Avoy
Book a demo
Avoy sits between your app and your TTS provider. Sensitive values are swapped for realistic fakes before anything leaves your network, then patched back into the audio locally.
Detection covers the data voice applications actually read aloud.
The listener hears the correct sentence. The provider only ever receives a plausible stand-in.
Sensitive spans are found in your text by NER and rules, or marked explicitly.
Each value becomes a realistic fake of the same type and shape.
A final check blocks any outbound request that still contains a real value.
The provider returns natural audio with word-level timings for the fake sentence.
The fake words are cut out and the real ones inserted, rendered on your own machine and converted to the provider's voice.
No change to your voices or prompts. Point your base URL at Avoy.
Compatible with the TTS API you already call. Change the base URL, keep your code.
Cached clips are encrypted at rest and isolated per provider, voice, model and settings.
Real values are rendered and converted on your infrastructure, never uploaded.
When a word must be fetched, it hides among decoys. The provider may learn vocabulary, never linkage.
Unknown proxy routes are blocked, not passed through. No sensitive text in logs or metrics.
English today. Turkish is in progress, with more providers on the roadmap.
Tell us your provider, languages and volume.
A sandbox Avoy instance set up for your voices and data types.
Send real traffic through it and listen to the result.