Skip to main content

TEXT TO SPEECH

Luna-TTS

Expressive speech. A consistent voice. Built for real-time applications.

Luna-TTS is the AI text-to-speech model from VUI Labs, built for natural expression, accurate delivery, and consistent voice identity. It brings speech generation and voice cloning to video narration, storytelling, and interactive voice applications. The Luna-TTS API supports system and cloned voices, expression tags, and complete or streaming audio output.

Hear the difference​

From an intimate line to a story, a conversation, or a game scene, listen to six examples from the VUI Labs showcase. Press play to explore the delivery, and expand the text to read along.

01Expressive voiceEnglish

Okay, okay, breathe. You don't have to be fine right now. I'll wait. I'm good at waiting.

0:00
0:08
Read the transcript

Okay, okay, breathe. You don't have to be fine right now. I'll wait. I'm good at waiting.

02A conversation that flowsEnglish

Thanks for calling Harbor and Vine. How can I help you today? Hi, I'd like to book a table for four this Friday around 7:30, if possible. Sure, let me just check the a…

0:00
1:29
Read the transcript

Thanks for calling Harbor and Vine. How can I help you today? Hi, I'd like to book a table for four this Friday around 7:30, if possible. Sure, let me just check the availability for you. I'm sorry, we don't have anything available at exactly 7:30, but I can offer you 7:15 or 7:45. Would either of those times work? Oh, okay. Let's go with 7:45, then. And could we possibly have a quieter table? It's actually my parents' anniversary, so we'd like somewhere a little more peaceful. Of course, I'd be happy to add that request. Just so you know, we can't guarantee the exact seating location, but we'll definitely do our best to arrange a quieter table for you. Also, may I ask if anyone in your party has any allergies or dietary requirements? Yes, actually. One person in our group is gluten-free. Got it. I'll make a note of that. And can I please get the name for the reservation and a mobile number? Sure, it's Daniel Carter, and the number is 415-555-0148. Thanks, Daniel. Let me confirm that for you. You're all booked for four people this Friday at 7:45 PM under the name Daniel Carter. I've added the request for a quieter table, and noted that one guest is gluten-free. You'll receive a text confirmation shortly. Is there anything else I can help you with today? That's everything. Thanks so much. You're very welcome. You're all set, and we look forward to welcoming you on Friday. Have a great day.

03AudiobooksEnglish

It is a truth universally acknowledged, that a single man in possession of a good fortune must be in want of a wife. At Longbourn, this truth became especially interes…

0:00
0:23
Read the transcript

It is a truth universally acknowledged, that a single man in possession of a good fortune must be in want of a wife. At Longbourn, this truth became especially interesting whenever a wealthy young gentleman arrived in the neighborhood. To Mrs. Bennet, such an arrival was not merely news— it was an opportunity.

04PodcastsEnglish

Welcome back to Read the Floor. Today, we're looking at the possession that never shows up in a box score. The point guard comes off the screen, sees the help defender…

0:00
0:20
Read the transcript

Welcome back to Read the Floor. Today, we're looking at the possession that never shows up in a box score. The point guard comes off the screen, sees the help defender lean one step toward the paint, and pauses just half a second. That pause changes everything. The corner opens, the defense rotates, and the pass is already on its way.

05DocumentariesEnglish

Every star begins in darkness. Inside a vast cloud of gas and dust, gravity gathers matter into a tightening sphere. As the pressure rises, the core grows hotter, unti…

0:00
0:14
Read the transcript

Every star begins in darkness. Inside a vast cloud of gas and dust, gravity gathers matter into a tightening sphere. As the pressure rises, the core grows hotter, until hydrogen nuclei begin to fuse.

06Game voiceoversEnglish

Attention, all shelter personnel. This is an Omega-level emergency broadcast. The Helios-class orbital cannon has locked onto Meridian Colony. Atmospheric defense has…

0:00
0:15
Read the transcript

Attention, all shelter personnel. This is an Omega-level emergency broadcast. The Helios-class orbital cannon has locked onto Meridian Colony. Atmospheric defense has failed. Surface evacuation is no longer possible.

What makes Luna-TTS different​

  • Expression beyond reading aloud. Generate speech with natural phrasing and add supported vocal expressions such as laughter, sighs, and breaths within a line.
  • A recognizable voice across your content. Choose a system voice or create a custom voice from a short reference recording, then reuse it across scripts.
  • Clear words. Consistent delivery. Bring accurate pronunciation, a recognizable voice, and natural rhythm to longer narratives and batch voiceover production.
  • Speech for interactive experiences. Stream audio as it is generated to build voice assistants, spoken replies, and product guidance.
  • A multilingual foundation. Trained on Chinese, English, Japanese, and Korean speech to learn the pronunciation, rhythm, and expression of different languages.

From a single ad to an entire course, we want every generation to carry clear pronunciation, a familiar voice, and expression that fits the content—so great speech becomes part of your everyday production workflow.

