Hello,
I build a personal voice assistant on the Gemini API (paid tier, Gemini 3.8 Flash-Lite TTS). Its voice was created with Voice Design from a text prompt, so it is fully synthetic and is not based on any real person’s voice.
The generated audio has a noticeable noise floor in pauses and some high-frequency hiss. We clean it on our side with a standard denoiser (ffmpeg afftdn) and a noise gate.
I have two questions:
-
Is it allowed to post-process Gemini TTS output this way (denoise, gate, EQ) before playing it to users?
-
We would like to create a cloned voice using these cleaned samples of our own Voice Design voice as the reference, so the model produces cleaner audio directly. The cloning flow asks for a consent recording from the voice owner. Since no human is behind this voice, what is the correct way to do this? Would a consent statement generated with the same designed voice be acceptable, or is there another procedure for synthetic voices created in our own project?
-
Can we use this designed voice and audio generated with it in our product long-term?
We don’t want to bypass any safeguard, so we would rather ask first.
Thank you,
Nick