| Total: 1123
In this paper, we present SANE-TTS, a stable and natural end-to-end multilingual TTS model. By the difficulty of obtaining multilingual corpus for given speaker, training multilingual TTS model with monolingual corpora is unavoidable. We introduce speaker regularization loss that improves speech naturalness during cross-lingual synthesis as well as domain adversarial training, which is applied in other multilingual TTS models. Furthermore, by adding speaker regularization loss, replacing speaker embedding with zero vector in duration predictor stabilizes cross-lingual inference. With this replacement, our model generates speeches with moderate rhythm regardless of source speaker in cross-lingual synthesis. In MOS evaluation, SANE-TTS achieves naturalness score above 3.80 both in cross-lingual and intralingual synthesis, where the ground truth score is 3.99. Also, SANE-TTS maintains speaker similarity close to that of ground truth even in cross-lingual inference. Audio samples are available on our web page.
The recently developed pitch-controllable text-to-speech (TTS) model, i.e. FastPitch, was conditioned for the pitch contours. However, the quality of the synthesized speech degraded considerably for pitch values that deviated significantly from the average pitch; i.e. the ability to control vocal pitch was limited. To address this issue, we propose two algorithms to improve the robustness of FastPitch. First, we propose a novel timbre-preserving pitch-shifting algorithm for natural pitch augmentation. Pitch-shifted speech samples sound more natural when using the proposed algorithm because the speaker's vocal timbre is maintained. Moreover, we propose a training algorithm that defines FastPitch using pitch-augmented speech datasets with different pitch ranges for the same sentence. The experimental results demonstrate that the proposed algorithms improve the pitch controllability of FastPitch.
Since neural Text-To-Speech models have achieved such high standards in terms of naturalness, the main focus of the field has gradually shifted to gaining more control over the expressiveness of the synthetic voices. One of these leverages is the control of the speaking rate that has become harder for a human operator to control since the introduction of neural attention networks to model speech dynamics. While numerous models have reintroduced an explicit duration control (ex: FastSpeech2), these models generally rely on additional tasks to complete during their training. In this paper, we show how an acoustic analysis of the internal embeddings delivered by the encoder of an unsupervised end-to-end TTS Tacotron2 model is enough to identify and control some acoustic parameters of interest. Specifically, we compare this speaking rate control with the duration control offered by a supervised FastSpeech2 model. Experimental results show that the control provided by embeddings reproduces a behaviour closer to natural speech data.
Three research directions that have recently advanced the text-to-speech (TTS) field are end-to-end architecture, prosody control modeling, and on-the-fly duration alignment of non-auto-regressive models. However, these three agendas have yet to be tackled at once in a single solution. Current studies are limited either by a lack of control over prosody modeling or by the inefficient training inherent in building a two-stage TTS pipeline. We propose TriniTTS, a pitch-controllable end-to-end TTS without an external aligner that generates natural speech by addressing the issues mentioned above at once. It eliminates the training inefficiency in the two-stage TTS pipeline by the end-to-end architecture. Moreover, it manages to learn the latent vector representing the data distribution of the speeches through performing tasks (alignment search, pitch estimation, waveform generation) simultaneously. Experimental results demonstrate that TriniTTS enables prosody modeling with user input parameters to generate deterministic speech, while synthesizing comparable speech to the state-of-the-art VITS. Furthermore, eliminating normalizing flow modules used in VITS increases the inference speed by 28.84% in CPU environment and by 29.16% in GPU environment.
In neural text-to-speech (TTS), two-stage system or a cascade of separately learned models have shown synthesis quality close to human speech. For example, FastSpeech2 transforms an input text to a mel-spectrogram and then HiFi-GAN generates a raw waveform from a mel-spectogram where they are called an acoustic feature generator and a neural vocoder respectively. However, their training pipeline is somewhat cumbersome in that it requires a fine-tuning and an accurate speech-text alignment for optimal performance. In this work, we present end-to-end text-to-speech (E2E-TTS) model which has a simplified training pipeline and outperforms a cascade of separately learned models. Specifically, our proposed model is jointly trained FastSpeech2 and HiFi-GAN with an alignment module. Since there is no acoustic feature mismatch between training and inference, it does not requires fine-tuning. Furthermore, we remove dependency on an external speech-text alignment tool by adopting an alignment learning objective in our joint training framework. Experiments on LJSpeech corpus shows that the proposed model outperforms publicly available, state-of-the-art implementations of ESPNet2-TTS on subjective evaluation (MOS) and some objective evaluations.
We present EdiTTS, an off-the-shelf speech editing methodology based on score-based generative modeling for text-to-speech synthesis. EdiTTS allows for targeted, granular editing of audio, both in terms of content and pitch, without the need for any additional training, task-specific optimization, or architectural modifications to the score-based model backbone. Specifically, we apply coarse yet deliberate perturbations in the Gaussian prior space to induce desired behavior from the diffusion model while applying masks and softening kernels to ensure that iterative edits are applied only to the target region. Through listening tests and speech-to-text back transcription, we show that EdiTTS outperforms existing baselines and produces robust samples that satisfy user-imposed requirements.