Listen to the details​

Clear pronunciation​

Names, numbers, and instructions need to land clearly. These seven boarding-announcement samples bring English and Chinese voices into an everyday listening scenario.

Ella · Take 1English
0:00
0:09
Ella · Take 2English
0:00
0:09
Olivia · Take 1English
0:00
0:09
Olivia · Take 2English
0:00
0:09
Listen to 3 more samples
Zhixia · Take 1Chinese
0:00
0:09
Zhixia · Take 2Chinese
0:00
0:09
Zhixia · Take 3Chinese
0:00
0:09

Rhythm and phrasing​

A voice comes alive through its pauses, emphasis, and changing pace. Explore seven readings across different voices and both languages.

Fiona · Take 1English
0:00
0:15
James · Take 1English
0:00
0:12
Jasper · Take 1English
0:00
0:18
Jinyu · Take 1Chinese
0:00
0:12
Listen to 3 more samples
Jinyu · Take 2Chinese
0:00
0:12
Xingran · Take 1Chinese
0:00
0:18
Xingyu · Take 1Chinese
0:00
0:19

Recognized for speech quality​

Luna-TTS has topped Hugging Face TTS Arena and reached the top three in Artificial Analysis Provider Voice TTS, earning listeners' preference in blind comparisons. From pronunciation to expression, we keep refining every part of the voice experience.

EvaluationAchievement
Hugging Face TTS ArenaNo. 1 in the overall Rating ranking
Artificial AnalysisNo. 3 in Provider Voice TTS
Seed-TTS-Eval, Chinese0.73% character error rate (CER)
Seed-TTS-Eval, English1.49% word error rate (WER)

How Luna-TTS works​

Parallel refinement of speech​

Luna-TTS uses masked diffusion over discrete speech tokens. Instead of committing to every speech token in a single left-to-right pass, the model predicts multiple masked positions, keeps confident predictions, and revisits uncertain ones over successive rounds. This lets it refine the utterance using broader context.

A compact backbone and staged training​

Luna-TTS builds on a 0.6B-parameter backbone and Luna-Codec. Its training pipeline uses approximately one million hours of speech, followed by annealing on approximately 100,000 hours of high-quality speech. Reinforcement learning prioritizes content correctness, then speaker similarity when correctness is close.

From full-context synthesis to streaming​

The real-time architecture uses block-causal attention to generate speech in chunks while retaining prior context. It balances natural delivery with fast responses, helping speech flow smoothly in a conversation. Explore the technology →

Choose your voice​

Start with a voice suited to your content. Compare the same short script across a few voices before generating a longer piece. Examples from the Luna-TTS system voice catalog include:

VoiceVoice IDStyle
FionafionaSoft, nuanced female voice with an intimate tone
JamesjamesPolished British male voice
AveryaveryFriendly, warm young American female voice

For a custom voice, the Voice Cloning API accepts a 3–60 second reference recording, up to 20 MiB. Use a clear recording of one speaker whose voice you are authorized to use, create the voice, and pass its saved voice_id to Luna-TTS.

You can also create a voice from a text description with the Voice Design API and use the saved voice_id for Luna-TTS synthesis.

How to use Luna-TTS​

Create an API key, choose an enabled voice, and set model_id to luna-tts-1.5. Replace both placeholders below before running the request. The response is saved as a WAV file.

export VUILABS_API_BASE_URL="https://api.vuilabs.ai"
export VUILABS_API_KEY="your-api-key"

curl --fail-with-body --request POST \
--url "$VUILABS_API_BASE_URL/v1/text-to-speech" \
--header "X-API-Key: $VUILABS_API_KEY" \
--header "Content-Type: application/json" \
--data '{
"model_id": "luna-tts-1.5",
"text": "Every story begins with a voice. Let yours be heard.",
"voice_id": "YOUR_PROJECT_VOICE_ID",
"language": "en",
"speed": 1,
"audio_format": "wav"
}' \
--dump-header luna-tts.headers \
--output luna-tts.response || { cat luna-tts.response >&2; exit 1; }

if ! grep -Eiq '^content-type:[[:space:]]*audio/wav([[:space:];]|$)' luna-tts.headers || [ ! -s luna-tts.response ]; then
cat luna-tts.response >&2
echo 'Expected a non-empty audio/wav response' >&2
exit 1
fi
mv luna-tts.response luna-tts.wav

See the API reference for the complete request contract and JavaScript and Python examples.

Shape the delivery​

A range of emotions​

Calm, joy, sadness, surprise—each delivery gives a line a different feeling. Explore 34 short samples grouped by voice. Within a voice, compare the same text number across emotions to hear the change.

