Edge TTS for PHP

Edge TTS turns text into natural-sounding speech with the online text-to-speech service of Microsoft Edge, without Edge, Windows or an API key.

Available in three languages

The same library for the Edge text-to-speech service, in Node.js, PHP and Go:

Language	Install	Repository
Node.js and browsers	npm install @andresaya/edge-tts	andresayac/edge-tts
PHP (this one)	composer require afaya/edge-tts	andresayac/edge-tts-php
Go	go get github.com/andresayac/edge-tts-go	andresayac/edge-tts-go

Features

• Text-to-speech with more than 300 neural voices in over 100 locales.
• Streaming: get the audio as it arrives, with a foreach or a callback.
• Word and sentence boundaries: know when each word and sentence is spoken.
• Subtitles: SRT or WebVTT, one cue per sentence.
• Voice filtering by language and gender, with the list cached for 10 minutes.
• Pitch, rate and volume control.
• Configurable: override the endpoint, headers and timeout without waiting for a release.
• No event loop: it works inside Laravel, Symfony, WordPress or a plain script.
• Verified TLS, on Windows too, without configuring a CA bundle.
• Command-line tool included.

Requirements

• PHP 7.4 or later (tested on 7.4, 8.1, 8.2, 8.3 and 8.4)
• The json and openssl extensions

Installation

composer require afaya/edge-tts

Quick start

<?php

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

use Afaya\EdgeTTS\Service\EdgeTTS;

$tts = new EdgeTTS();
$tts->synthesize('Hello, world!', 'en-US-AriaNeural');

echo $tts->toFile('hello'); // hello.mp3, the extension comes from the format

Command-line interface

# Synthesize text to hello.mp3
php vendor/bin/edge-tts synthesize --text "Hello, world!" -o hello

# Read the text from a file, with another voice and prosody, and write subtitles too
php vendor/bin/edge-tts synthesize -f story.txt --voice es-CO-SalomeNeural -r +10% -p -5Hz -o story --subtitles story.srt

# Read from stdin, WebM/Opus output, WebVTT subtitles
cat story.txt | php vendor/bin/edge-tts synthesize --file=- --format webm-24khz-16bit-mono-opus -o story --subtitles story.vtt

# List the voices, optionally filtered
php vendor/bin/edge-tts voice-list
php vendor/bin/edge-tts voice-list --locale es-CO --gender female

Option	Description	Default
-t, --text	Text to convert to audio
-f, --file	Input file (text); --file=- reads from stdin
--voice	Voice to use	en-US-AriaNeural
-r, --rate	Rate of speech	0%
-l, --volume	Volume of speech	0%
-p, --pitch	Pitch of speech	0Hz
-o, --output	Output file name; the extension is added when missing	output_<timestamp>
--format	Output format	audio-24khz-48kbitrate-mono-mp3
--subtitles	Write subtitles: .srt, or .vtt for WebVTT

edge-tts:synthesize and edge-tts:voice-list still work as names of the commands.

API

Configuration

Every option is optional. The defaults come from Constants, so when Microsoft changes the service you can set the new values yourself instead of waiting for a release.

$tts = new EdgeTTS([
    'wssUrl'          => 'wss://api.msedgeservices.com/tts/cognitiveservices/websocket/v1',
    'voicesUrl'       => 'https://speech.platform.bing.com/consumer/speech/synthesize/readaloud/voices/list',
    'secMsGecVersion' => '1-155.0.4283.45',
    'headers'         => ['User-Agent' => '...'], // added to the default headers
    'timeout'         => 30,                      // seconds without data before giving up
    'voicesCacheTtl'  => 600,                     // seconds to reuse the voice list, 0 to always fetch it
]);

A misspelled option throws InvalidArgumentException.

Voices

$voices   = $tts->getVoices();                  // every voice
$english  = $tts->getVoicesByLanguage('en');    // en-US, en-GB, en-IN...
$american = $tts->getVoicesByLanguage('en-US');
$female   = $tts->getVoicesByGender('Female');
$voice    = $tts->checkVoice('en-US-AriaNeural'); // throws InvalidArgumentException when it does not exist

foreach ($voices as $voice) {
    echo $voice['ShortName'] . ' ' . $voice['FriendlyName'] . PHP_EOL; // en-US-AriaNeural Aria (Neural) - en-US
}

Each voice has Name, OriginalName, ShortName, DisplayName, LocalName, Gender, Locale, LocaleName, SecondaryLocaleList, VoiceType, VoiceTag and FriendlyName.

Synthesis

use Afaya\EdgeTTS\Config\Constants;

$tts->synthesize('Hello, world!', 'en-US-AriaNeural', [
    'rate'         => '+50%',  // -100% and up
    'volume'       => '-10%',  // -100% to +100%
    'pitch'        => '+20Hz', // -100Hz to +100Hz
    'outputFormat' => Constants::OUTPUT_FORMAT['AUDIO_24KHZ_96KBITRATE_MONO_MP3'],
    'timeout'      => 60,      // seconds, for this call only
]);

Every option is optional, and numbers work too: 'rate' => 50 is +50% and 'pitch' => -5 is -5Hz.