For text-to-speech (TTS) synthesis, prosodic structure prediction (PSP) plays an important role in producing natural and intelligible speech. Although inter-utterance linguistic information can influence the speech interpretation of the target utterance, previous works on PSP mainly focus on utilizing intrautterance linguistic information of the current utterance only. This work proposes to use inter-utterance linguistic information to improve the performance of PSP. Multi-level contextual information, which includes both inter-utterance and intrautterance linguistic information, is extracted by a hierarchical encoder from character level, utterance level and discourse level of the input text. Then a multi-task learning (MTL) decoder predicts prosodic boundaries from multi-level contextual information. Objective evaluation results on two datasets show that our method achieves better F1 scores in predicting prosodic word (PW), prosodic phrase (PPH) and intonational phrase (IPH). It demonstrates the effectiveness of using multi-level contextual information for PSP. Subjective preference tests also indicate the naturalness of synthesized speeches are improved.
We propose SpeechPainter, a model for filling in gaps of up to one second in speech samples by leveraging an auxiliary textual input. We demonstrate that the model performs speech inpainting with the appropriate content, while maintaining speaker identity, prosody and recording environment conditions, and generalizing to unseen speakers. Our approach significantly outperforms baselines constructed using adaptive TTS, as judged by human raters in side-by-side preference and MOS tests.
Grapheme-to-phoneme (G2P) conversion is an indispensable part of the Chinese Mandarin text-to-speech (TTS) system, and the core of G2P conversion is to solve the problem of polyphone disambiguation, which is to pick up the correct pronunciation for several candidates for a Chinese polyphonic character. In this paper, we propose a Chinese polyphone BERT model to predict the pronunciations of Chinese polyphonic characters. Firstly, we create 741 new Chinese monophonic characters from 354 source Chinese polyphonic characters by pronunciation. Then we get a Chinese polyphone BERT by extending a pre-trained Chinese BERT with 741 new Chinese monophonic characters and adding a corresponding embedding layer for new tokens, which is initialized by the embeddings of source Chinese polyphonic characters. In this way, we can turn the polyphone disambiguation task into a pre-training task of the Chinese polyphone BERT. Experimental results demonstrate the effectiveness of the proposed model, and the polyphone BERT model obtain 2% (from 92.1% to 94.1%) improvement of average accuracy compared with the BERT-based classifier model, which is the prior state-of-the-art in polyphone disambiguation.
End-to-end TTS requires a large amount of speech/text paired data to cover all necessary knowledge, particularly how to pronounce different words in diverse contexts, so that a neural model may learn such knowledge accordingly. But in real applications, such high demand of training data is hard to be satisfied and additional knowledge often needs to be injected manually. For example, to capture pronunciation knowledge on languages without regular orthography, a complicated grapheme-to-phoneme pipeline needs to be built based on a large structured pronunciation lexicon, leading to extra, sometimes high, costs to extend neural TTS to such languages. In this paper, we propose a framework to learn to automatically extract knowledge from unstructured external resources using a novel Token2Knowledge attention module. The framework is applied to build a TTS model named Neural Lexicon Reader that extracts pronunciations from raw lexicon texts in an end-to-end manner. Experiments show the proposed model significantly reduces pronunciation errors in low-resource, end-to-end Chinese TTS, and the lexicon-reading capability can be transferred to other languages with a smaller amount of data.
In this study, we tackle massively multilingual grapheme-to-phoneme conversion through implementing G2P models based on ByT5. We have curated a G2P dataset from various sources that covers around 100 languages and trained large-scale multilingual G2P models based on ByT5. We found that ByT5 operating on byte-level inputs significantly outperformed the token-based mT5 model in terms of multilingual G2P. Pairwise comparison with monolingual models in these languages suggests that multilingual ByT5 models generally lower the phone error rate by jointly learning from a variety of languages. The pretrained model can further benefit low resource G2P through zero-shot prediction on unseen languages or provides pretrained weights for finetuning, which helps the model converge to a lower phone error rate than randomly initialized weights. To facilitate future research on multilingual G2P, we make available our code and pretrained multilingual G2P models at: https://github.com/lingjzhu/CharsiuG2P.
We propose a new task of synthesizing speech directly from semi-structured documents where the extracted text tokens from OCR systems may not be in the correct reading order due to the complex document layout. We refer to this task as layout-informed document-level TTS and present the DocSpeech dataset which consists of 10K audio clips of a single-speaker reading layout-enriched Word document. For each document, we provide the natural reading order of text tokens, their corresponding bounding boxes, and the audio clips synthesized in the correct reading order. We also introduce DocLayoutTTS, a Transformer encoder-decoder architecture that generates speech in an end-to-end manner given a document image with OCR extracted text. Our architecture simultaneously learns text reordering and mel-spectrogram prediction in a multi-task setup. Moreover, we take advantage of curriculum learning to progressively learn longer, more challenging document-level text utilizing both \texttt{DocSpeech} and LJSpeech datasets. Our empirical results show that the underlying task is challenging. Our proposed architecture performs slightly better than competitive baseline TTS models with a pre-trained model providing reading order priors. We release samples of the DocSpeech dataset.
