Publications

We introduce a novel token-based text-to-speech (TTS) model with 0.8B parameters, trained on a mix of real and synthetic data totaling 650k hours, to address issues like pronunciation accuracy and style consistency. This model integrates a latent variable sequence with enhanced acoustic information into the TTS system, reducing errors and style changes. Our training includes data augmentation for improved timbre consistency, and we use a few-shot voice conversion model to generate diverse voices. This approach enables learning of one-to-many mappings in speech, ensuring both diversity and timbre consistency. Our model outperforms VALL-E in pronunciation, style maintenance, and timbre continuity.

We propose a new end-to-end (E2E) method for automatic speaker verification, specifically tailored for scenarios with multiple enrollment utterances. Unlike conventional systems, which separately optimize front-end models like TDNN for speaker embeddings and back-end models like PLDA for scoring, our approach aims to overcome local optimization limits by jointly optimizing these components. Our model incorporates frame-level and utterance-level attention mechanisms to leverage the relationships among multiple utterances. Additionally, we enhance optimization through data augmentation techniques, including conventional noise augmentation with MUSAN and RIRs datasets, and a novel speaker embedding-level mixup strategy.

This paper introduces HiFi-WaveGAN, a system designed for real-time synthesis of high-quality 48kHz singing voices from full-band mel-spectrograms. It improves upon WaveNet with a generator, incorporates elements from HiFiGAN and UnivNet, and introduces an auxiliary spectrogram-phase loss to enhance high-frequency reconstruction and accelerate training. HiFi-WaveGAN outperforms other neural vocoders like Parallel WaveGAN and HiFiGAN in quality metrics, with faster training and better high-frequency modeling.

This paper presents CrossSinger, a cross-lingual singing voice synthesizer based on Xiaoicesing2. It tackles the challenge of creating a multi-singer high-fidelity singing voice synthesis system with cross-lingual capabilities using only monolingual singers during training. The system unifies language representation, incorporates language information, and removes singer biases. Experimental results show that CrossSinger can synthesize high-quality songs for different singers in various languages, including code-switching cases.

This paper presents XiaoiceSing2, an enhanced singing voice synthesis system that addresses over-smoothing issues in middle- and high-frequency areas of mel-spectrograms. It employs a generative adversarial network (GAN) with improved model architecture to capture finer details.

The ability of countermeasure models to generalize from seen speech synthesis methods to unseen ones has been investigated in the ASVspoof challenge. However, a new mismatch scenario in which fake audio may be generated from real audio with unseen genres has not been studied thoroughly. To this end, we first use five different vocoders to create a new dataset called CN-Spoof based on the CN-Celeb1&2 datasets. Then, we design two auxiliary objectives for regularization via meta-optimization and a genre alignment module, respectively, and combine them with the main anti-spoofing objective using learnable weights for multiple loss terms.

Recent advances in Transformer-based models have enhanced end-to-end automatic speech recognition (ASR), enabling deployment on smart devices. However, these models often require large numbers of parameters. To address this, we introduce a block-reusing strategy for speech Transformers (BRST) and an adapter module (ADM) for efficient parameter use in small footprint ASR systems, without sacrificing accuracy. Tested on the AISHELL-1 corpus, our method achieved low character error rates with significantly fewer parameters, demonstrating both efficiency and effectiveness, particularly with the inclusion of ADM.

We introduce SSI-Net, our submission to the ICASSP 2023 Speech Signal Improvement (SSI) Challenge, designed for real-time communication systems. SSI-Net features a multi-stage architecture, beginning with a time-domain restoration generative adversarial network (TRGAN) for initial speech restoration. In the second stage, we use a lightweight multi-scale temporal frequency convolutional network with axial self-attention (MTFAA-Lite) for fullband speech enhancement. In subjective tests on the SSI Challenge blind test set, SSI-Net achieved a P.835 mean opinion score (MOS) of 3.190 and a P.804 MOS of 3.178, ranking 3rd in tracks 1&2.

Visual speech (i.e., lip motion) is highly related to auditory speech due to the co-occurrence and synchronization in speech production. This paper investigates this correlation and proposes a cross-modal speech co-learning paradigm. The primary motivation of our cross-modal co-learning method is modeling one modality aided by exploiting knowledge from another modality. Specifically, two cross-modal boosters are introduced based on an audio-visual pseudo-siamese structure to learn the modality-transformed correlation. Inside each booster, a max-feature-map embedded Transformer variant is proposed for modality alignment and enhanced feature generation. The network is co-learned both from scratch and with pretrained models. Experimental results on the test scenarios demonstrate that our proposed method achieves around 60% and 20% average relative performance improvement over baseline unimodal and fusion systems, respectively.

Audio Deep Synthesis Detection, Spectro-temporal, Domain Adaptation, Self-Supervised Learning, Frame transition, Greedy Fusion.

Current state-of-the-art automatic speaker verification (ASV) systems are vulnerable to presentation attacks, and several countermeasures (CMs), which distinguish bona fide trials from spoofing ones, have been explored to protect ASV. However, ASV systems and CMs are generally developed and optimized independently without considering their inter-relationship. In this paper, we propose a new spoofing-aware ASV back-end module that efficiently computes a combined ASV score based on speaker similarity and CM score.

We propose a fake audio detection (FAD) system using speaker anonymization (SA) to enhance the distinction between synthetic and natural speech. The key strategy is to remove speaker-specific information from acoustic speech before it’s processed by a deep neural network (DNN)-based FAD system. We employed the McAdams-coefficient-based SA algorithm to ensure that speaker information doesn’t influence DNN learning. Our system, based on a light convolutional neural network bidirectional long short-term memory (LCNN-BLSTM), was tested on the Audio Deep Synthesis Detection Challenge (ADD2022) datasets. Results indicate a 17.66% performance improvement in the first track of ADD2022 by eliminating speaker information from the acoustic speech.

Probabilistic linear discriminant analysis (PLDA) or cosine similarity has been widely used in traditional speaker verification systems as a back-end technique to measure pairwise similarities. To make better use of multiple enrollment utterances, we propose a novel attention back-end model that is applied to the utterance-level features. Specifically, we use scaled-dot self-attention and feed-forward self-attention networks as architectures that learn the intra-relationships of enrollment utterances.

speaker recognition, audio-visual, lip biometrics, deep learning, visual speech.