Text to speech API with designed and preset voices
OpenAI's speech endpoint with Qwen3-TTS and ZONOS2: preset voices, voices designed from a description, emotions and 10 to 40+ languages, per minute.
Last updated 2026-11-05
Black Friday deposit bonus
Every qualifying top-up earns free credit, on top of any other offer.
Ends November 30
Send text, get audio back. Lovely Router serves OpenAI's speech endpoint, POST /v1/audio/speech, with two open-weight voice models on our own GPUs: Qwen3-TTS and ZONOS2. Code written for OpenAI's SDK works by changing the base URL and the key, and on top of OpenAI's fields you can describe a brand-new voice in words, set the language and, on ZONOS2, pick an emotion. Both models cost $0.036 per minute of audio, billed by the second, from the same prepaid balance as your model calls.
Hand it to your AI agent
Copy this prompt into Claude Code, Cursor, Codex or any coding agent. It tells the agent where the docs are and how to use your API key.
Add text to speech to this project with the Lovely Router speech API (OpenAI-compatible). - Docs (read first): https://lovelyrouter.com/docs/models/speech.md and https://lovelyrouter.com/docs/api/audio.md - POST https://lovelyrouter.com/v1/audio/speech with `model`, `input`, `voice` (a preset, an OpenAI voice name or a saved voice id) and optional `instructions`, `response_format`, `speed`, `language`. The answer is the audio file. - The first request after an idle spell can take a while (the model wakes up): use a long client timeout. - Auth: read my API key from the LOVELY_API_KEY environment variable and send it as `Authorization: Bearer $LOVELY_API_KEY`. Never print or commit the key. If it is not set, ask me to create one at https://lovelyrouter.com/app/keys. - Billing is prepaid: every call is charged to my balance, so test with a small request before running anything large. - Every docs page is plain Markdown; the full index is https://lovelyrouter.com/llms.txt. My task: <the text, the language and the voice you want>
Pick a model
qwen3-tts | zonos2 | |
|---|---|---|
| Price | $0.036 per minute | $0.036 per minute |
| Steer the delivery with | instructions in plain words: emotion, pace, volume, a whisper, an accent | emotion: happy, sad, angry, surprised, or weights from 0 to 1 |
| New voice from a description | Yes: leave out voice and describe it | No |
| Preset voices | 9 | 3: american_female, american_male, british_female |
| Languages | 10: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian | 40+, best in English, Chinese and Japanese |
| Sample rate | 24 kHz | 44.1 kHz |
| Speed once warm | About twice real time | About real time per request, faster when several run at once |
| Licence | Apache-2.0, Qwen team | Apache-2.0, Zyphra |
Both models also clone a voice from a short recording; that has its own page, the voice cloning API.
Your first request
Point the OpenAI SDK at https://lovelyrouter.com/v1. The response is the audio file itself, MP3 unless you ask for another format, and the x-lovely-audio-seconds header says how long it is.
curl https://lovelyrouter.com/v1/audio/speech \
-H "Authorization: Bearer $LOVELY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen3-tts",
"input": "Welcome back! Your order has shipped and should arrive on Thursday.",
"voice": "serena",
"instructions": "Warm and cheerful, a little fast"
}' \
--output welcome.mp3A request takes up to 4,096 characters of text. A long text is spoken sentence by sentence in the same voice.
Four ways to choose the voice
1. A preset. qwen3-tts has nine: vivian, serena, uncle_fu, dylan, eric, ryan, aiden, ono_anna and sohee, each with a native language among Chinese (including Beijing and Sichuan dialects), English, Japanese and Korean. Every voice speaks every language the model supports, best in its own.
2. An OpenAI voice name. alloy, nova, onyx and the other OpenAI names map to the closest preset, so an app built for OpenAI's voices keeps working unchanged.
3. A voice designed from a description. Leave voice out and describe the speaker in instructions (age, gender, timbre, pace, mood, accent), and qwen3-tts makes a new voice for it:
{
"model": "qwen3-tts",
"input": "Ahoy! Lower the sails, we're coming into port.",
"instructions": "A gravelly old pirate captain with a booming voice, amused and a little drunk"
}A designed voice comes out a little different each time. When you find one you like, save the audio as a voice and use its id from then on.
4. No voice at all. Without voice or instructions, the model picks a preset that suits language: on qwen3-tts, vivian for Chinese, ono_anna for Japanese, sohee for Korean, else ryan.