English voice 01English · 5 clips
Sad · Text 1
0:00
0:02
Surprised · Text 1
0:00
0:01
Calm · Text 3
0:00
0:01
Sad · Text 3
0:00
0:01
Surprised · Text 3
0:00
0:01
English voice 02English · 3 clips
Calm · Text 3
0:00
0:01
Sad · Text 3
0:00
0:01
Surprised · Text 3
0:00
0:01
English voice 03English · 3 clips
Calm · Text 1
0:00
0:01
Fearful · Text 1
0:00
0:01
Surprised · Text 1
0:00
0:01
English voice 04English · 2 clips
Calm · Text 3
0:00
0:01
Surprised · Text 3
0:00
0:01
QingyanChinese · 6 clips
Calm
0:00
0:01
Happy
0:00
0:01
Sad
0:00
0:02
Angry
0:00
0:01
Fearful
0:00
0:01
Surprised
0:00
0:01
QiyanChinese · 5 clips
Calm
0:00
0:01
Happy
0:00
0:01
Sad
0:00
0:02
Fearful
0:00
0:01
Surprised
0:00
0:01
ShuyanChinese · 3 clips
Happy
0:00
0:01
Sad
0:00
0:02
Fearful
0:00
0:01
Chinese voice 01Chinese · 4 clips
Calm · Text 1
0:00
0:01
Sad · Text 1
0:00
0:01
Angry · Text 1
0:00
0:01
Fearful · Text 1
0:00
0:01
Chinese voice 02Chinese · 3 clips
Calm · Text 1
0:00
0:01
Sad · Text 1
0:00
0:02
Angry · Text 1
0:00
0:01

Add vocal expressions​

Place supported tags directly in text where the sound should occur. For example:

{
"model_id": "luna-tts-1.5",
"voice_id": "YOUR_PROJECT_VOICE_ID",
"text": "[sighs] That was a long day. [chuckles] But we made it.",
"language": "en"
}

Supported tags include [laughs], [chuckles], [sighs], and [gasps]. Use the official tag list when writing scripts.

Listen to four samples that bring the small sounds of speech into the performance.

Grace · Take 1English
0:00
0:07
Grace · Take 3English
0:00
0:06
Xingran · Take 1Chinese
0:00
0:10
Xingran · Take 3Chinese
0:00
0:08

Adjust pacing and volume​

Set speed between 0.6 and 1.2. Synchronous requests also support volume between -5 and 5; omit volume for streaming requests. Start at speed: 1 and volume: 0, then adjust one setting at a time. Text normalization is enabled by default; set text_normalization: false to preserve the input text before synthesis.

Stream audio into your application​

Use POST /v1/text-to-speech/stream with model_id: "luna-tts-1.5" to receive audio progressively. The response contains ordered audio bytes that your application can consume before generation completes.

Complete and streaming responses both support WAV and PCM. For raw PCM playback, configure the decoder using the sample rate, channel count, and bit depth in the response headers. See the streaming examples.

What you can build​

Use caseHow to use Luna-TTS
Video and product narrationTurn a script into a voiceover and adjust pacing for the edit.
Stories and educational contentReuse an approved voice across passages and lessons.
Voice assistantsConvert your application's text replies into streaming speech.
Guided and accessible experiencesRead instructions, notifications, and on-screen content aloud.

For longer content, split scripts at sentence or paragraph boundaries. The current API accepts up to 500 weighted characters per request: Han characters (including extensions), hiragana, katakana, Hangul, and ー/ー count as 2; CR/LF count as 0; other Unicode code points count as 1. 250 pure CJK characters are accepted; 251 return 400/10013 before synthesis or billing. Reuse the same voice and controls across segments, then check the transitions. See the request fields and request limits.

A voice that carries the passage​

Listen beyond a single sentence. These six longer passages let you hear how the voice, rhythm, and delivery hold together as the narration unfolds.

Grace · Take 1English
0:00
0:32
Grace · Take 2English
0:00
0:32
Grace · Take 3English
0:00
0:32
Jinyu · Take 1Chinese · Slow
0:00
0:31
Listen to 2 more samples
Qiyan · Take 1Chinese
0:00
0:29
Qiyan · Take 2Chinese
0:00
0:30

Frequently asked questions​

What is Luna-TTS?​

Luna-TTS is the VUI Labs AI text-to-speech model for expressive speech generation and voice cloning. It supports content production and interactive applications through complete and streaming audio output.

Which languages does Luna-TTS support?​

Luna-TTS is trained on Chinese, English, Japanese, and Korean speech. Choose your language and voice in the Console, and see the voice selection guide to get started.

Does Luna-TTS support voice cloning?​

Yes. Create a custom voice from a reference recording with the Voice Cloning API, then reuse its voice_id with model_id: "luna-tts-1.5".

Can I use Luna-TTS in a real-time application?​

Yes. The streaming endpoint lets your application receive and process audio as it is generated, bringing timely spoken replies to assistants, customer service, and guided experiences.

How is Luna-TTS priced?​

Speech generation is billed by text usage. See Pricing for current rates, free allowances, character-counting rules, and the organization voice-slot rules for free cloning.

Learn more​

Ready to build? Generate your first audio, explore the API reference, or contact support.