Speech and transcription

The package adds Workers AI to Laravel AI’s Audio and Transcription APIs. Both need no extra config, and work on the direct route and the gateway route. The exception is nova-3.

Generate speech

use Laravel\Ai\Audio;

$audio = Audio::of('Your order has shipped.')->generate(provider: 'workers-ai');

$path = $audio->store('audio', 'public');
$bytes = (string) $audio;
$mime = $audio->mime; // audio/mpeg

The default model is @cf/deepgram/aura-2-en. To change it, set models.audio.default on the provider entry in config/ai.php, or pass model: to generate().

Choose a voice with voice(). male() and female() pick orion and luna, or nestor and carina on the Spanish model.

Audio::of('Hello.')->voice('asteria')->generate(provider: 'workers-ai');

Audio::of('Hello.')
    ->withProviderOptions(['lang' => 'fr'])
    ->generate(provider: 'workers-ai', model: '@cf/myshell-ai/melotts');

withProviderOptions() needs laravel/ai 1.0 or later. On Aura models, it also accepts encoding, container, sample_rate and bit_rate. These models have no input for instructions(), so that method has no effect here.

Transcribe audio

use Laravel\Ai\Transcription;

$transcript = Transcription::fromPath(storage_path('app/call.mp3'))
    ->language('en')
    ->generate(provider: 'workers-ai');

$transcript->text;

foreach ($transcript->segments as $segment) {
    $segment->text;
    $segment->startSeconds;
    $segment->endSeconds;
}

Transcription::fromStorage(), fromUpload() and fromBase64() also work.

The default model is @cf/openai/whisper-large-v3-turbo. To change it, set models.transcription.default on the provider entry in config/ai.php, or pass model: to generate().

Accuracy varies, so test your own audio. In tests on 2026-11-05 with synthetic speech, whisper-tiny-en heard “invoice” as “inblice”, and whisper-large-v3-turbo once heard it as “in-place”.

nova-3 cannot run through AI Gateway, so it throws when a gateway is set. Use a provider entry with no gateway for that call. The Whisper models work through a gateway, with key alone. See Non-JSON models.

Label speakers

Only nova-3 labels speakers. diarize() on another model throws before a request. A speaker-labeled call therefore cannot go through a gateway. Speakers are named speaker_0, speaker_1 and so on.

$transcript = Transcription::fromPath(storage_path('app/call.mp3'))
    ->diarize()
    ->generate(provider: 'workers-ai', model: '@cf/deepgram/nova-3');

foreach ($transcript->segments as $segment) {
    echo "{$segment->speaker}: {$segment->text}\n";
}

Long audio

Laravel AI sends a 30-second timeout unless you set one. For long audio, set a longer one:

Transcription::fromPath($path)->timeout(300)->generate(provider: 'workers-ai');

Omit the provider name

Laravel AI’s defaults are openai until you change them. Set them in config/ai.php to omit provider::

'default_for_audio' => 'workers-ai',
'default_for_transcription' => 'workers-ai',

Test your code

use Laravel\Ai\Audio;
use Laravel\Ai\Transcription;

Audio::fake();
Transcription::fake(['Hello']);

// run your code, then:
Audio::assertGenerated(fn ($prompt) => $prompt->contains('shipped'));
Transcription::assertGenerated(fn ($prompt) => true);

Cost

Cloudflare bills speech and transcription in neurons. The package reports them with WorkersAiUsageReported. See Cost tracking.