Text to speech

Pendra turns text into natural-sounding speech — voice announcements, audio versions of written content, spoken responses for assistants and IVR systems. It's OpenAI-compatible, so existing text-to-speech code works against Pendra by swapping the base URL and key.

Generate speech

Send a model and the input text (up to 4,096 characters) and you get back a WAV file, ready to save or play. Pick a voice from the model's preset voices, or leave it out to use the model's default. The Pendra SDKs (0.17 and later) also tell you which worker generated it, as speech.pendra.worker; the OpenAI SDKs work too, since the endpoint is OpenAI-compatible.

from pendra import Pendra

client = Pendra()

speech = client.audio.speech.create(
    model="your-speech-model",  # an id from /models?type=speech
    input="Good morning, and welcome to today's briefing.",
)
speech.write_to_file("speech.wav")

import Pendra from 'pendra';
import { writeFileSync } from 'node:fs';

const client = new Pendra();

const speech = await client.audio.speech.create({
  model: 'your-speech-model', // an id from /models?type=speech
  input: "Good morning, and welcome to today's briefing.",
});
writeFileSync('speech.wav', speech.audio);

Voices

Each speech model ships with a set of preset voices. The /models?type=speech entry for a model lists its voices and the default_voice used when a request omits one — so you can present the choices in your product and default sensibly. Passing a voice the model doesn't have returns a clear error.

Output format

Speech is returned as a WAV file (24 kHz, 16-bit mono) — playable everywhere and easy to convert if you need something smaller for delivery (e.g. ffmpeg -i speech.wav speech.mp3).

Finding speech models

The speech models available to you are the ones your connected workers serve — list them, with their voices, at /models?type=speech, and pass one of their ids as model.