Token-based text-to-speech (TTS) models have emerged as a promising avenue for generating natural and realistic speech, yet they grapple with low pronunciation accuracy, speaking style and timbre inconsistency, and a substantial need for diverse training data. In response, we introduce a novel hierarchical acoustic modeling approach complemented by a tailored data augmentation strategy and train it on the combination of real and synthetic data, scaling the data size up to 650k hours, leading to the zero-shot TTS model with 0.8B parameters. Specifically, our method incorporates a latent variable sequence containing supplementary acoustic information based on refined self-supervised learning (SSL) discrete units into the TTS model by a predictor. This significantly mitigates pronunciation errors and style mutations in synthesized speech. During training, we strategically replace and duplicate segments of the data to enhance timbre uniformity. Moreover, a pretrained few-shot voice conversion model is utilized to generate a plethora of voices with identical content yet varied timbres. This facilitates the explicit learning of utterance-level one-to-many mappings, enriching speech diversity and also ensuring consistency in timbre. Comparative experiments (Demo page: https://anonymous.4open.science/w/ham-tts/)demonstrate our model's superiority over VALL-E in pronunciation precision and maintaining speaking style, as well as timbre continuity.
It is challenging to build a multi-singer high-fidelity singing voice synthesis system with cross-lingual ability by only using monolingual singers in the training stage. In this paper, we propose CrossSinger, which is a cross-lingual singing voice synthesizer based on Xiaoicesing2. Specifically, we utilize International Phonetic Alphabet to unify the representation for all languages of the training data. Moreover, we leverage conditional layer normalization to incorporate the language information into the model for better pronunciation when singers meet unseen languages. Additionally, gradient reversal layer (GRL) is utilized to remove singer biases included in lyrics since all singers are monolingual, which indicates singer's identity is implicitly associated with the text. The experiment is conducted on a combination of three singing voice datasets containing Japanese Kiritan dataset, English NUS-48E dataset, and one internal Chinese dataset. The result shows CrossSinger can synthesize high-fidelity songs for various singers with cross-lingual ability, including code-switch cases.
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. The results on our cross-genre evaluation dataset for anti-spoofing show that the proposed method significantly improved the generalization ability of the countermeasures compared with the baseline system in the genre mismatch scenario.
Transformer-based models have recently made significant achievements in the application of end-to-end (E2E) automatic speech recognition (ASR). It is possible to deploy the E2E ASR system on smart devices with the help of Transformer-based models. While these models still have the disadvantage of requiring a large number of model parameters. To overcome the drawback of universal Transformer models for the application of ASR on edge devices, we propose a solution that can reuse the block in Transformer models for the occasion of the small footprint ASR system, which meets the objective of accommodating resource limitations without compromising recognition accuracy. Specifically, we design a novel block-reusing strategy for speech Transformer (BRST) to enhance the effectiveness of parameters and propose an adapter module (ADM) that can produce a compact and adaptable model with only a few additional trainable parameters accompanying each reusing block. We conducted an experiment with the proposed method on the public AISHELL-1 corpus, and the results show that the proposed approach achieves the character error rate (CER) of 9.3%/6.63% with only 7.6M/8.3M parameters without and with the ADM, respectively. In addition, we also make a deeper analysis to show the effect of ADM in the general block-reusing method.
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 LRSLip3, GridLip, LomGridLip, and VoxLip datasets demonstrate that our proposed method achieves 60% and 20% average relative performance improvement over independently trained audio-only/visual-only and baseline fusion systems, respectively.
XiaoiceSing is a singing voice synthesis (SVS) system that aims at generating 48kHz singing voices. However, the mel-spectrogram generated by it is over-smoothing in middle- and high-frequency areas due to no special design for modeling the details of these parts. In this paper, we propose XiaoiceSing2, which can generate the details of middle- and high-frequency parts to better construct the full-band mel-spectrogram. Specifically, in order to alleviate this problem, XiaoiceSing2 adopts a generative adversarial network (GAN), which consists of a FastSpeech-based generator and a multi-band discriminator. We improve the feed-forward Transformer (FFT) block by adding multiple residual convolutional blocks in parallel with the self-attention block to balance the local and global features. The multi-band discriminator contains three sub-discriminators responsible for low-, middle-, and high-frequency parts of the mel-spectrogram, respectively. Each sub-discriminator is composed of several segment discriminators (SD) and detail discriminators (DD) to distinguish the audio from different aspects. The experiment on our internal 48kHz singing voice dataset shows XiaoiceSing2 significantly improves the quality of the singing voice over XiaoiceSing.
