01
Upload a voice sample
Use a short, clean recording of the speaker whose voice you are authorised to use.
Recommended length: 3-6 seconds
Radiance · Voice Cloning
Create a reusable AI voice from a short reference recording for voice agents, multilingual applications, narration, customer experiences, media, and personalised speech. Provide a clean voice sample, enter new text, and use Shunya Radiance to generate speech that follows the reference speaker's voice characteristics - including across supported languages.
Welcome to Shunya. This is my reference voice sample.
3-6 seconds of clean speechYour appointment has been confirmed for tomorrow morning at ten.
Generated from new textVoice Cloning
Traditional voice recording requires the speaker to record every sentence that an application may eventually need. Voice cloning changes that workflow. Provide a short reference recording and Radiance can use that reference to generate entirely new speech.
3-6 sec
Recommended reference audio for voice cloning
1
Reference voice used to generate new sentences
Cross-lingual
Cloned voices can synthesize supported languages
11
Expression styles available through the Radiance speech stack
| Capability | Radiance Voice Cloning |
|---|---|
| Voice Cloning | ✓ |
| Short Reference Audio | 3-6 sec |
| New Text Generation | ✓ |
| Cross-lingual Synthesis | ✓ |
| Expression Control | ✓ |
| Speed Control | ✓ |
| Batch Generation | ✓ |
| Streaming TTS | ✓ |
| Multiple Audio Formats | ✓ |
| API Integration | ✓ |
| Private Deployment Options | ✓ |
Try it yourself
A useful voice clone should not only repeat the reference sentence. Give it something new to say.
01
Use a short, clean recording of the speaker whose voice you are authorised to use.
Recommended length: 3-6 seconds
02
Enter the exact words spoken in the reference audio.
Welcome to Shunya. This is my reference voice sample.
03
Now enter something completely different.
Your appointment has been confirmed for tomorrow morning at ten.
04
Radiance uses the reference voice to synthesize the new sentence.
Clone this voice in the playground.
How voice cloning works
Voice cloning uses characteristics from a reference recording to guide speech synthesis.
01
Provide a clean sample of the speaker.
02
Provide the exact transcript of what was spoken in that reference clip.
03
The reference audio establishes the voice characteristics used during synthesis.
04
Write a sentence that was never recorded by the original speaker.
05
Radiance generates the new sentence using the reference voice.
Reference audio
The speaker sample
Voice characteristics
Timbre · Prosody · Accent
New text
Words the speaker never recorded
Generated speech
The new sentence, in the reference voice
The words can change. The voice identity remains connected to the reference.
The reference sample matters
Voice cloning quality depends heavily on the audio used as the reference. A noisy or distorted recording gives the model less reliable information about the speaker.
Use a recording where the speaker can be heard clearly.
Avoid traffic, music, other speakers, and environmental noise.
The speaker should remain at a stable distance from the microphone.
Avoid large echoing rooms where possible.
The speaker should talk naturally but clearly.
Use a sample rate of at least 16 kHz where possible.
3-6 seconds
Recommended length
Long recordings are not automatically better. A short, clean, representative voice sample is more useful than a longer recording filled with noise or inconsistent speech.
Reference text
Voice cloning uses both the audio containing the reference speaker and the exact transcript of the reference recording. These two pieces of information work together.
reference_wav
Hello, welcome to our customer support service.
reference_text
Hello, welcome to our customer support service.
New text
A sentence the reference recording never included.
Generated voice
The new sentence is synthesized using characteristics from the reference speaker.
Match the transcript.The reference_text should accurately represent what is spoken in the reference recording. Incorrect reference text can reduce the quality of the resulting voice clone.
No fixed script
The value of voice cloning is not repeating the reference recording. It is generating speech that did not exist before.
ReferenceWelcome to our service.
Your payment has been processed successfully.
Your appointment is scheduled for Monday at four PM.
Your order has left our warehouse and should arrive tomorrow.
Is there anything else I can help you with today?
