoro/kokoro-v0_19-int8.onnx	87.68 MiB	Dynamically quantized INT8 model used by default
kokoro/voices-v1.0.bin	24.91 MiB	Kokoro voice profiles (NPZ format)

The v1.0 quantized model is the smallest model that passed Chinese ASR regression and is used by default. The v0.19 INT8 model was produced from the FP32 graph with ONNX Runtime dynamic quantization (QInt8, MatMulConstBOnly=true). It keeps the same tokens/style/speed -> audio signature as the original model. The UINT8 model is the ready-made model_uint8.onnx published by onnx-community/Kokoro-82M-ONNX; it was verified with the application's local CPU integration test.

Sources:

• hexgrad/Kokoro-82M
• nazdridoy/kokoro-tts model release

Kitten TTS Nano 0.8

File	Size	Purpose
kitten/kitten_tts_nano_v0_8.onnx	54.14 MiB	Kitten TTS ONNX model
kitten/voices.npz	3.13 MiB	Eight voice profiles
kitten/config.json	688 B	Voice aliases and speed priors

Source:

• KittenML/kitten-tts-nano-0.8

Piper Plus Japanese

File	Size	Purpose
piper-ja/model.onnx	37.39 MiB	Piper Plus つくよみちゃん VITS model
piper-ja/model.onnx.json	8.37 KiB	Japanese role, six-language IDs and phoneme inventory

Source:

• ayousanz/piper-plus-tsukuyomi-chan

The voice follows the Tsukuyomi-chan corpus terms and is not covered by the Apache-2.0 license used by this application's source code.

Usage

Download the files into the directory layout expected by the application:

models/
  kokoro/
    kokoro-v1.0-quantized.onnx
    kokoro-v0_19.onnx
    kokoro-v0_19-uint8.onnx
    kokoro-v0_19-int8.onnx
    voices-v1.0.bin
  kitten/
    kitten_tts_nano_v0_8.onnx
    voices.npz
    config.json
  piper-ja/
    model.onnx
    model.onnx.json

Kokoro is optional and is enabled only when its model paths are present in application.yml.

SHA-256

fbae9257e1e05ffc727e951ef9b9c98418e6d79f1c9b6b13bd59f5c9028a1478  kokoro/kokoro-v1.0-quantized.onnx
dece567789190ebe987bd245d95c09d5ac86de28ff0c325c2e3faaf3de04442c  kokoro/kokoro-v0_19.onnx
9748ecc9462b89bff04a7c7307d78e084f88ba4786083b36b083841599d38cea  kokoro/kokoro-v0_19-uint8.onnx
6dd7312add9cdddbcdaee461904ba75f69240c69404b3914369246a42af2ecbc  kokoro/kokoro-v0_19-int8.onnx
d19762d46cf0e6648cb28a7711df1637aad15818185d13f4ff840d57f2f6dfed  kokoro/voices-v1.0.bin
320564d2615f235de972ca27a7f39551c94185cfa24ca85b07a29084135f1e5e  kitten/kitten_tts_nano_v0_8.onnx
8aa7cee235abb0739cb51e6559685f65a4dacd95568833d05699b1633f519b3f  kitten/voices.npz
b66006ccbeccd4de5fc3c9272059c47f5725df7215fd889785c03602652fab64  kitten/config.json
7c8e909967370a01230ca173fd4541364f30aa3178597b6000122bec92597644  piper-ja/model.onnx

License

The Kokoro and Kitten model repositories declare the Apache-2.0 license. Piper Plus つくよみちゃん retains the separate voice and corpus conditions from its upstream model card. Model files are redistributed without modifying their learned weights, except for the optional INT8 quantization of Kokoro. Refer to the upstream model cards and licenses for details.