OpenRouter Text-to-Speech - Complete Documentation

OpenRouter supports text-to-speech (TTS) via a dedicated /api/v1/audio/speech endpoint that is compatible with the OpenAI Audio Speech API. Send text and receive a raw audio byte stream in your chosen format.

Model Discovery

You can find TTS models in several ways:

Via the API

Use the output_modalities query parameter on the Models API to discover TTS models:

# List only TTS models
curl "https://openrouter.ai/api/v1/models?output_modalities=speech"

On the Models Page

Visit the Models page and filter by output modalities to find models capable of speech synthesis. Look for models that list "speech" in their output modalities.

API Usage

Send a POST request to /api/v1/audio/speech with the text you want to synthesize. The response is a raw audio byte stream (not JSON), so you can pipe it directly to a file or audio player.

Basic Example

Request Parameters

Parameter	Type	Required	Description
`model`	string	Yes	The TTS model to use (e.g., `openai/gpt-4o-mini-tts-2025-12-15`, `mistralai/voxtral-mini-tts-2603`)
`input`	string	Yes	The text to synthesize into speech. Some models treat it as a prompt that can also describe non-speech audio; see Non-Speech Prompts below
`voice`	string	Provider-dependent	Voice identifier. Available voices vary by model, so check each model’s page on the Models page for supported voices. Omit this parameter only when the selected provider documents a default voice; otherwise an explicit voice is required.
`response_format`	string	No	Audio output format: `mp3` or `pcm`. Defaults to `pcm`
`speed`	number	No	Playback speed multiplier. Honored by models that support it (e.g., OpenAI TTS). Other providers either ignore it or return a 400 for a non-default value when the model has no speed control. Defaults to `1.0`
`input_references`	array	No	Reference content for stateless voice cloning or voice design: one to three `input_audio` parts (each optionally paired with a `text` transcript), or exactly one `image_url` part. See Voice Cloning below
`provider`	object	No	Provider-specific passthrough configuration

When voice is omitted, OpenRouter only forwards the request to providers whose adapter supports a provider-side default voice. For other providers, the request is rejected with a validation error.

Voice Cloning

Some models support stateless voice cloning: you send a short sample of reference audio directly with the TTS request, and the generated speech mimics that voice. No separate voice-creation or upload step is required. Pass the reference audio as a base64 input_audio part in input_references (a data:audio/...;base64, URI also works), and optionally include its transcript as a text part:

{
  "model": "fish-audio/s2.1-pro",
  "input": "Hello from my cloned voice!",
  "response_format": "mp3",
  "input_references": [
    { "type": "input_audio", "input_audio": { "data": "data:audio/wav;base64,UklGRuQXDAB..." } },
    { "type": "text", "text": "This is the transcript of the reference audio." }
  ]
}

Instead of inline base64, input_audio can carry a public url. OpenRouter downloads the file and forwards the bytes to the provider; the provider never receives your URL:

{
  "type": "input_audio",
  "input_audio": { "url": "https://example.com/samples/narrator.wav" }
}

Not every provider of a voice-cloning model supports voice cloning, and only some accept more than one clip or an image reference. Requests that carry references are only routed to endpoints whose capability flags allow them. Check these fields on the endpoints API:

Field	Meaning
`supports_voice_cloning`	The endpoint accepts `input_audio` reference clips
`supports_multiple_audio_references`	The endpoint accepts more than one `input_audio` clip per request
`supports_image_reference`	The endpoint accepts an `image_url` reference

Multiple Reference Clips

Models such as Seed Audio 1.0 accept up to three input_audio clips in one request, so a single generation can switch between voices. Clips are numbered in the order you send them, and you address them from input with the placeholders @Audio1, @Audio2, and @Audio3. If you also send transcripts with more than one clip, place each clip’s text part immediately after the clip it describes:

{
  "model": "bytedance-seed/seed-audio-1-0",
  "input": "@Audio1 Welcome back to the show. @Audio2 Thanks for having me, it is great to be here.",
  "response_format": "mp3",
  "input_references": [
    { "type": "input_audio", "input_audio": { "data": "data:audio/wav;base64,UklGRuQXDAB..." } },
    { "type": "input_audio", "input_audio": { "url": "https://example.com/samples/guest.mp3" } }
  ]
}

Placeholder rules, as enforced by Seed Audio 1.0 (currently the only model that accepts more than one clip; a model that accepts a single clip takes input verbatim and does not validate placeholders):

With a single clip, the placeholder is optional: the whole input is spoken in that voice. If you include one, it must be @Audio1.
With more than one clip, every clip must be referenced at least once. A request that supplies two clips but never mentions @Audio2 is rejected with a 400.
A placeholder that points at a clip you did not send (for example @Audio3 with two clips) is rejected with a 400, as is a placeholder in a request with no audio references.
Transcripts are optional and provider-dependent: Fish Audio sends each clip’s transcript to the model, while Seed Audio 1.0 ignores them. With a single clip the text part may appear before or after it; with multiple clips a text part must immediately follow its clip, and a transcript with no preceding clip is rejected.

