Tacotron-based text-to-speech (TTS) systems directly synthesize speech from text input. Such frameworks typically consist of a feature prediction network that maps character sequences to frequency-domain acoustic features, followed by a waveform reconstruction algorithm or a neural vocoder that generates the time-domain waveform from acoustic features. As the loss function is usually calculated only for frequency-domain acoustic features, that doesn't directly control the quality of the generated time-domain waveform. To address this problem, we propose a new training scheme for Tacotron-based TTS, referred to as WaveTTS, that has 2 loss functions: 1) time-domain loss, denoted as the waveform loss, that measures the distortion between the natural and generated waveform; and 2) frequency-domain loss, that measures the Mel-scale acoustic feature loss between the natural and generated acoustic features. WaveTTS ensures both the quality of the acoustic features and the resulting speech waveform. To our best knowledge, this is the first implementation of Tacotron with joint time-frequency domain loss. Experimental results show that the proposed framework outperforms the baselines and achieves high-quality synthesized speech.
Targeting at both high efficiency and performance, we propose AlignTTS to predict the mel-spectrum in parallel. AlignTTS is based on a Feed-Forward Transformer which generates mel-spectrum from a sequence of characters, and the duration of each character is determined by a duration predictor.Instead of adopting the attention mechanism in Transformer TTS to align text to mel-spectrum, the alignment loss is presented to consider all possible alignments in training by use of dynamic programming. Experiments on the LJSpeech dataset show that our model achieves not only state-of-the-art performance which outperforms Transformer TTS by 0.03 in mean option score (MOS), but also a high efficiency which is more than 50 times faster than real-time.
While neural end-to-end text-to-speech (TTS) is superior to conventional statistical methods in many ways, the exposure bias problem in the autoregressive models remains an issue to be resolved. The exposure bias problem arises from the mismatch between the training and inference process, that results in unpredictable performance for out-of-domain test data at run-time. To overcome this, we propose a teacher-student training scheme for Tacotron-based TTS by introducing a distillation loss function in addition to the feature loss function. We first train a Tacotron2-based TTS model by always providing natural speech frames to the decoder, that serves as a teacher model. We then train another Tacotron2-based model as a student model, of which the decoder takes the predicted speech frames as input, similar to how the decoder works during run-time inference. With the distillation loss, the student model learns the output probabilities from the teacher model, that is called knowledge distillation. Experiments show that our proposed training scheme consistently improves the voice quality for out-of-domain test data both in Chinese and English systems.
Neural network based end-to-end text to speech (TTS) has significantly improved the quality of synthesized speech. Prominent methods (e.g., Tacotron 2) usually first generate mel-spectrogram from text, and then synthesize speech from mel-spectrogram using vocoder such as WaveNet. Compared with traditional concatenative and statistical parametric approaches, neural network based end-to-end models suffer from slow inference speed, and the synthesized speech is usually not robust (i.e., some words are skipped or repeated) and lack of controllability (voice speed or prosody control). In this work, we propose a novel feed-forward network based on Transformer to generate mel-spectrogram in parallel for TTS. Specifically, we extract attention alignments from an encoder-decoder based teacher model for phoneme duration prediction, which is used by a length regulator to expand the source phoneme sequence to match the length of target mel-sprectrogram sequence for parallel mel-sprectrogram generation. Experiments on the LJSpeech dataset show that our parallel model matches autoregressive models in terms of speech quality, nearly eliminates the skipped words and repeated words, and can adjust voice speed smoothly. Most importantly, compared with autoregressive models, our model speeds up the mel-sprectrogram generation by 270x. Therefore, we call our model FastSpeech. We will release the code on Github.
Most deep learning-based models for speech enhancement have mainly focused on estimating the magnitude of spectrogram while reusing the phase from noisy speech for reconstruction. This is due to the difficulty of estimating the phase of clean speech. To improve speech enhancement performance, we tackle the phase estimation problem in three ways. First, we propose Deep Complex U-Net, an advanced U-Net structured model incorporating well-defined complex-valued building blocks to deal with complex-valued spectrograms. Second, we propose a polar coordinate-wise complex-valued masking method to reflect the distribution of complex ideal ratio masks. Third, we define a novel loss function, weighted source-to-distortion ratio (wSDR) loss, which is designed to directly correlate with a quantitative evaluation measure. Our model was evaluated on a mixture of the Voice Bank corpus and DEMAND database, which has been widely used by many deep learning models for speech enhancement. Ablation experiments were conducted on the mixed dataset showing that all three proposed approaches are empirically valid. Experimental results show that the proposed method achieves state-of-the-art performance in all metrics, outperforming previous approaches by a large margin.
Although end-to-end neural text-to-speech (TTS) methods (such as Tacotron2) are proposed and achieve state-of-the-art performance, they still suffer from two problems: 1) low efficiency during training and inference; 2) hard to model long dependency using current recurrent neural networks (RNNs). Inspired by the success of Transformer network in neural machine translation (NMT), in this paper, we introduce and adapt the multi-head attention mechanism to replace the RNN structures and also the original attention mechanism in Tacotron2. With the help of multi-head self-attention, the hidden states in the encoder and decoder are constructed in parallel, which improves the training efficiency. Meanwhile, any two inputs at different times are connected directly by self-attention mechanism, which solves the long range dependency problem effectively. Using phoneme sequences as input, our Transformer TTS network generates mel spectrograms, followed by a WaveNet vocoder to output the final audio results. Experiments are conducted to test the efficiency and performance of our new network. For the efficiency, our Transformer TTS network can speed up the training about 4.25 times faster compared with Tacotron2. For the performance, rigorous human tests show that our proposed model achieves state-of-the-art performance (outperforms Tacotron2 with a gap of 0.048) and is very close to human quality (4.39 vs 4.44 in MOS).
