We trained HierSpeech++ with a batch size of 160 for 1,000k steps on eight NVIDIA A6000 GPUs.

Following VITS [35], we utilize a variational autoencoder and a monotonic alignment search (MAS) to align the text and speech internally

We propose a hierarchical speech synthesizer as the backbone speech synthesizer for HierSpeech++

In this study, we propose HierSpeech++, a human-level zeroshot speech synthesis model in terms of naturalness and voice similarity.

Recently, neural audio codec model, have replaced conventional acoustic representations with a high-compressed audio codec.

For a fair comparison, we trained all model with the same dataset (LT460, train-clean-460 subsets of LibriTTS) without YourTTS.

HierVST has significantly improved a voice style transfer performance of the E2E model; therefore so we conduct ablation studies by building up on HierVST

We utilized LibriTTS dataset [90] to train the hierarchical speech synthesizer.

We found a simple trick to transfer the style even with a one second speech prompt by introducing style prompt replication (SPR).

In this work, we propose HierSpeech++, which achieves a human-level high-quality zero-shot speech synthesis performance.

Although our model improves the zero-shot speech synthesis performance significantly, our model also synthesizes the noisy environmental information.

We compared the zero-shot TTS performance of our model with Vall-E, NaturalSpeech 2, and StyleTTS 2.

We compare the performance of zero-shot TTS according to different prompt lengths of 1s, 3s 5s, and 10s.

We compared the zero-shot TTS performance of HierSpeech++ with other baselines: YourTTS, VITS-based end-to-end TTS model and many more.