This makes voice cloning useful for applications where the content changes dynamically.
One voice, more languages
A voice may begin in English. The application may later need Hindi, Tamil, Bengali, Telugu, or anothersupported language. Radiance supports cross-lingual synthesis for cloned voices.
One speaker
Reference voice
Your appointment is confirmed.
आपकी अपॉइंटमेंट कन्फर्म हो गई है।
உங்கள் சந்திப்பு உறுதிசெய்யப்பட்டுள்ளது.
আপনার অ্যাপয়েন্টমেন্ট নিশ্চিত করা হয়েছে।
Same voice identity.The spoken language changes. The voice remains connected to the same reference speaker.
Cross-lingual voice cloning: This can help multilingual products maintain a more consistent voice experience instead of assigning a completely different speaker to every language.
Expressive cloned voices
Voice identity and speaking style are different controls. The same cloned voice may need to sound conversational in one interaction
and more narrative in another. Radiance provides multiple expression styles through the same speech stack.
For voice agents, customer interactions, and everyday dialogue.
For stories, learning material, articles, and long-form speech.
For structured informational delivery.
For positive confirmations and friendly interactions.
For higher-energy speech.
For more subdued delivery.
11
Expression styles
The application can vary how speech is spoken without changing who is speaking.
Real-time voice cloning workflows
Voice cloning becomes particularly useful when combined with applications that generate new text continuously. A voice agent is one example.
Listen → Think → Speak: A cloned voice can become the speaking layer of a conversational AI application.
Built-in voice vs. cloned voice
Not every application needs voice cloning.
Use a built-in voice when
Use a cloned voice when
| Requirement | Built-in voice | Cloned voice |
|---|---|---|
| Reference Audio Needed | - | ✓ |
| Ready Immediately | ✓ | - |
| Unique Speaker Identity | - | ✓ |
| Dynamic New Text | ✓ | ✓ |
| Expression Control | ✓ | ✓ |
| Cross-lingual Speech | ✓ | ✓ |
| Streaming Workflows | ✓ | ✓ |
Different entry point.Same Radiance speech infrastructure.
Responsible voice cloning
Voice cloning should only be used when you have the appropriate rights and permission to use the reference voice.
01
The person whose voice is being cloned should understand and authorise its synthetic use.
02
Make sure you have the rights necessary to use the original reference recording.
03
A voice approved for one project should not automatically be assumed to be approved for every other use.
04
Treat reference recordings as identity-related source material and manage access appropriately.
Control the voice: Voice cloning is a production capability. Consent and authorised use should be part of the workflow, not an afterthought.
Real-world use cases
Give conversational AI a specific authorised voice identity while responses change dynamically.
Maintain a recognisable voice across automated conversations, support journeys, and notifications.
Use one authorised voice across supported languages instead of changing speaker identity between markets.
Generate authorised speech for translated or adapted content while maintaining a consistent speaker identity.
Create new educational speech without recording every lesson sentence individually.
Build personalised speech experiences using an authorised reference voice.
Maintain a consistent synthetic character voice across changing dialogue.
Give internal or customer-facing assistants a controlled organisational voice identity.
Generate changing announcements through the same approved voice.
Create audio from newly generated text without scheduling another recording session.
Developers
Radiance does not require a separate voice-registration step for the inline cloning workflow. The reference can be supplied directly during synthesis.
{
"model": "radiance-indic",
"input": "Your request has been completed successfully.",
"voice": "Varun",
"reference_wav": "<base64-encoded audio>",
"reference_text": "Hello, welcome to our customer support service."
}Input
The sentence the cloned voice should speak.
reference_wav
Base64-encoded WAV, FLAC, or OGG clip of the reference speaker.
reference_text
The exact text spoken in the reference audio.
model
Generated speech using characteristics from the reference voice.
Important.For the inline workflow, the reference is used for the request. A separate permanent voice-registration step is not required.
Shunya-specific parameters: Voice cloning uses reference_wav and reference_text. The voice field is still required, and the reference audio overrides that preset speaker. These fields are available through the Shunya API.