Neural waveform models such as the WaveNet are used in many recent text-to-speech systems, but the original WaveNet is quite slow in waveform generation because of its autoregressive (AR) structure. Although faster non-AR models were recently reported, they may be prohibitively complicated due to the use of a distilling training method and the blend of other disparate training criteria. This study proposes a non-AR neural source-filter waveform model that can be directly trained using spectrum-based training criteria and the stochastic gradient descent method. Given the input acoustic features, the proposed model first uses a source module to generate a sine-based excitation signal and then uses a filter module to transform the excitation signal into the output speech waveform. Our experiments demonstrated that the proposed model generated waveforms at least 100 times faster than the AR WaveNet and the quality of its synthetic speech is close to that of speech generated by the AR WaveNet. Ablation test results showed that both the sine-wave excitation signal and the spectrum-based training criteria were essential to the performance of the proposed model.
Although end-to-end neural text-to-speech (TTS) methods (such as Tacotron2) are proposed and achieve state-of-the-art performance, they still suffer from two problems: 1) low efficiency during training and inference; 2) hard to model long dependency using current recurrent neural networks (RNNs). Inspired by the success of Transformer network in neural machine translation (NMT), in this paper, we introduce and adapt the multi-head attention mechanism to replace the RNN structures and also the original attention mechanism in Tacotron2. With the help of multi-head self-attention, the hidden states in the encoder and decoder are constructed in parallel, which improves the training efficiency. Meanwhile, any two inputs at different times are connected directly by self-attention mechanism, which solves the long range dependency problem effectively. Using phoneme sequences as input, our Transformer TTS network generates mel spectrograms, followed by a WaveNet vocoder to output the final audio results. Experiments are conducted to test the efficiency and performance of our new network. For the efficiency, our Transformer TTS network can speed up the training about 4.25 times faster compared with Tacotron2. For the performance, rigorous human tests show that our proposed model achieves state-of-the-art performance (outperforms Tacotron2 with a gap of 0.048) and is very close to human quality (4.39 vs 4.44 in MOS).
Sequence-to-sequence attention-based models have recently shown very promising results on automatic speech recognition (ASR) tasks, which integrate an acoustic, pronunciation and language model into a single neural network. In these models, the Transformer, a new sequence-to-sequence attention-based model relying entirely on self-attention without using RNNs or convolutions, achieves a new single-model state-of-the-art BLEU on neural machine translation (NMT) tasks. Since the outstanding performance of the Transformer, we extend it to speech and concentrate on it as the basic architecture of sequence-to-sequence attention-based model on Mandarin Chinese ASR tasks. Furthermore, we investigate a comparison between syllable based model and context-independent phoneme (CI-phoneme) based model with the Transformer in Mandarin Chinese. Additionally, a greedy cascading decoder with the Transformer is proposed for mapping CI-phoneme sequences and syllable sequences into word sequences. Experiments on HKUST datasets demonstrate that syllable based model with the Transformer performs better than CI-phoneme based counterpart, and achieves a character error rate (CER) of \emph{$28.77\%$}, which is competitive to the state-of-the-art CER of $28.0\%$ by the joint CTC-attention based encoder-decoder network.
This paper proposes a method for generating speech from filterbank mel frequency cepstral coefficients (MFCC), which are widely used in speech applications, such as ASR, but are generally considered unusable for speech synthesis. First, we predict fundamental frequency and voicing information from MFCCs with an autoregressive recurrent neural net. Second, the spectral envelope information contained in MFCCs is converted to all-pole filters, and a pitch-synchronous excitation model matched to these filters is trained. Finally, we introduce a generative adversarial network -based noise model to add a realistic high-frequency stochastic component to the modeled excitation signal. The results show that high quality speech reconstruction can be obtained, given only MFCC information at test time.
Attention-based encoder-decoder architectures such as Listen, Attend, and Spell (LAS), subsume the acoustic, pronunciation and language model components of a traditional automatic speech recognition (ASR) system into a single neural network. In our previous work, we have shown that such architectures are comparable to state-of-the-art ASR systems on dictation tasks, but it was not clear if such architectures would be practical for more challenging tasks such as voice search. In this work, we explore a variety of structural and optimization improvements to our LAS model which significantly improve performance. On the structural side, we show that word piece models can be used instead of graphemes. We introduce a multi-head attention architecture, which offers improvements over the commonly-used single-head attention. On the optimization side, we explore techniques such as synchronous training, scheduled sampling, label smoothing, and minimum word error rate optimization, which are all shown to improve accuracy. We present results with a unidirectional LSTM encoder for streaming recognition. On a 12,500 hour voice search task, we find that the proposed changes improve the WER of the LAS system from 9.2% to 5.6%, while the best conventional system achieve 6.7% WER. We also test both models on a dictation dataset, and our model provide 4.1% WER while the conventional system provides 5% WER.