1. Why Creators & Enterprises Ask This Question
As generative AI tools become ubiquitous in content production, voice actors, authors, and corporate product teams are rightfully protective of their intellectual property and biometric likeness:
- Unreleased Intellectual Property: Authors synthesizing audiobooks and corporate teams drafting marketing scripts cannot risk sensitive narrative drafts leaking into public datasets.
- Voice Likeness Protection: Creators cloning their voice want ironclad assurances that their biometric vocal timbre will not be appropriated to train public foundational models.
- Enterprise Regulatory Compliance: GDPR, CCPA, and enterprise SOC compliance require clear transparency regarding data processing versus data retention.
2. The Crucial Difference: Inference vs Model Training
Understanding AI data privacy requires understanding two distinct technical operations:
Real-Time Inference (What Voocaly Does)
During inference, your text is passed into memory through fixed mathematical neural weights to compute an output waveform. Once the audio bytes are rendered and delivered, the model weights remain unchanged. Your text is not learned or baked into the network.
Model Training & Fine-Tuning
Training is the computational process of adjusting internal neural weights using backpropagation across vast datasets. If a vendor trains on your data, phrases from your scripts or nuances of your voice can be reconstructed by other users.
Voocaly maintains zero training pipelines for submitted user scripts or audio recordings.
3. Voocaly's Data Governance by Content Type
How each category of customer data is handled across its lifecycle:
Text Scripts & Dialogue Inputs
Transmitted over TLS-encrypted connections. For registered users, script text is stored in private cloud Firestore databases strictly accessible by your authenticated user ID. It is never added to model training corpora.
Uploaded Voice Cloning Audio
Reference audio samples are stored in access-controlled cloud storage buckets under user-scoped paths. They are used solely to extract speaker embeddings for your custom clone and are never utilized for general foundational model pre-training.
Generated Audio Files
Delivered as an in-memory stream to your browser and saved to your workspace History. You can manually delete any generated audio item from your account history at any time with immediate effect, or purge all items via self-service account deletion.
Guest Demo Sessions
Unauthenticated visitors creating demo audio in the studio have their history stored exclusively in their local browser using IndexedDB and localStorage. Zero database records or cloud files are created on our servers.
4. Data Lifecycle & Retention Matrix
| Data Type | Storage Location | Used to Train Models? | Retention & Deletion |
|---|---|---|---|
| Guest Audio & Text | Client Browser (IndexedDB) | No (0% training) | Cleared when browser cache is wiped |
| User Text Scripts | Firestore (Private User Collection) | No (0% training) | Deleted manually or on account closure |
| Generated Audio Files | In-Memory Delivery / Workspace History | No (0% training) | Retained in History; manual instant deletion or account closure |
| Cloning Reference Audio | Access-Controlled Server Storage | No (0% training) | Permanently erased upon clone deletion or account closure |
5. The Distinction: First-Party Platform vs External Inference
To maintain absolute transparency, it is essential to distinguish between Voocaly's own platform policies and upstream inference services:
Voocaly itself develops and operates the application layer, user accounts, and voice management services. Voocaly does not train or fine-tune AI models on your scripts, reference audio recordings, or generated outputs. This guarantee is strictly about Voocaly's own first-party data handling.
When synthesizing audio, requests are processed by upstream commercial inference services. These providers operate under their own independent terms of service and commercial API agreements. Voocaly does not make representations about external provider data policies; users seeking information on any upstream provider's data handling should review that provider's own documentation directly.
Voocaly's architecture is designed to minimize data exposure at every tier, routing only the minimum necessary payload to external inference services and retaining no intermediary copies of synthesis payloads beyond the immediate request lifecycle.
6. Self-Service Account & Data Erasure
We believe users should possess complete sovereignty over their digital footprints. Under our self-service account deletion workflow (/api/account/delete):
- Authentication tokens are immediately revoked.
- Active subscriptions are canceled and unlinked via Paddle.
- All Firestore user profile documents and generation history batches are permanently deleted.
- All stored audio files and cloning samples are recursively purged from cloud storage buckets.
- Custom cloned voice registrations are deleted.
Statutory invoice records are retained exclusively by Paddle as Merchant of Record for international tax and accounting compliance.