You prepare the text
You write the material and choose the voice authorised for use.
Products Voice Studio
Voice production Specialised instrument
From text, a voice for your material.
A studio for generating, listening to, and downloading voice materials through text-to-speech models. The interface brings together voice selection and settings useful for production.
- 1Text and voice
- 2Generare audio
- 3Listen and export
Voice Studio
From information to work done.
Reglezi
You adjust the parameters and options available to the voice model.
You revise
You listen, correct, and download the material for the project.
Where it becomes useful.
Narrative
Explanatory materials, presentations, and audio content.
Production
A generation and review flow for content teams.
Experimentare
Comparing interpretations of a text before use.
The tool uses the voice models of the integrated providers. Costs, rights to the voices and the provider’s rules apply.
Voice Studio in detail
What you can do with this project.
Voice Studio is the chain from text to audio file: you write the text, choose the voice, add emotion and pause markers, adjust the model, listen, correct and download. The provider is ElevenLabs — we name it explicitly, because that determines which voices can be used, how much it costs and what you are allowed to do with the result.
The part that is not visible and is in fact the reason the tool exists: the API key never reaches the browser. The interface sends only the text to a proprietary server function, and that function holds the key and talks to ElevenLabs. Without this layer, any page that generates voice exposes its key to anyone who opens the browser console.
It is not Grai. Grai is vocal model research, without a product. Voice Studio is production with an integrated provider, available today. The two do not share any code.
Four models, chosen for different trade-offs
Eleven v3 for expressiveness and emotion markers, Multilingual v2 for stability, Turbo v2.5 for cost, Flash v2.5 for minimal latency. They are not four versions of the same thing: v3 accepts markers that the others ignore, and Flash sacrifices nuance for speed.
Acting markers, not just text
On Eleven v3 you can insert 14 emotion states into the text (from [excited] and [whispers] to [sarcastic] and [tired]), 9 non-verbal and pause markers ([laughs], [sighs], [clears throat], [pause], [breathes]) and 7 delivery markers ([rushed], [drawn out], [singing], [muttering], [quietly]). They are inserted at the cursor position, because placement matters just as much as the marker.
Adjustments that are visible in the result
Stability, similarity to the source voice, style exaggeration, speed between 0.7 and 1.2, plus speaker boost. On v3, stability is not a continuous slider, but three steps — Creative, Natural, Robust — because the model behaves discretely, not continuously, and a fine slider would suggest control that does not exist.
Five export formats, including one for telephony
MP3 at 44.1 kHz 128 or 192 kbps, MP3 22 kHz 32 kbps for small files, PCM 44.1 kHz for raw WAV that goes into an edit, and μ-law 8 kHz — the telephony format, if the material ends up in an IVR or a switchboard.
Eight forceable languages, or detection from text
Romanian, Russian, English, Ukrainian, French, Italian, Spanish, German — or left on automatic. Forcing the language matters in mixed texts, where automatic detection chooses wrongly precisely on proper names.
Data and operation
What enters the system. What needs to be checked.
- The key stays in one place
- In an encrypted environment variable on the server, never in code, in git or in the frontend. The browser only calls our own function. The key gives ElevenLabs only synthesis permission, not access to the whole account, plus a monthly character limit — because the real defence of a key is not its secrecy, but how much harm it can do if it escapes.
- The text reaches the provider
- This is the direct implication of using an external provider: the text you give for synthesis goes to ElevenLabs. A text with personal data, unpublished prices or client information is not sent without authorisation. It is not a formality — it is the only compliance decision the tool transfers to the user.
- The voices are cloned, so consent is required
- The voices fixed in the interface are not library voices, but cloned voices from samples of real people, created through ElevenLabs' cloning interface. That means it is not a matter of product licensing, but of the person's agreement. In the Republic of Moldova, Law 195/2024 treats the voice as biometric data from 23 August 2026.
- Nothing is kept
- Generations remain in the browser tab memory, at most the last 12, and disappear on reload. There is no database, no server-side history, no account. The audio response is served with Cache-Control: no-store.
- What you are allowed to do with the result
- The ElevenLabs subscription grants commercial use rights over the generated material, but their usage policy explicitly forbids using the output to train, test or build a competing system. In practice: the audio file can go into a client material; it cannot become training data for your own model.
