Haitong Zhang

Haitong Zhang, Yue Lin

Recently, few-shot voice cloning has achieved a significant improvement. However, most models for few-shot voice cloning are single-modal, and multi-modal few-shot voice cloning has been understudied. In this paper, we propose to use multi-modal learning to improve the few-shot voice cloning performance. Inspired by the recent works on unsupervised speech representation, the proposed multi-modal system is built by extending Tacotron2 with an unsupervised speech representation module. We evaluate our proposed system in two few-shot voice cloning scenarios, namely few-shot text-to-speech(TTS) and voice conversion(VC). Experimental results demonstrate that the proposed multi-modal learning can significantly improve the few-shot voice cloning performance over their counterpart single-modal systems.

Haitong Zhang, Yue Lin

Recently, sequence-to-sequence (seq-to-seq) models have been successfully applied in text-to-speech (TTS) to synthesize speech for single-language text. To synthesize speech for multiple languages usually requires multi-lingual speech from the target speaker. However, it is both laborious and expensive to collect high-quality multi-lingual TTS data for the target speakers. In this paper, we proposed to use low-quality code-switched found data from the non-target speakers to achieve cross-lingual voice cloning for the target speakers. Experiments show that our proposed method can generate high-quality code-switched speech in the target voices in terms of both naturalness and speaker consistency. More importantly, we find that our method can achieve a comparable result to the state-of-the-art (SOTA) performance in cross-lingual voice cloning.

Haoyue Zhan, Xinyuan Yu, Haitong Zhang et al.

In this paper, we study the disentanglement of speaker and language representations in non-autoregressive cross-lingual TTS models from various aspects. We propose a phoneme length regulator that solves the length mismatch problem between IPA input sequence and monolingual alignment results. Using the phoneme length regulator, we present a FastPitch-based cross-lingual model with IPA symbols as input representations. Our experiments show that language-independent input representations (e.g. IPA symbols), an increasing number of training speakers, and explicit modeling of speech variance information all encourage non-autoregressive cross-lingual TTS model to disentangle speaker and language representations. The subjective evaluation shows that our proposed model can achieve decent naturalness and speaker similarity in cross-language voice cloning.

Haitong Zhang, Haoyue Zhan, Yang Zhang et al.

International Phonetic Alphabet (IPA) has been widely used in cross-lingual text-to-speech (TTS) to achieve cross-lingual voice cloning (CL VC). However, IPA itself has been understudied in cross-lingual TTS. In this paper, we report some empirical findings of building a cross-lingual TTS model using IPA as inputs. Experiments show that the way to process the IPA and suprasegmental sequence has a negligible impact on the CL VC performance. Furthermore, we find that using a dataset including one speaker per language to build an IPA-based TTS system would fail CL VC since the language-unique IPA and tone/stress symbols could leak the speaker information. In addition, we experiment with different combinations of speakers in the training dataset to further investigate the effect of the number of speakers on the CL VC performance.

Haitong Zhang

This paper presents the description of our submitted system for Voice Conversion Challenge (VCC) 2020 with vector-quantization variational autoencoder (VQ-VAE) with WaveNet as the decoder, i.e., VQ-VAE-WaveNet. VQ-VAE-WaveNet is a nonparallel VAE-based voice conversion that reconstructs the acoustic features along with separating the linguistic information with speaker identity. The model is further improved with the WaveNet cycle as the decoder to generate the high-quality speech waveform, since WaveNet, as an autoregressive neural vocoder, has achieved the SoTA result of waveform generation. In practice, our system can be developed with VCC 2020 dataset for both Task 1 (intra-lingual) and Task 2 (cross-lingual). However, we only submit our system for the intra-lingual voice conversion task. The results of VCC 2020 demonstrate that our system VQ-VAE-WaveNet achieves: 3.04 mean opinion score (MOS) in naturalness and a 3.28 average score in similarity ( the speaker similarity percentage (Sim) of 75.99%) for Task 1. The subjective evaluations also reveal that our system gives top performance when no supervised learning is involved. What's more, our system performs well in some objective evaluations. Specifically, our system achieves an average score of 3.95 in naturalness in automatic naturalness prediction and ranked the 6th and 8th, respectively in ASV-based speaker similarity and spoofing countermeasures.

Haitong Zhang, Yue Lin

Recently, sequence-to-sequence models with attention have been successfully applied in Text-to-speech (TTS). These models can generate near-human speech with a large accurately-transcribed speech corpus. However, preparing such a large data-set is both expensive and laborious. To alleviate the problem of heavy data demand, we propose a novel unsupervised pre-training mechanism in this paper. Specifically, we first use Vector-quantization Variational-Autoencoder (VQ-VAE) to ex-tract the unsupervised linguistic units from large-scale, publicly found, and untranscribed speech. We then pre-train the sequence-to-sequence TTS model by using the<unsupervised linguistic units, audio>pairs. Finally, we fine-tune the model with a small amount of<text, audio>paired data from the target speaker. As a result, both objective and subjective evaluations show that our proposed method can synthesize more intelligible and natural speech with the same amount of paired training data. Besides, we extend our proposed method to the hypothesized low-resource languages and verify the effectiveness of the method using objective evaluation.

Haitong Zhang, Yongping Du, Jiaxin Sun et al.

Word embeddings are substantially successful in capturing semantic relations among words. However, these lexical semantics are difficult to be interpreted. Definition modeling provides a more intuitive way to evaluate embeddings by utilizing them to generate natural language definitions of corresponding words. This task is of great significance for practical application and in-depth understanding of word representations. We propose a novel framework for definition modeling, which can generate reasonable and understandable context-dependent definitions. Moreover, we introduce usage modeling and study whether it is possible to utilize embeddings to generate example sentences of words. These ways are a more direct and explicit expression of embedding's semantics for better interpretability. We extend the single task model to multi-task setting and investigate several joint multi-task models to combine usage modeling and definition modeling together. Experimental results on existing Oxford dataset and a new collected Oxford-2019 dataset show that our single-task model achieves the state-of-the-art result in definition modeling and the multi-task learning methods are helpful for two tasks to improve the performance.

Haitong Zhang, Junbo Zhang, Yujun Wang

In this paper, we propose an attention-based end-to-end model for multi-channel keyword spotting (KWS), which is trained to optimize the KWS result directly. As a result, our model outperforms the baseline model with signal pre-processing techniques in both the clean and noisy testing data. We also found that multi-task learning results in a better performance when the training and testing data are similar. Transfer learning and multi-target spectral mapping can dramatically enhance the robustness to the noisy environment. At 0.1 false alarm (FA) per hour, the model with transfer learning and multi-target mapping gain an absolute 30% improvement in the wake-up rate in the noisy data with SNR about -20.

Haitong Zhang, Junbo Zhang, Yujun Wang

In this paper, we propose a sequence-to-sequence model for keyword spotting (KWS). Compared with other end-to-end architectures for KWS, our model simplifies the pipelines of production-quality KWS system and satisfies the requirement of high accuracy, low-latency, and small-footprint. We also evaluate the performances of different encoder architectures, which include LSTM and GRU. Experiments on the real-world wake-up data show that our approach outperforms the recently proposed attention-based end-to-end model. Specifically speaking, with 73K parameters, our sequence-to-sequence model achieves $\sim$3.05\% false rejection rate (FRR) at 0.1 false alarm (FA) per hour.