SpeechifyAI provider for the Vercel AI SDK

The SpeechifyAI provider for the Vercel AI SDK adds SpeechifyAI's text-to-speech models to the AI SDK's unified speech interface — switch from OpenAI, ElevenLabs, or Deepgram to SpeechifyAI with a one-line model swap. It is a thin bridge over the official @speechify/api client.

Setup

npm install @speechify/vercel ai

Get an API key from the SpeechifyAI Platform and set it as SPEECHIFY_API_KEY.

Usage

import { speechify } from '@speechify/vercel';
import { generateSpeech } from 'ai';

const { audio } = await generateSpeech({
  model: speechify.speech(), // simba-3.2, voice "geffen_32"
  text: 'Hello from Speechify!',
});

To customize the API key, base URL, or headers, create your own provider instance:

import { createSpeechify } from '@speechify/vercel';

const speechify = createSpeechify({
  apiKey: process.env.MY_SPEECHIFY_KEY,
});

Models

| Model ID | Description | | --- | --- | | simba-3.2 | Latest-generation streaming-native model, lower latency and richer expressivity (default). English only; serves a curated voice set. | | simba-english | English-optimized TTS | | simba-multilingual | Multilingual TTS — use this for non-English input | | simba-3.0 | Earlier Simba 3 model, still available |

Voices

The voice option takes a SpeechifyAI voice ID, including your own cloned voices. Defaults to geffen_32 when omitted. The simba-3.2 curated set is currently: beatrice_32, dominic_32, edmund_32, geffen_32, harper_32, hugh_32, imogen_32, wyatt_32; voices like george and scott work with the 1.6 models. List available voices with GET /v1/voices (docs) — each voice's models field shows which models it supports; simba-3.2 serves a curated subset of shared voices (cloned voices are not in it).

Output formats

The standard outputFormat option accepts either a simple format name — mp3, wav, ogg, aac, pcm — or a SpeechifyAI codec string with sample rate and bitrate, e.g. mp3_24000_128, pcm_16000, ulaw_8000 (useful for telephony). Defaults to mp3 when omitted.

Speed, emotion, and SSML

SpeechifyAI controls prosody through SSML. The standard speed option is implemented by wrapping your text in <speak><prosody rate="...">:

await generateSpeech({
  model: speechify.speech(),
  text: 'A bit faster please.',
  speed: 1.25,
});

If your text is already SSML (starts with <speak), it is sent unchanged — set rate, pitch, and emotion directly in your markup. The instructions option is not supported and produces a warning.

Provider options

await generateSpeech({
  model: speechify.speech(),
  text: 'Hello!',
  providerOptions: {
    speechify: {
      ssml: true,                    // treat text as SSML, disable speed wrapping
      outputFormat: 'mp3_24000_128', // codec string, overrides outputFormat
      loudnessNormalization: true,
      textNormalization: true,
    },
  },
});

Speech marks (word timing)

SpeechifyAI returns word- and sentence-level timing data with every generation, available via provider metadata:

const result = await generateSpeech({
  model: speechify.speech(),
  text: 'Timed speech.',
});

const { speechMarks, billableCharactersCount } =
  result.providerMetadata.speechify;

Limitations

• Text-to-speech only — SpeechifyAI has no transcription API, so there is no transcription model.
• The AI SDK speech interface is request/response; SpeechifyAI's streaming endpoint (/v1/audio/stream) is not yet exposed through this provider.