Output formats

Constants::OUTPUT_FORMAT key	Format	Extension
AUDIO_24KHZ_48KBITRATE_MONO_MP3 (default)	audio-24khz-48kbitrate-mono-mp3	.mp3
AUDIO_24KHZ_96KBITRATE_MONO_MP3	audio-24khz-96kbitrate-mono-mp3	.mp3
WEBM_24KHZ_16BIT_MONO_OPUS	webm-24khz-16bit-mono-opus	.webm

These are the only formats the service accepts. It refuses the other Azure formats (WAV, PCM, OGG, 48 kHz MP3...) with "Unsupported Edge output format".

Matching the voice to the text

Voices are locale specific. A voice can only read text written in the script of its own language, so passing Chinese, Cyrillic, Arabic or Devanagari text to an English voice returns no audio. synthesize() then throws EmptyAudioException:

use Afaya\EdgeTTS\Exception\EmptyAudioException;

try {
    $tts->synthesize('这是一个测试', 'en-US-AriaNeural');
} catch (EmptyAudioException $e) {
    // Use a Chinese voice instead, such as zh-CN-XiaoxiaoNeural.
}

Streaming

stream() yields the audio as it arrives, so you can start playing or sending it before the synthesis ends:

header('Content-Type: audio/mpeg');

foreach ($tts->stream('Long text to stream...', 'en-US-AriaNeural') as $chunk) {
    echo $chunk;
    flush();
}

Leaving the loop early closes the connection. If you prefer a callback:

$tts->synthesizeStream('Long text to stream...', 'en-US-AriaNeural', [], function (string $chunk) {
    file_put_contents('out.mp3', $chunk, FILE_APPEND);
});

Everything in one result

generate() returns the audio, the boundaries and the duration in a SynthesisResult, and leaves the instance as it was:

$result = $tts->generate('¿Cómo estás? ¡Qué bien!', 'es-CO-SalomeNeural');

$result->audio;              // binary string
$result->format;             // 'mp3'
$result->outputFormat;       // 'audio-24khz-48kbitrate-mono-mp3'
$result->wordBoundaries;
$result->sentenceBoundaries;
$result->duration;           // estimated, in seconds

$result->toFile('greeting'); // greeting.mp3
$result->toSRT();            // or toVTT()

Getting the audio

After synthesize(), stream() or synthesizeStream():

$raw    = $tts->toRaw();            // binary string
$base64 = $tts->toBase64();         // base64 string
$path   = $tts->toFile('out/audio'); // out/audio.mp3
$path   = $tts->toFile('out/audio.mp3'); // out/audio.mp3, the extension is not repeated
$path   = $tts->toFile('out/audio', 'bin'); // out/audio.bin
$stream = $tts->toStream();         // php://memory resource

$info = $tts->getAudioInfo();       // ['size' => 46944, 'format' => 'mp3', 'estimatedDuration' => 3.912]
$seconds = $tts->getDuration();     // estimated from the size and bitrate
$bytes = $tts->getSizeBytes();

Word and sentence boundaries

foreach ($tts->getWordBoundaries() as $word) {
    // Offsets and durations are in 100-nanosecond units.
    printf("%.2fs %s\n", $word['offset'] / 1e7, $word['text']);
}

$tts->getSentenceBoundaries(); // the same, one per sentence
$tts->saveMetadata('words.json');

The service leaves the opening punctuation out of its sentences, so "¿Cómo estás?" comes back as "Cómo estás?". The library puts it back from your text: ¿ ¡ « " ' “ ‘ ( [.

Subtitles

file_put_contents('speech.srt', $tts->toSRT()); // one cue per sentence
file_put_contents('speech.vtt', $tts->toVTT());

use Afaya\EdgeTTS\Service\Subtitles;

Subtitles::toSRT($tts->getWordBoundaries());    // one cue per word

1
00:00:00,050 --> 00:00:01,600
¿Cómo estás?

2
00:00:01,600 --> 00:00:02,950
¡Qué bien!

Errors

Exception	When
InvalidArgumentException	An empty text, an invalid pitch, rate or volume, or checkVoice() with an unknown voice
Afaya\EdgeTTS\Exception\EmptyAudioException	The voice cannot read the text
Afaya\EdgeTTS\Exception\HandshakeException	The service refused the connection; getStatusCode() has the HTTP status
Afaya\EdgeTTS\Exception\EdgeTTSException	Any other error of the service or the network, such as "The service refused the request: Unsupported voice xx-XX-NobodyNeural."
RuntimeException	Asking for audio or metadata before a synthesis

Every exception of the service extends EdgeTTSException, which extends RuntimeException.

A machine whose clock is off by more than a few minutes is refused by the service. The client then corrects its clock from the response and tries once more.

Upgrading

Version 1.7 stopped working when Microsoft moved the service to a new endpoint; 1.8.3 and later use the new one. What changed in each version, and what to look at when upgrading, is in the CHANGELOG.

Examples

• examples/basic.php: synthesize to a file and list voices.
• examples/stream.php: stream the audio with a generator and with a callback.
• examples/subtitles.php: generate the audio and its subtitles in one call.

Testing

composer install
composer test

The tests replay a recorded session of the service, through a fake and through a local server on real sockets, so they need no network.

License

GPL-3.0-or-later. See LICENSE.txt.

Acknowledgments

• https://github.com/rany2/edge-tts
• https://github.com/hasscc/hass-edge-tts