MaskCycleGAN-based Whisper to Normal Speech Conversion
K. Rohith Gupta, K. Ramnath, S. Johanan Joysingh, P. Vijayalakshmi, T. Nagarajan
TL;DR
The paper addresses whisper-to-normal speech conversion by applying MaskCycleGAN, a mask-guided, cycle-consistent GAN, to spectrogram translation with a MelGAN vocoder for waveform synthesis. It introduces pre-processing via a Voice Activity Detection (VAD) step and tunes mask parameters (frame window and mask size) to preserve speaker-specific characteristics while injecting normal-speech harmonics in voiced regions. On the wTIMIT US subset (24 speakers), the approach is evaluated with PESQ, G-Loss, and MOS, showing that VAD-enabled masking and larger masks improve both objective and subjective quality, achieving a MOS of 4.1 and a best PESQ of 3.159. The results demonstrate that careful mask parameter tuning and VAD pre-processing make MaskCycleGAN a viable method for whisper-to-normal speech conversion with practical implications for inclusive communication and quiet environments.
Abstract
Whisper to normal speech conversion is an active area of research. Various architectures based on generative adversarial networks have been proposed in the recent past. Especially, recent study shows that MaskCycleGAN, which is a mask guided, and cyclic consistency keeping, generative adversarial network, performs really well for voice conversion from spectrogram representations. In the current work we present a MaskCycleGAN approach for the conversion of whispered speech to normal speech. We find that tuning the mask parameters, and pre-processing the signal with a voice activity detector provides superior performance when compared to the existing approach. The wTIMIT dataset is used for evaluation. Objective metrics such as PESQ and G-Loss are used to evaluate the converted speech, along with subjective evaluation using mean opinion score. The results show that the proposed approach offers considerable benefits.