Audio

Two directions, one page: Audio::of() turns text into speech, and Transcription::fromPath() (or storage, upload) turns audio back into text. OpenAI and ElevenLabs cover both; Gemini, Mistral, and Groq cover transcription alongside their other capabilities.

Speech

Pick a voice (male(), female(), or voice('name')), optionally coach delivery with instructions(), then generate. The response stringifies to its audio bytes and stores to the configured filesystem:

use Crustum\Ai\Audio;

$audio = Audio::of('I love coding with CakePHP.')->generate();
echo $audio->mimeType() . PHP_EOL;
echo strlen((string)$audio) . " bytes\n";

audio/mp3
56 bytes

store(), storeAs('audio.mp3'), storePublicly(), and storePubliclyAs() persist the bytes; timeout() bounds the request. Generation queues like everything else. queue()->then(fn ($audio) => $audio->store()) runs the store after the worker finishes.

Transcription

Any transcribable audio file transcribes with generate(); diarize() adds per-speaker segments alongside the raw text:

use Crustum\Ai\Transcription;

$transcript = Transcription::fromPath('/var/www/audio.mp3')->generate();
echo (string)$transcript . PHP_EOL;

harness-transcript

Uploads and stored files work the same way (fromUpload($request->getData(...)), fromStorage('audio.mp3')). Transcription queues with a TranscriptionResponse in the then() callback, and takes provider options and headers per request like the other builders.

See also

• Providers — capability matrix
• ElevenLabs — voices and Scribe
• OpenAI — the transcription default