Speech to Text and Text to Speech Quickstart

Text to speech and speech to text are two halves of the same workflow: one turns text into spoken audio, the other turns audio back into text. They feed into each other, so this quickstart runs both as a round-trip. You generate speech from a sentence, then transcribe that audio and check the text comes back.

1. Set up your project

   Create a new directory and initialize a Node.js project:

   Terminal

   mkdir ai-speech-demo
   cd ai-speech-demo
   pnpm init

2. Install dependencies

   Install the canary AI SDK, the AI Gateway provider, and development dependencies:

   npmyarnpnpmbun

   Terminal

   npm install ai@canary @ai-sdk/gateway@canary dotenv @types/node tsx typescript

   Terminal

   yarn add ai@canary @ai-sdk/gateway@canary dotenv @types/node tsx typescript

   Terminal

   pnpm add ai@canary @ai-sdk/gateway@canary dotenv @types/node tsx typescript

   Terminal

   bun add ai@canary @ai-sdk/gateway@canary dotenv @types/node tsx typescript

3. Set up your API key

   Go to the AI Gateway API Keys page in your Vercel dashboard and click Create key to generate a new API key.

   Create a .env.local file and save your API key:

   .env.local

   AI_GATEWAY_API_KEY=your_ai_gateway_api_key

   Instead of using an API key, you can use OIDC tokens to authenticate your requests.

4. Run the round-trip

   Create an index.ts file. It generates speech from a sentence, saves the audio, then transcribes that same audio back to text:

   index.ts

   import {
     experimental_generateSpeech as generateSpeech,
     experimental_transcribe as transcribe,
   } from 'ai';
   import { gateway } from '@ai-sdk/gateway';
   import { writeFile } from 'node:fs/promises';
   import 'dotenv/config';
    
   async function main() {
     const text = 'Thanks for trying out AI Gateway.';
    
     // Text to speech
     const speech = await generateSpeech({
       model: gateway.speechModel('openai/tts-1'),
       text,
       voice: 'alloy',
       outputFormat: 'mp3',
     });
     await writeFile('speech.mp3', speech.audio.uint8Array);
     console.log('Saved speech.mp3');
    
     // Speech to text: transcribe the audio we just generated
     const transcript = await transcribe({
       model: gateway.transcriptionModel('openai/whisper-1'),
       audio: speech.audio.uint8Array,
     });
     console.log('Transcript:', transcript.text);
   }
    
   main().catch(console.error);

   Run your script:

   Terminal

   pnpm tsx index.ts

   You get speech.mp3 with the spoken sentence, and the transcript prints back the text you started with.

5. Next steps

   • Read the Text to Speech reference for voices, formats, and the full list of request options
   • Read the Speech to Text reference for segments, timestamps, and provider options
   • For live, two-way voice conversations, follow the Realtime quickstart

The two calls are independent. Use experimental_generateSpeech alone to add voiceovers or spoken responses, and experimental_transcribe alone to transcribe recordings, voice notes, or call audio. You can also call the REST endpoints directly without the AI SDK; see the Text to Speech and Speech to Text references.