Responsible AI
Generated audio can be mistaken for a human voice. This page describes the safety limits we build into the product.
Last updated August 18, 2026
Paxa Flash#
What Paxa Flash does, what it will never do, and the rules that apply to everyone who uses it.
Every voice was recorded with consent.
Paxa Labs does not scrape or harvest voice data. Each contributor agreed to the use of their recordings for AI training before recording began.
Paxa Flash does not clone voices.
It ships a fixed catalog of 26 voices. The API accepts no reference audio and offers no other way to imitate a person.
The model speaks only what you write.
Paxa Flash generates speech from text you supply. It generates no text of its own.
Our approach#
Paxa Flash is a general-purpose commercial speech model built by Paxa Labs. It is designed for accessibility and assistive tools, voice interfaces and agents, media narration, education, and localization.
Safety constraints are part of the product design. The catalog is fixed, the training data is documented below, and the rules on this page apply to every account.
Training data
Paxa Flash was trained on three kinds of data.
- Text drawn from publicly available internet sources, which teaches the model the structure, variety, and pronunciation of real-world language.
- Synthetic speech generated with appropriately licensed third-party models, used under their license terms.
- Original voice recordings made for AI training with each contributor's consent, which improve pronunciation and the coverage of accents and dialects.
Boundaries
- Voice cloning
- The model offers none.
- Reference audio
- No endpoint accepts it.
- Voice catalog
- Fixed at 26 voices.
- Customer input
- Discarded once the request is served. We store no submitted text and no generated audio.
- Generated text
- The model produces none.
Acceptable use#
Customers may not use Paxa Flash to do any of the following.
- Impersonate a real person without their consent.
- Commit or facilitate fraud, scams, or social engineering.
- Bypass or attack voice-based authentication.
- Create disinformation, including fabricated statements and false endorsements.
- Harass, defame, or intimidate a person.
- Send unlawful robocalls or spam.
- Produce sexual content involving minors, or any other unlawful content.
Violations result in suspension or termination. The acceptable use section of the Terms of Service carries the binding wording.
Limitations#
Paxa Flash can mispronounce words and produce audio artifacts. Do not use it for safety-critical, medical, or emergency applications.
Disclosure#
We recommend that customers disclose AI-generated audio wherever a listener could reasonably mistake it for a human voice.
Customers are responsible for complying with transparency law in their own jurisdiction, including the synthetic media obligations of the EU AI Act.
Report misuse#
Report misuse of our models to the address below. Include the account, the request, and what you observed.