voicevox

README

voicevox

English | 日本語

Go bindings for VOICEVOX core, for Japanese text to speech that runs locally.

• Uses purego and ffi, so CGo is not needed.
• Targets voicevox_core 0.17.0.

Installation

libffi

On Linux, install libffi.

sudo apt install libffi8

On macOS and Windows x64, the ffi package ships libffi, so you don't need to install it.

VOICEVOX runtime

Get the official downloader from the voicevox_core releases, then run it:

curl -sSfL https://github.com/VOICEVOX/voicevox_core/releases/download/0.17.0/download-linux-x64 -o download
chmod +x download
./download -o ~/voicevox_core

The downloader asks you to accept the VOICEVOX voice model terms and the VOICEVOX ONNX Runtime terms. To get only the talk models, add --models-pattern '[0-9]*.vvm'.

This creates the following layout:

voicevox_core/
  c_api/lib/libvoicevox_core.so
  onnxruntime/lib/libvoicevox_onnxruntime.so.<version>
  dict/open_jtalk_dic_utf_8-1.11/
  models/vvms/*.vvm

LoadOnnxruntime uses the recommended ONNX Runtime version if it is in the directory. Otherwise it uses the newest version it finds there.

Example

package main

import (
	"os"

	"github.com/hybridgroup/voicevox/pkg/voicevox"
)

func main() {
	voicevox.Load("voicevox_core/c_api/lib")

	ort, _ := voicevox.LoadOnnxruntime("voicevox_core/onnxruntime/lib")
	ojt, _ := voicevox.NewOpenJtalk("voicevox_core/dict/open_jtalk_dic_utf_8-1.11")

	syn, _ := voicevox.NewSynthesizer(ort, ojt, voicevox.DefaultInitializeOptions())
	ojt.Delete()
	defer syn.Delete()

	model, _ := voicevox.OpenVoiceModelFile("voicevox_core/models/vvms/0.vvm")
	syn.LoadVoiceModel(model, voicevox.DefaultLoadVoiceModelOptions())
	model.Delete()

	wav, _ := syn.TTS("こんにちは", 0, voicevox.DefaultTTSOptions())
	os.WriteFile("out.wav", wav, 0644)
}

The output is a 24 kHz, 16-bit, mono WAV file.

To change speed, pitch, intonation or volume:

1. Call CreateAudioQuery.
2. Edit speedScale, pitchScale, intonationScale or volumeScale in the returned JSON.
3. Pass the JSON to Synthesis.

A full example is in examples/hello:

go run ./examples/hello -data ~/voicevox_core -style 0 -text "こんにちは"

GPU

To use CUDA, download the CUDA build of ONNX Runtime and the CUDA libraries into the same directory:

./download -o ~/voicevox_core --devices cuda --only onnxruntime additional-libraries

The ONNX Runtime CUDA provider needs to find the CUDA and cuDNN libraries in additional_libraries. Call LoadAdditionalLibraries before LoadOnnxruntime, then set the acceleration mode when you create the synthesizer:

voicevox.LoadAdditionalLibraries("voicevox_core/additional_libraries")
ort, _ := voicevox.LoadOnnxruntime("voicevox_core/onnxruntime/lib")

opts := voicevox.DefaultInitializeOptions()
opts.AccelerationMode = voicevox.AccelerationModeGPU
syn, err := voicevox.NewSynthesizer(ort, ojt, opts)

NewSynthesizer returns an error if the GPU can't be used. Call syn.IsGPUMode() to see which mode is active, and ort.SupportedDevices() to see which devices the ONNX Runtime build supports.

go run ./examples/hello -data ~/voicevox_core -gpu

Tests

VOICEVOX_LIB=~/voicevox_core/c_api/lib VOICEVOX_DIR=~/voicevox_core go test ./...

VOICEVOX_LIB is required. Tests that need voice models are skipped when VOICEVOX_DIR is not set. To also run the GPU test, set VOICEVOX_TEST_GPU=1.

Credits

The VOICEVOX voice model terms require a credit for each character you use, for example "VOICEVOX:ずんだもん". Check the terms for each character.

License

MIT. VOICEVOX core, its voice models and its ONNX Runtime each have their own licenses and terms.