Audio tags mid-sentence
Put a tag such as [whispers], [laughs] or [short pause] exactly where the change should happen. Google lists more than 200, for emotion, pacing and sounds.
The Gemini 3.1 Flash TTS API: Google's expressive speech in 30 voices and 70+ languages, steered with audio tags, at 40% below Google's list price.
gemini-3.1-flash-ttsUpdated from openai import OpenAIclient = OpenAI( base_url="https://api.zurelay.com/v1", api_key="YOUR_ZURELAY_KEY",)speech = client.audio.speech.create( model="gemini-3.1-flash-tts", voice="Kore", input="Welcome to Zurelay. Your video is ready.", instructions="Say it warmly",)speech.write_to_file("welcome.mp3")Base URL https://api.zurelay.com/v1. Works with any OpenAI SDK. Speech API docs
Pricing
Pay as you go from prepaid credit, with no subscription. Top up from $10. Requests that fail are never billed.
| Rate | Zurelay | Google list | You save |
|---|---|---|---|
Text in per 1M tokens | $0.60 | 40%off | |
Audio out per 1M tokens | $12.00 | 40%off |
Billed on the text you send and the audio you get back: 1,920 audio tokens a minute, so about $0.023 a minute of speech. Requests that fail cost nothing.
Every Text-to-speech model and priceMove the sliders to your monthly usage.
$1,200 a year
Cost
Typical workloads at Zurelay's live prices, next to the list price.
About a minute of speech.
About 30 minutes of speech.
About 10 hours of speech.
Overview
Gemini 3.1 Flash TTS is Google's expressive, steerable text-to-speech model, released in preview on April 15, 2026. It reads text aloud in 30 voices and more than 70 languages, and audio tags written into the text change the delivery mid-sentence. The Zurelay Gemini 3.1 Flash TTS API serves it through an OpenAI-compatible speech endpoint at 40% below Google's list price, about $0.023 a minute of speech.
Google launched Gemini 3.1 Flash TTS in preview on April 15, 2026 as a cost-efficient, expressive and steerable speech model. At launch Google called it its most natural and expressive model to date, and it scored an Elo of 1,211 on the Artificial Analysis TTS leaderboard, a ranking built from thousands of blind human preferences. On Zurelay it uses the model ID gemini-3.1-flash-tts, and also accepts google/gemini-3.1-flash-tts and Google's own ID, gemini-3.1-flash-tts-preview.
Its headline feature is audio tags: more than 200 of them, written in square brackets right in the text, such as [whispers], [laughs], [excitement], [slow] or [short pause]. Put a tag exactly where the change should happen, so a line can turn from calm to excited mid-sentence, and keep text or punctuation between two tags. Tags are always in English, but Google says they can be combined with text in other languages. Zurelay sends your input to the model as written, tags included.
Zurelay serves it through an OpenAI-compatible endpoint: POST https://api.zurelay.com/v1/audio/speech with up to 5,000 characters of input, one of Gemini's 30 voices (Kore by default, or an OpenAI name such as alloy or nova), and optional instructions that set the style in plain words, such as "Say cheerfully" or "Read slowly, like a bedtime story". The reply is the audio file itself: MP3 by default, or WAV or raw PCM. A sentence usually takes a few seconds and a paragraph about 15 seconds. Each request speaks in one voice. Google's own API can also voice a multi-speaker dialogue in one request and stream audio as it's made, but neither is available through Zurelay.
Speech is billed on tokens, as on Google's API: the text you send, instructions included, at $0.60 per 1M tokens, and the audio you get back at $12.00 per 1M tokens. Gemini 3.1 Flash TTS counts 32 audio tokens for each second of speech, 1,920 a minute, so a minute costs about $0.023 plus the text. Each reply's x-zurelay-input-tokens and x-zurelay-output-tokens headers show what it was billed on, every request is in your request log, and requests that fail cost nothing.
Strengths
Put a tag such as [whispers], [laughs] or [short pause] exactly where the change should happen. Google lists more than 200, for emotion, pacing and sounds.
At launch Google called it its most natural and expressive model to date, and it scored an Elo of 1,211 on the Artificial Analysis TTS leaderboard.
The language is picked up from the text, so there's nothing to set. Tags stay in English and can be mixed with text in other languages.
Describe the delivery in instructions the way you'd brief a voice actor: mood, pace, accent. The direction shapes the reading and isn't spoken.
Use cases
Spoken game menus and clear audio for people who rely on screen readers, two of the uses Google highlights for the model.
Describe what happens on screen in film, TV or video in a voice that matches the scene, with tags like [whispers] for tense moments.