Entertainment-oriented singing voice synthesis (SVS) requires a vocoder to generate high-fidelity (e.g. 48kHz) audio. However, most text-to-speech (TTS) vocoders cannot work well in this scenario even if the neural vocoder for TTS has achieved significant progress. In this paper, we propose HiFi-WaveGAN which is designed for synthesizing the 48kHz high-quality singing voices from the full-band mel-spectrogram in real-time. Specifically, it consists of a generator improved from WaveNet, a multi-period discriminator same to HiFiGAN, and a multi-resolution spectrogram discriminator borrowed from UnivNet. To better reconstruct the high-frequency part from the full-band mel-spectrogram, we design a novel auxiliary spectrogram-phase loss to train the neural network, which can also accelerate the training process. The experimental result shows that our proposed HiFi-WaveGAN significantly outperforms other neural vocoders such as Parallel WaveGAN (PWG) and HiFiGAN in the mean opinion score (MOS) metric for the 48kHz SVS task. And a comparative study of HiFi-WaveGAN with/without phase loss term proves that phase loss indeed improves the training speed. Besides, we also compare the spectrogram generated by our HiFi-WaveGAN and PWG, which shows our HiFi-WaveGAN has a more powerful ability to model the high-frequency parts.
The Audio Deep Synthesis Detection (ADD) Challenge has been held to detect generated human-like speech. With our submitted system, this paper provides an overall assessment of track 1 (Low-quality Fake Audio Detection) and track 2 (Partially Fake Audio Detection). In this paper, spectro-temporal artifacts were detected using raw temporal signals, spectral features, as well as deep embedding features. To address track 1, low-quality data augmentation, domain adaptation via finetuning, and various complementary feature information fusion were aggregated in our system. Furthermore, we analyzed the clustering characteristics of subsystems with different features by visualization method and explained the effectiveness of our proposed greedy fusion strategy. As for track 2, frame transition and smoothing were detected using self-supervised learning structure to capture the manipulation of PF attacks in the time domain. We ranked 4th and 5th in track 1 and track 2, respectively.
Audio-visual (AV) lip biometrics is a promising authentication technique that leverages the benefits of both the audio and visual modalities in speech communication. Previous works have demonstrated the usefulness of AV lip biometrics. However, the lack of a sizeable AV database hinders the exploration of deep-learning-based audio-visual lip biometrics. To address this problem, we compile a moderate-size database using existing public databases. Meanwhile, we establish the DeepLip AV lip biometrics system realized with a convolutional neural network (CNN) based video module, a time-delay neural network (TDNN) based audio module, and a multimodal fusion module. Our experiments show that DeepLip outperforms traditional speaker recognition models in context modeling and achieves over 50% relative improvements compared with our best single modality baseline, with an equal error rate of 0.75% and 1.11% on the test datasets, respectively.
A back-end model is a key element of modern speaker verification systems. Probabilistic linear discriminant analysis (PLDA) has been widely used as a back-end model in speaker verification. However, it cannot fully make use of multiple utterances from enrollment speakers. In this paper, we propose a novel attention-based back-end model, which can be used for both text-independent (TI) and text-dependent (TD) speaker verification with multiple enrollment utterances, and employ scaled-dot self-attention and feed-forward self-attention networks as architectures that learn the intra-relationships of the enrollment utterances. In order to verify the proposed attention back-end, we combine it with two completely different but dominant speaker encoders, which are time delay neural network (TDNN) and ResNet trained using the additive-margin-based softmax loss and the uniform loss, and compare them with the conventional PLDA or cosine scoring approaches. Experimental results on a multi-genre dataset called CN-Celeb show that the performance of our proposed approach outperforms PLDA scoring with TDNN and cosine scoring with ResNet by around 14.1% and 7.8% in relative EER, respectively. Additionally, an ablation experiment is also reported in this paper for examining the impact of some significant hyper-parameters for the proposed back-end model.