Image References

Some models can design a voice from a picture instead of an audio sample. Send exactly one image_url part, either as a base64 data URI or as a public http(s) URL that OpenRouter downloads and forwards as bytes. JPEG, PNG, and WebP are accepted:

{
  "model": "bytedance-seed/seed-audio-1-0",
  "input": "Hi there, I will be your guide through the museum today.",
  "response_format": "mp3",
  "input_references": [
    { "type": "image_url", "image_url": { "url": "https://example.com/portraits/guide.png" } }
  ]
}

Audio and image references cannot be combined in one request, and @AudioN placeholders are rejected when no audio clip is present.

Limits and Requirements

Supported audio formats for reference clips are provider-specific.
input_references accepts one to three input_audio parts (each optionally paired with a text transcript) or exactly one image_url part, never both kinds in the same request. An empty array is treated as no reference.
Each inline reference is limited to 20 MiB of base64 (15 MiB of decoded media); a larger data value is rejected with a 400. Remote downloads are capped at the same 15 MiB and a larger file is rejected with a 413.
Remote url values must be public http(s) URLs of at most 2048 characters. OpenRouter fetches them server-side and forwards only the bytes, never the URL.
A text transcript is limited to 10,000 characters.
Seed Audio 1.0 treats both voice (a Seed speaker ID) and input_references as the reference voice, so a request that sends both is rejected with a 400. Use one or the other.

When Seed Audio 1.0 rejects a reference or prompt for reasons of its own, such as a clip it cannot decode, OpenRouter returns a 400 whose message begins with Provider rejected the request: followed by the provider’s explanation, with any base64 payload redacted. Other provider failures are reported with OpenRouter’s own messages.

Non-Speech Prompts

For most TTS models, input is read aloud verbatim. Seed Audio 1.0 instead treats input as a prompt that is forwarded unchanged to the model, so it can describe the delivery of the speech, or describe sound that is not speech at all: sound effects, ambience, or a scene. Omit voice and input_references when you want the model to infer everything from the prompt:

{
  "model": "bytedance-seed/seed-audio-1-0",
  "input": "Heavy rain falling on a tin roof with distant rolling thunder, no voices.",
  "response_format": "mp3"
}

Seed Audio 1.0 caps generated audio at 120 seconds per request; see Seed Audio 1.0 below for its other limits.

Provider-Specific Options

You can pass provider-specific options using the provider parameter. Options are keyed by provider slug, and only the options for the matched provider are forwarded:

{
  "model": "openai/gpt-4o-mini-tts-2025-12-15",
  "input": "Hello world",
  "voice": "alloy",
  "provider": {
    "options": {
      "openai": {
        "instructions": "Speak in a warm, friendly tone."
      }
    }
  }
}

Seed Audio 1.0

Seed Audio 1.0 takes no provider options; every field it reads is derived from the standard parameters. It enforces these limits with a 400:

input is limited to 3000 characters.
speed must be between 0.5 and 2.0.
Generated audio is capped at 120 seconds per request. A prompt that would produce more than that is rejected before any audio is returned.

Azure (MAI-Voice)

Azure serves microsoft/mai-voice-2, microsoft/mai-voice-2.1, and microsoft/mai-voice-2.1-flash. Azure TTS uses SSML internally, but this is fully abstracted, so you only need the standard parameters. The voice parameter takes a full Azure voice ID with the model suffix (:MAI-Voice-2, :MAI-Voice-2.1, or :MAI-Voice-2.1-Flash, e.g., en-US-Harper:MAI-Voice-2.1), and the voice’s locale sets the synthesis language. Set response_format to mp3 or pcm (24 kHz mono). The full list of voices for each model is in the supported_voices field of the models API.

