about Cosy Voice

[Ozawakun97]

In the generation process of Cosy Voice a flow matching module is employed to convert Speech Tokens to Mel Spectrum my question is why is flow model employed instead of diffusion model？

[aluminumbox]

well maybe @ZhihaoDU can answer this question, have you done some ablation study?

[ZhihaoDU]

In the generation process of Cosy Voice a flow matching module is employed to convert Speech Tokens to Mel Spectrum my question is why is flow model employed instead of diffusion model？

In general, flow matching models require much less iterations (5-10 steps) than denoising diffusion models to obtain a satisfactory Mel spectrum. We choose flow matching rather than denoising diffusion models for efficiency reasons.

[Ozawakun97]

In the generation process of Cosy Voice a flow matching module is employed to convert Speech Tokens to Mel Spectrum my question is why is flow model employed instead of diffusion model？

In general, flow matching models require much less iterations (5-10 steps) than denoising diffusion models to obtain a satisfactory Mel spectrum. We choose flow matching rather than denoising diffusion models for efficiency reasons.

Thanks for answering