How to use kRouter as a unified text-to-speech API

Most developers think of AI routers as just "the thing that switches between GPT and Claude." But the fragmentation in the AI space is not limited to text generation.

If you want your application to speak, you are facing a mess of incompatible APIs. OpenAI has one format. Google Gemini TTS has another. NVIDIA NIM expects its own schema. MiniMax has an entirely different authentication and payload structure. Every provider means a new SDK, a new set of voice enums, and a new billing dashboard.

kRouter solves this by standardizing everything into the OpenAI /v1/audio/speech format. One endpoint, one SDK, four providers.

The 4 TTS Providers

kRouter currently supports text-to-speech from four providers, all accessible through the same OpenAI-compatible endpoint:

The Standardized Endpoint

Once you configure your providers in kRouter, you use the exact same curl command (or OpenAI SDK call) to generate audio from any of them:

curl http://localhost:20128/v1/audio/speech \
  -H "Authorization: Bearer sk-krouter-local" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax/minimax-tts",
    "input": "This is routing through kRouter",
    "voice": "alloy"
  }' --output output.mp3

Change the model field to any of the four providers. kRouter intercepts the payload, translates the voice enum and input parameters into the target provider's specific JSON schema, sends the request, and streams the binary audio back to you.

Voice Mapping

Every provider has its own voice naming scheme. kRouter maps the standard OpenAI voice names (alloy, echo, fable, onyx, nova, shimmer) to each provider's equivalent. You write "voice": "alloy" and kRouter figures out the correct enum for Gemini, NVIDIA, or MiniMax behind the scenes.

If you need a provider-specific voice that has no OpenAI equivalent, you can pass the raw voice ID prefixed with the provider namespace (e.g., "voice": "nvidia/custom-voice-id"). kRouter will pass it through without mapping.

Code Examples

OpenAI SDK (Python)

from openai import OpenAI
 
client = OpenAI(
    base_url="http://localhost:20128/v1",
    api_key="sk-krouter-local"
)
 
# Use Gemini's free tier TTS
response = client.audio.speech.create(
    model="gemini/gemini-2.5-flash-preview-tts",
    voice="nova",
    input="Welcome to the kRouter podcast."
)
response.stream_to_file("intro.mp3")

Node.js

const response = await fetch("http://localhost:20128/v1/audio/speech", {
  method: "POST",
  headers: {
    "Authorization": "Bearer sk-krouter-local",
    "Content-Type": "application/json"
  },
  body: JSON.stringify({
    model: "openai/tts-1-hd",
    voice: "shimmer",
    input: "This is production-quality narration."
  })
});
const buffer = await response.arrayBuffer();
fs.writeFileSync("narration.mp3", Buffer.from(buffer));

Beyond TTS: STT and Image Generation

kRouter does not stop at text-to-speech. The same unified approach applies to two other modality endpoints:

Speech-to-Text via /v1/audio/transcriptions -- upload an audio file and get a transcript back. kRouter translates between OpenAI Whisper format and any configured STT provider.

Image Generation via /v1/images/generations -- generate images using DALL-E, Gemini Imagen, or other configured providers, all through the standard OpenAI images endpoint.

Each of these endpoints has a corresponding Agent Skill that teaches AI agents how to call them automatically. Feed the skill URL to Claude Code, and your agent can generate audio, transcribe meetings, or create images without you writing a single line of API integration code.

Why This Matters

1. No vendor lock-in. Rip out provider-specific TTS SDKs from your codebase. Use the standard OpenAI SDK for everything.
2. Cost arbitrage. If OpenAI's TTS gets expensive, instantly fall back to Gemini's free tier or MiniMax's ultra-cheap voices by changing one string in your combo configuration.
3. 3-tier auto-fallback. Set up a combo like openai/tts-1-hd → gemini/tts → minimax/tts. If your OpenAI quota runs dry, kRouter's Zenith engine seamlessly routes to the next provider. Your users never hear silence.
4. Agentic workflows. Give a terminal AI agent the krouter-tts Skill URL and it instantly knows how to generate audio files on your local machine using whatever providers you have configured.

Get Started

If you are already running kRouter for coding, open the dashboard at http://localhost:20128/dashboard and check the Skills tab to see all the endpoints you did not know you had.

If you are new to kRouter, install it and configure your first TTS provider in under 2 minutes:

npm install -g @sifxprime/krouter

See the full install guide and changelog for the latest provider additions.