Usage - VOICEVOX Core for PHP - Knowledge of Laravel

Basic speech synthesis

The following talk.php demonstrates how to synthesize speech from text and write it to a WAV file.

PHP FFI is usually disabled in web server environments such as FPM. Run this script from the CLI with php talk.php.

<?php

require __DIR__ . '/vendor/autoload.php';

use Revolution\Voicevox\Core\Enums\AccelerationMode;
use Revolution\Voicevox\Core\Onnxruntime;
use Revolution\Voicevox\Core\OpenJtalk;
use Revolution\Voicevox\Core\Synthesizer;
use Revolution\Voicevox\Core\VoiceModelFile;

// Adjust the path to match your voicevox_core installation
$voicevoxCoreDir = getenv('HOME') . '/.local/voicevox_core';
// The recommended ONNX Runtime is 1.23.2. If it does not match your installed version, specify the filename directly
$onnxruntimeFilename = $voicevoxCoreDir . '/onnxruntime/lib/' . Onnxruntime::libRecommendedVersionedFilename();
$dictDir = $voicevoxCoreDir . '/dict/open_jtalk_dic_utf_8-1.11';
$vvmPath  = $voicevoxCoreDir . '/models/vvms/0.vvm';

// Text and style ID to synthesize
$text    = 'この音声は、ボイスボックスを使用して、出力されています。';
$styleId = 0;
$outPath = './output.wav';

// Initialize
$onnxruntime = Onnxruntime::loadOnce($onnxruntimeFilename);
$openJtalk   = new OpenJtalk($dictDir);
$synthesizer = new Synthesizer($onnxruntime, $openJtalk, AccelerationMode::Auto);

// Load the voice model
$model = VoiceModelFile::open($vvmPath);
$synthesizer->loadVoiceModel($model);

// Synthesize speech
$audioQuery = $synthesizer->createAudioQuery($text, $styleId);
$wav        = $synthesizer->synthesis($audioQuery, $styleId);

file_put_contents($outPath, $wav);
echo 'Wrote ' . $outPath . PHP_EOL;

Run it:

php talk.php

One-step synthesis with `tts`

If you want to skip the two-step createAudioQuery + synthesis flow, use the tts() method:

$wav = $synthesizer->tts($text, $styleId);
file_put_contents('./output.wav', $wav);

Synthesis from kana notation

You can also synthesize from AquesTalk-style kana notation:

// Using ttsFromKana
$wav = $synthesizer->ttsFromKana("コノオンセイワ'、ボイスボックスオ'/シヨーシテ'、シュツリョクサレテイマ'ス。", $styleId);
file_put_contents('./output.wav', $wav);

Using a user dictionary

Use UserDict to register custom word pronunciations:

use Revolution\Voicevox\Core\UserDict;
use Revolution\Voicevox\Core\Enums\UserDictWordType;

$userDict = new UserDict();
$uuid = $userDict->addWord(
    surface: 'ボイボ',
    pronunciation: 'ボイボ',
    accentType: 0,
    wordType: UserDictWordType::ProperNoun,
    priority: 10,
);

// Save the dictionary
$userDict->save('./mydict.json');

// Apply the dictionary to OpenJTalk
$openJtalk->useUserDict($userDict);

Adjusting AudioQuery

You can fine-tune the intonation and tempo by modifying the JSON returned by createAudioQuery before synthesis:

$audioQueryJson = $synthesizer->createAudioQuery($text, $styleId);

// Decode and adjust the JSON
$audioQuery = json_decode($audioQueryJson, true);
$audioQuery['speedScale'] = 1.2;   // Speed up by 1.2x
$audioQuery['pitchScale'] = 0.05;  // Raise pitch slightly

$wav = $synthesizer->synthesis(json_encode($audioQuery), $styleId);
file_put_contents('./output.wav', $wav);

Retrieving accent phrases

You can retrieve and inspect accent phrase information for a text:

$accentPhrasesJson = $synthesizer->createAccentPhrases($text, $styleId);
$accentPhrases = json_decode($accentPhrasesJson, true);

// Overwrite mora pitch and phoneme length with another style
$updatedJson = $synthesizer->replaceMoraData($accentPhrasesJson, $styleId);

GPU mode

In environments where a GPU is available, specify AccelerationMode::Gpu for faster synthesis:

use Revolution\Voicevox\Core\Enums\AccelerationMode;

$synthesizer = new Synthesizer($onnxruntime, $openJtalk, AccelerationMode::Gpu);

// Check whether GPU mode is active
if ($synthesizer->isGpuMode()) {
    echo 'Running in GPU mode' . PHP_EOL;
}

Using multiple voice models

Load multiple .vvm files and switch between style IDs to use different character voices:

$model0 = VoiceModelFile::open($voicevoxCoreDir . '/models/vvms/0.vvm');
$model1 = VoiceModelFile::open($voicevoxCoreDir . '/models/vvms/1.vvm');

$synthesizer->loadVoiceModel($model0);
$synthesizer->loadVoiceModel($model1);

// Inspect speaker metadata included in the models
$metas = json_decode($synthesizer->metas(), true);

// Unload a model when it is no longer needed
$synthesizer->unloadVoiceModel($model0->id());

To control what happens when a model with the same ID is already loaded, pass an OnExistingVoiceModelId as the second argument of loadVoiceModel(). The default Error throws an exception, Reload reloads the model and Skip does nothing:

use Revolution\Voicevox\Core\Enums\OnExistingVoiceModelId;

$synthesizer->loadVoiceModel($model0, OnExistingVoiceModelId::Skip);

Frame-based PCM rendering

Create an AudioFeature with createAudioFeature(), then slice() it by frame range and render() each slice to turn long text into audio piece by piece. This example writes PCM in chunks of about 0.3 seconds:

use Revolution\Voicevox\Core\VoicevoxCore;

$audioQuery   = $synthesizer->createAudioQuery($text, $styleId);
$audioFeature = $synthesizer->createAudioFeature($audioQuery, $styleId);

$frameRate        = (new VoicevoxCore)->getFrameRate();
$chunkFrameLength = max(1, (int) round(0.3 * $frameRate));
$totalFrames      = $audioFeature->length();

$stream = fopen('./talk.pcm', 'wb');

for ($start = 0; $start < $totalFrames; $start += $chunkFrameLength) {
    $end = min($start + $chunkFrameLength, $totalFrames);
    fwrite($stream, $audioFeature->slice($start, $end)->render());
}

fclose($stream);

render() returns raw 16-bit PCM, not a WAV file. This path requires a VVM v2 model. The C API exposes AudioFeature creation and range rendering, but not the higher-level streaming_synthesis API.

GitHub Copilot

SDK

AI SDK

VOICEVOX

Starter Kits

Feeds and RSS

Middleware

Usage - VOICEVOX Core for PHP

Basic speech synthesis

One-step synthesis with `tts`

Synthesis from kana notation

Using a user dictionary

Adjusting AudioQuery

Retrieving accent phrases

GPU mode

Using multiple voice models

Frame-based PCM rendering

​Basic speech synthesis

​One-step synthesis with tts

​Synthesis from kana notation

​Using a user dictionary

​Adjusting AudioQuery

​Retrieving accent phrases

​GPU mode

​Using multiple voice models

​Frame-based PCM rendering

Basic speech synthesis

One-step synthesis with `tts`

Synthesis from kana notation

Using a user dictionary

Adjusting AudioQuery

Retrieving accent phrases

GPU mode

Using multiple voice models

Frame-based PCM rendering