Non-autoregressive real-time Accent Conversion model with voice cloning

Abstract

Currently, the development of Foreign Accent Conversion (FAC) models utilizes deep neural network architectures, as well as ensembles of neural networks for speech recognition and speech generation. The use of these models is limited by architectural features, which does not allow flexible changes in the timbre of the generated speech and requires the accumulation of context, leading to increased delays in generation and makes these systems unsuitable for use in real-time multi-user communication scenarios. We have developed the non-autoregressive model for real-time accent conversion with voice cloning. The model generates native-sounding L1 speech with minimal latency based on input L2 accented speech. The model consists of interconnected modules for extracting accent, gender, and speaker embeddings, converting speech, generating spectrograms, and decoding the resulting spectrogram into an audio signal. The model has the ability to save, clone and change the timbre, gender and accent of the speaker's voice in real time. The results of the objective assessment show that the model improves speech quality, leading to enhanced recognition performance in existing ASR systems. The results of subjective tests show that the proposed accent and gender encoder improves the generation quality. The developed model demonstrates high-quality low-latency accent conversion, voice cloning, and speech enhancement capabilities, making it suitable for real-time multi-user communication scenarios.