{
  "model": "microsoft/mai-voice-2.1",
  "input": "Welcome to the event!",
  "voice": "en-US-Harper:MAI-Voice-2.1",
  "response_format": "mp3"
}

Google (Gemini TTS)

Gemini 3.8 TTS models read input verbatim, so delivery directions written into the text may be spoken aloud. Pass the style as speech_metadata in provider options instead. It is attached to the input text part of the upstream request, and any other options are forwarded as generation config:

{
  "model": "google/gemini-3.8-flash-lite-tts",
  "input": "Have a wonderful day!",
  "voice": "Kore",
  "response_format": "pcm",
  "provider": {
    "options": {
      "google-ai-studio": {
        "speech_metadata": {
          "style": "warm and friendly"
        }
      }
    }
  }
}

Option	Type	Description
`speech_metadata.style`	string	Sustained delivery style for the input (e.g., `cheerful and friendly`, `whispering`).

Response Format

The TTS endpoint returns a raw audio byte stream, not JSON. The response includes the following headers:

Header	Description
`Content-Type`	The MIME type of the audio. `audio/mpeg` for `mp3` format, `audio/pcm` for `pcm` format
`X-Generation-Id`	The unique generation ID for the request, useful for tracking and debugging

Output Formats

Format	Content-Type	Description
`mp3`	`audio/mpeg`	Compressed audio, smaller file size. Good for storage and playback
`pcm`	`audio/pcm`	Uncompressed raw audio. Lower latency, suitable for real-time streaming pipelines

Pricing

Most TTS models are priced per character of input text. Some audio generation models, such as Seed Audio 1.0, are priced per second of generated audio instead. Pricing varies by model and provider. You can check the per-character or per-second cost for each model on the Models page or via the Models API, where per-second models report their rate under pricing.completion.

OpenAI SDK Compatibility

The TTS endpoint is fully compatible with the OpenAI SDK. You can use the OpenAI client libraries by pointing them at OpenRouter’s base URL:

Best Practices

Choose the right format: Use mp3 for storage and general playback. Use pcm for real-time streaming pipelines where latency matters
Voice selection: Different providers offer different voices. Check the model’s documentation or experiment with available voices to find the best fit for your use case
Input length: For very long texts, consider splitting the input into smaller segments and concatenating the audio output. This can improve reliability and reduce latency for the first audio chunk
Speed parameter: The speed parameter is only supported by certain providers (e.g., OpenAI). Providers that don’t support it either ignore it or reject a non-default value with a 400 when the model has no speed control, so omit it unless the model documents speed

Troubleshooting

Empty or corrupted audio file?

Verify the response_format matches how you’re saving the file (e.g., don’t save pcm output with a .mp3 extension)
Check the response status code, since non-200 responses return JSON error bodies, not audio

Model not found?

Use the Models page to find available TTS models
Verify the model slug is correct (e.g., openai/gpt-4o-mini-tts-2025-12-15, not gpt-4o-mini-tts)

Voice not available?

Available voices vary by provider. Check the provider’s documentation for supported voice identifiers
Each model has its own set of voices, so check the model’s page on the Models page for the full list

Overview

Models & Routing

Tool Calling

Features

Observability

Authentication

Ori

Developer Tools

Privacy

Best Practices

Community

Text-to-Speech

Model Discovery

Via the API

On the Models Page

API Usage

Basic Example

Request Parameters

Voice Cloning

Multiple Reference Clips

Image References

Limits and Requirements

Non-Speech Prompts

Provider-Specific Options

Seed Audio 1.0

Azure (MAI-Voice)

Google (Gemini TTS)

Response Format

Output Formats

Pricing

OpenAI SDK Compatibility

Best Practices

Troubleshooting

​Model Discovery

​Via the API

​On the Models Page

​API Usage

​Basic Example

​Request Parameters

​Voice Cloning

​Multiple Reference Clips

​Image References

​Limits and Requirements

​Non-Speech Prompts

​Provider-Specific Options

​Seed Audio 1.0

​Azure (MAI-Voice)

​Google (Gemini TTS)

​Response Format

​Output Formats

​Pricing

​OpenAI SDK Compatibility

​Best Practices

​Troubleshooting

Model Discovery

Via the API

On the Models Page

API Usage

Basic Example

Request Parameters

Voice Cloning

Multiple Reference Clips

Image References

Limits and Requirements

Non-Speech Prompts

Provider-Specific Options

Seed Audio 1.0

Azure (MAI-Voice)

Google (Gemini TTS)

Response Format

Output Formats

Pricing

OpenAI SDK Compatibility

Best Practices

Troubleshooting