From exploration to implementation
How we prepare a project with Voice Studio.
We clarify the material and who listens to it
A radio spot, a presentation narration and an IVR message have different format, rhythm and length requirements. The export format is chosen here, not at the end.
We settle the voice rights first
Which voice is used and on the basis of which agreement. If a new voice is needed, cloning requires samples and the person's consent, not just an audio file.
We generate, listen and correct the text, not the settings
Most pronunciation issues are solved by rewriting the text, not by moving the sliders: abbreviations are expanded, numbers are written in words where it is ambiguous, and proper names are tested in isolation before being placed in a sentence.
We hand over the file and note what is permitted
The checked export, plus a note about the voice used and the limits of use — so the material can be reused after six months without restarting the discussion about rights.
Questions worth clarifying.
Which provider do you use?
ElevenLabs. Our server function calls api.elevenlabs.io directly, the synthesis endpoint. It is not an abstract provider: the available voices, the cost, the output formats and the rules for using the result depend on it.
Which voices can I use?
The ones fixed in the interface, which are cloned voices from provided samples — not library voices. A new voice is created by uploading samples through the ElevenLabs cloning interface, with their 11 MB per file limit; above that, we automatically recompress to mono 22 kHz 64 kbps. The non-technical part is the one that matters: cloning a person's voice requires that person's consent. In the Republic of Moldova, voice is biometric data as of 23 August 2026, so a public recording does not authorise anything.
How much text can I send at once?
5,000 characters per request, a limit imposed by us, not by the provider. The reason is cost: a wrong request with text one hundred times longer is a bill, not an error. The function has a 60-second execution time, which is enough for a block of this size.
In which formats does it export?
MP3 at 44.1 kHz 128 kbps (default) or 192 kbps, MP3 22 kHz 32 kbps for small size, PCM 44.1 kHz for raw WAV, and μ-law 8 kHz for telephony. The last one is the one people forget and then discover that the material sounds poor in the phone exchange.
Can I use the generated audio to train my own voice?
No. An ElevenLabs subscription gives you commercial use rights over the material, but their usage policy explicitly forbids using the output to train, test or develop a competing system. A file delivered to a client is fine; the same file as training data for your own model is not. It is a distinction that is paid for if ignored, and it is why Grai is not trained on anything coming out of here.
Is what I generate stored anywhere?
No. The history lives in the tab memory, at most 12 generations, and disappears on reload. There is no account, database or server storage; the audio response is explicitly marked as non-cacheable. If you want the file, you download it at that moment.
Can I use it on my own?
Not today. The deployment is behind Vercel authentication and, optionally, an additional access token required as a header with each request. It is an internal working tool for producing materials, not a service with registration. What we deliver is the audio material, not access to the panel.
Illustrative example
The narration of a presentation material, in two languages
A use case, without client data or attributed commercial results.
Initial situation
The text of a product presentation, which must be read in Romanian and Russian by the same voice, with the same energy.
How it works
Eleven v3 is chosen for expressiveness and the language is forced explicitly for each variant, instead of being left to automatic detection — in texts with proper names, detection chooses wrongly exactly where it is heard most clearly. Stability is set to Natural; pauses are marked in the text with [pause] and [short pause] at the exact position, not at the end of the sentence. The text is listened to and corrected, not the sliders.
Rezultatul
Two MP3 files at 44.1 kHz, the same voice, the same energy, downloaded at the moment of generation. If the material is to go into an edit, PCM is used instead of MP3, so it is not compressed twice.
What is required:The voice must be one for which the cloned person's consent exists. The text must not contain information that is not allowed to reach an external provider.
Collaboration possibilities
Voice Studio, in your organisation’s context.
Audio content processing
Transcription and voice generation projects on materials for which there are usage rights and a documented purpose.
Private companies
We define a pilot around a real process: users, data, integrations, costs and acceptance criteria. Expansion follows after evaluating the result.
State institutions and companies
We establish the requirements for accessibility, hosting, data protection and interoperability. Any connection to AGE or STISC services requires validation of eligibility, access and approvals.
These are adaptation scenarios, not statements about existing contracts or partnerships. The proposed functions are confirmed within the project scope.
Discuss a pilotPart of an ecosystem.
What would you want to work better?
Tell us about your process. Together we decide what is worth building, what we can connect, and how we check the result.