Direct the delivery
On qwen3-tts, instructions (up to 1,000 characters) tells a preset or saved voice how to speak, written in any language the way you would brief a narrator: "Slow and conspiratorial, half whispering." On zonos2, emotion changes the delivery while the speaker stays the same: "emotion": "sad", or a blend such as {"happy": 0.7, "surprised": 0.3}. Sending instructions to zonos2, or emotion to qwen3-tts, returns 400 naming the field.
Set language to an ISO code such as en, zh, ja or pt-br when the text is short or full of numbers; the default auto detects it. speed from 0.25 to 4.0 changes the pace without changing the pitch, and seed makes the output repeatable.
Output formats
response_format | You get |
|---|---|
mp3 (default) | MP3, 128 kbps |
opus | Ogg Opus, 64 kbps |
aac | AAC (ADTS) |
flac | FLAC |
wav | WAV, 16-bit |
pcm | Raw 16-bit little-endian samples, 24 kHz mono, as OpenAI sends them |
The audio arrives when it is complete: there is no streaming, and stream_format: "sse" returns 400.
Cold starts, plainly
Both models run on GPUs that sleep when nobody uses them. The first request after a quiet spell waits while the model loads: about 45 seconds on qwen3-tts and up to 2 minutes on zonos2. After that the speeds in the table apply. Give your HTTP client a timeout that covers the load, and send a batch of lines one after another rather than one per hour if latency matters. You are never charged for the wait, only for the seconds of audio made.
What people build with it
Order updates, onboarding tips and notifications in a fixed preset, in each user's language.
One designed voice per character, each kept by saving it as a voice.
The same sentence in ten languages from one preset, slowed down with speed.
An agent writes a reply with any chat model and speaks it with qwen3-tts on the same key.
Characters in video shots can speak in a saved voice.
What you pay
$0.036 per minute of audio on both models, billed by the second, at least one second per request. The speed you choose changes how many seconds you get back, and that is what you pay for. A request that fails is not charged, and there is no hold: a request only needs a positive balance. Each call appears in your logs with the seconds made. There is no subscription; see pricing for the rest of the price list on the same balance.
Need volume at the lowest price? kokoro-82m, hexgrad's small open-weight model, speaks American and British English, Spanish, French, Hindi, Italian, Japanese, Portuguese and Chinese with 54 preset voices (af_heart, bm_george, zf_xiaoxiao, …), no wake-up wait, and costs $0.008 per minute of audio, billed by the second. It has no instructions, emotions or cloning, and returns MP3, WAV or PCM.
Need acting rather than narration? drama-3, Fish Audio's expressive model, takes directions in [brackets] ([whispering], [voice breaking]) anywhere in the text, clones a voice from a WAV or MP3 clip, and speaks 83 languages, at $0.09 per minute of audio. It returns MP3 at normal speed only.
| Model | Input → output | Price |
|---|---|---|
| Qwen3-TTS 1.7B qwen3-tts | text→speech | $0.036 per minute |
| ZONOS2 zonos2 | text→speech | $0.036 per minute |
| Kokoro 82M kokoro-82m | text→speech | $0.008 per minute |
| Fish Audio Drama 3 drama-3 | text→speech | $0.09 per minute |
Live prices in USD, the same the API charges.
Frequently asked questions
Is this compatible with OpenAI's speech API?+
Yes. It is the same endpoint, POST /v1/audio/speech, with the same fields and output formats, so the official OpenAI SDKs work against https://lovelyrouter.com/v1. OpenAI voice names map to the closest preset. The extra fields (language, emotion, seed, cloning) are optional.
How much does text to speech cost?+
$0.036 per minute of audio for both models, billed by the second with a one-second minimum per request. Failed requests are free.
Which languages are supported?+
qwen3-tts speaks 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. zonos2 speaks more than 40, best in English, Chinese and Japanese.
Why did my first request take so long?+
The model was asleep and had to load: about 45 seconds for qwen3-tts, up to 2 minutes for zonos2. Requests right after that are fast. The wait is not billed.
How long can the text be?+
Up to 4,096 characters per request, and up to 1,000 characters of instructions. Split a longer script into several requests in the same voice.
Can I stream the audio as it is made?+
Not yet. The audio comes back in one response once it is complete.
Every field and error is in the audio API reference; the speech guide has the voice list and more examples.