Stack pacing and emotion tags to build suspense. For long text, Google suggests having Gemini 3.1 Flash-Lite add the tags first, on the same Zurelay key.
Fraud alerts and travel updates that move from serious to reassuring, with [slow] to read out card or flight numbers.
Get started
No waitlist and no new SDK. If your code already talks to OpenAI, it already talks to Zurelay.
Sign up, add credit and create an API key. Set a monthly budget or a rate limit per key if you like.
Change the base URL. Everything else in your code stays the same.
https://api.zurelay.com/v1POST your text and a voice to /v1/audio/speech and get the audio back as MP3, WAV or PCM.
gemini-3.1-flash-ttsWorks with the tools you already use
Compare
$0.30 / $5.40 per 1M tokens
Pick Gemini 3.8 Flash TTS, Google's newest speech model, for more than 130 languages at a lower price per minute, when you don't need instructions.
Gemini 3.8 Flash TTS APIfrom $0.03 per image
Nano Banana 2.1 by Google, 11% below list price.
Nano Banana 2.1 APIfrom $0.035 per image
Nano Banana Pro by Google, 74% below list price.
Nano Banana Pro APIZurelay charges $0.60 per 1M tokens of text in and $12.00 per 1M tokens of audio out. Google's list price is $1.00 and $20.00, so you save 40%. Gemini 3.1 Flash TTS counts 32 audio tokens for each second of speech, 1,920 a minute, so a minute costs about $0.023 plus the text you send. Requests that fail cost nothing.
Create a Zurelay API key, set the SDK's base URL to https://api.zurelay.com/v1 and call client.audio.speech.create with model set to gemini-3.1-flash-tts, a voice, your text as input and, if you like, instructions. The reply is the audio file itself: save it with write_to_file in Python, or write its bytes to a file in Node. To try voices and styles first, use the Speech mode in the Playground at https://zurelay.com/app/playground.
All 30 of Gemini's prebuilt voices, such as Kore, Puck, Charon, Leda, Aoede and Sulafat, with Kore as the default. OpenAI voice names work too and map to a Gemini voice of similar character: alloy speaks as Kore, nova as Leda, onyx as Orus and shimmer as Callirrhoe. The speech docs at https://zurelay.com/docs/speech list every voice with the character Google gives it.
Describe it in instructions, in plain words, as you'd brief a voice actor: "Say cheerfully", "Whisper" or "Read slowly, like a bedtime story". Instructions can be up to 2,000 characters and aren't read out. Google recommends asking for an accent this way rather than relying on the language. speed only accepts 1, so ask for the pace in instructions too.
Yes. Write them in square brackets inside input, exactly where the change should happen, such as [whispers] before a secret or [short pause] before a reveal. Zurelay sends your text to the model as written, tags included. Google lists more than 200 tags, including [excitement], [curiosity], [slow], [fast] and [laughs]. Tags are in English, and Google asks for text or punctuation between two of them. For long text, Google suggests letting Gemini 3.1 Flash-Lite add the tags for you.
More than 70, according to Google. There's no language setting: the language is picked up from the text, so a Spanish sentence is spoken in Spanish. Audio tags stay in English and can be combined with text in other languages.
Each request takes up to 5,000 characters of input and 2,000 of instructions; split longer text into several requests. Audio comes back as MP3 by default, or set response_format to wav, or to pcm for raw 24 kHz, 16-bit mono samples. Google lists an input limit of 8,192 tokens for the model. Zurelay doesn't store the audio, so save the file you get back.
Not in one request on Zurelay. Google's own API can voice a multi-speaker dialogue in one call, but Zurelay speaks each request in a single voice. For a dialogue, make each line with its own voice and join the audio in order; raw PCM is the simplest to join.
A sentence usually takes a few seconds and a paragraph about 15 seconds. The whole file comes back in one reply: Google's own API can stream 3.1 Flash TTS audio as it's made, but streaming isn't available through Zurelay. For a voice assistant, send a sentence at a time so playback starts sooner.
Yes. Google released it in preview on April 15, 2026 and has announced no shutdown date. On Zurelay, Google's ID, gemini-3.1-flash-tts-preview, works as well as gemini-3.1-flash-tts.
Yes. Your requests run on Google's Gemini 3.1 Flash TTS, and Zurelay doesn't swap in a different or smaller model. Speech is generated fresh on every call, so the same text and voice give a slightly different reading each time. Google watermarks all audio from the model with SynthID, an imperceptible mark that identifies it as AI-generated. Zurelay is an independent service and is not affiliated with Google.
Create a key in seconds, point the SDK you already use at Zurelay, and every request costs up to 90% less from the first token.