Explore Voice Cloning APIControl the generated voice
After the voice is established, the application still needs control over the generated output.
Voice identity
The authorised speaker clip that establishes who the generated voice sounds like.
Speech control
Choose how the cloned voice speaks and which audio encoding the application receives.
Final audio
Speech synthesized from new text while staying connected to the reference voice.
Text
Choose what the voice says.
Language
Choose the supported language being spoken.
Expression
Choose how the sentence is delivered.
Speed
Control speaking speed.
| Format | Common use |
|---|---|
| MP3 | Web playback and general applications |
| WAV | High-quality audio workflows |
| PCM | Low-latency and real-time pipelines |
| OGG Opus | Efficient streaming and web audio |
| FLAC | Lossless audio |
| μ-law | Telephony and IVR |
| A-law | Telephony and SIP systems |
The voice clone becomes one component of a configurable speech-generation workflow.
Multilingual voice experiences
Voice cloning can be combined with the rest of the Shunya speech stack.
Speech stack
Example
One voice layer.The same voice workflow can become part of multilingual assistants, translated media, customer support, and other speech applications.
Built for production
A convincing five-second demo is not the same as operating a voice product. Production applications need control overreference quality, generated text, language, expressions, audio formats, APIs, deployment, and authorised use.
| Requirement | Radiance |
|---|---|
| Voice Cloning | ✓ |
| Short Reference Audio | 3-6 sec |
| Reference Transcript | ✓ |
| Dynamic New Speech | ✓ |
| Cross-lingual Synthesis | ✓ |
| Expression Control | ✓ |
| Speed Control | ✓ |
| Batch Synthesis | ✓ |
| Streaming TTS | ✓ |
| Multiple Audio Formats | ✓ |
| Telephony Formats | ✓ |
| API Integration | ✓ |
| Voice Agent Integration | ✓ |
| Private Deployment Options | ✓ |
One voice
Same voice identity
Voice cloning becomes infrastructure for applications that need a consistent authorised speaker rather than a one-time audio effect.
Questions
Voice cloning uses a reference recording of a speaker to generate new synthetic speech that follows characteristics of that reference voice.
Yes. Radiance supports voice cloning using reference audio and its transcript.
Shunya's current documentation recommends approximately 3-6 seconds of clean reference audio.
Use clear speech with minimal background noise, limited reverb, consistent volume, and a sample rate of at least 16 kHz where possible.
Yes. The inline cloning workflow uses both reference_wav and reference_text. The reference text should accurately match what the speaker says in the audio.
Yes. The reference establishes the voice characteristics, while the new input text determines what speech is generated.
In the inline voice-cloning workflow, there is no separate registration step. The reference audio and transcript are supplied as part of the synthesis request.
Yes. Radiance supports cross-lingual synthesis for cloned voices across supported languages.
The voice cloning workflow sits inside Radiance, which supports multiple expression styles for speech generation.
Yes. Radiance provides configurable speech-speed control.
Yes. A cloned voice can be used as the TTS layer after an LLM or conversational agent generates its response.
Yes. Cross-lingual synthesis allows one voice identity to be used across supported languages.
Radiance supports MP3, WAV, PCM, OGG Opus, FLAC, μ-law, and A-law output.
The standard OpenAI speech SDK interface does not expose Shunya-specific voice-cloning parameters such as reference_wav and reference_text. Use the Shunya SDK or REST API when those capabilities are required.
You should only clone and synthesize a person's voice when you have the appropriate rights to the reference recording and the voice owner's consent for the intended use.
Shunya's current pricing page lists custom and cloned voices under Enterprise availability. Check the current plan or speak with Shunya before planning a production deployment around voice cloning.
Get Started
Don't test voice cloning with our speaker. Test an authorised voice your application actually needs. Provide a cleanreference clip, enter its transcript, then give the voice something completely new to say. Try another sentence.Change the language. Change the expression. Hear whether the voice identity stays consistent.