Zero Shot Voice Conditioning for Denoising Diffusion TTS Models

This post presents "Zero Shot Voice Conditioning for Denoising Diffusion TTS Models", a method for speech synthesis from unseen speakers using only a single reference with no training.

Abstract:

We present a novel way of conditioning a pretrained denoising diffusion speech model to produce speech in the voice of a novel person unseen during training. The method requires a short (~3 seconds) sample from the target person, and generation is steered at inference time, without any training steps. At the heart of the method lies a sampling process that combines the estimation of the denoising model with a low-pass version of the new speaker’s sample. The objective and subjective evaluations show that our sampling method can generate a voice similar to that of the target speaker in terms of frequency, with an accuracy comparable to state-of-the-art methods, and without training.

Samples generated from our mathod using the Grad-TTS Diffusion Denoising probabalistic model with Iterative latent variable sampling.

• Grad-TTS - Sample generated from the baseline Grad-TTS model without our sampling method.
• Reference - Speech of a reference speaker.
• Our - sample from the Grad-TTS model using our sampling method. Conditioned on the reference.

Voice Conversion Examples

Non Speech References Examples

N2 Values Scale

ILVR stopping step constant at 6

ILVR Stopping Step Values Scale

N1 constant at 1, N2 constant at 18

References

[1]. Popov V, Vovk I, Gogoryan V, Sadekova T, Kudinov M. Grad-tts: A diffusion probabilistic model for text-to-speech. InInternational Conference on Machine Learning 2021 Jul 1 (pp. 8599-8608). PMLR.

[2]. Choi J, Kim S, Jeong Y, Gwon Y, Yoon S. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938. 2021 Aug 6.