Recently, leveraging BERT pre-training to improve the phoneme encoder in text to speech (TTS) has drawn increasing attention. However, the works apply pre-training with character-based units to enhance the TTS phoneme encoder, which is inconsistent with the TTS fine-tuning that takes phonemes as input. Pre-training only with phonemes as input can alleviate the input mismatch but lack the ability to model rich representations and sematic information due to limited phoneme vocabulary. In this paper, we propose Mixed-Phoneme BERT, a novel variant of the BERT model that uses mixed phoneme and sup-phoneme representations to enhance the learning capability. Specifically, we merge the adjacent phonemes into sup-phonemes and combine the phoneme sequence and the merged sup-phoneme sequence as the model input, which can enhance the model capacity to learn rich contextual representations. Experiment results demonstrate that our proposed Mixed-Phoneme BERT significantly improves the TTS performance with 0.30 CMOS gain compared with the FastSpeech 2 baseline. The Mixed-Phoneme BERT achieves $3\times$ inference speedup and similar voice quality to the previous TTS pre-trained model PnG BERT.
An unsupervised text-to-speech synthesis (TTS) system learns to generate speech waveforms corresponding to any written sentence in a language by observing: 1) a collection of untranscribed speech waveforms in that language; 2) a collection of texts written in that language without access to any transcribed speech. Developing such a system can significantly improve the availability of speech technology to languages without a large amount of parallel speech and text data. This paper proposes an unsupervised TTS system based on an alignment module that outputs pseudo-text and another synthesis module that uses pseudo-text for training and real text for inference. Our unsupervised system can achieve comparable performance to the supervised system in seven languages with about 10-20 hours of speech each. A careful study on the effect of text units and vocoders has also been conducted to better understand what factors may affect unsupervised TTS performance. The samples generated by our models can be found at https://cactuswiththoughts.github.io/UnsupTTS-Demo, and our code can be found at https://github.com/lwang114/UnsupTTS.
In recent years, parallel end-to-end speech synthesis systems have outperformed the 2-stage TTS approaches in audio quality and latency. A parallel end-to-end speech like VITS can generate the audio with high MOS comparable to ground truth and achieve low latency on GPU. However, the VITS still has high latency when synthesizing long utterances on CPUs. Therefore, in this paper, we propose a streaming method for the parallel speech synthesis model like VITS to synthesize with the long texts effectively on CPU. Our system has achieved human-like speech quality in both the non-streaming and streaming mode on the in-house Vietnamese evaluation set, while the synthesis speed of our system is approximately four times faster than that of the VITS in the non-streaming mode. Furthermore, the customer perceived latency of our system in streaming mode is 25 times faster than the VITS on computer CPU. Our system in non-streaming mode achieves a MOS of 4.43 compared to ground-truth with MOS 4.56; it also has high-quality speech with a MOS of 4.35 in streaming mode. Finally, we release a Vietnamese single accent dataset used in our experiments.
Automatically predicting the outcome of subjective listening tests is a challenging task. Ratings may vary from person to person even if preferences are consistent across listeners. While previous work has focused on predicting listeners' ratings (mean opinion scores) of individual stimuli, we focus on the simpler task of predicting subjective preference given two speech stimuli for the same text. We propose a model based on anti-symmetric twin neural networks, trained on pairs of waveforms and their corresponding preference scores. We explore both attention and recurrent neural nets to account for the fact that stimuli in a pair are not time aligned. To obtain a large training set we convert listeners' ratings from MUSHRA tests to values that reflect how often one stimulus in the pair was rated higher than the other. Specifically, we evaluate performance on data obtained from twelve MUSHRA evaluations conducted over five years, containing different TTS systems, built from data of different speakers. Our results compare favourably to a state-of-the-art model trained to predict MOS scores.
Background music (BGM) plays an essential role in audiobooks, which can enhance the immersive experience of audiences and help them better understand the story. However, well-designed BGM still requires human effort in the text-to-speech (TTS) audiobook production, which is quite time-consuming and costly. In this paper, we introduce an automatic soundtracking system for TTS-based audiobooks. The proposed system divides the soundtracking process into three tasks: plot partition, plot classification, and music selection. The experiments shows that both our plot partition module and plot classification module outperform baselines by a large margin. Furthermore, TTS-based audiobooks produced with our proposed automatic soundtracking system achieves comparable performance to that produced with the human soundtracking system. To our best of knowledge, this is the first work of automatic soundtracking system for audiobooks. Demos are available on https://acst1223.github.io/interspeech2022/main.
This study aims at designing an environment-aware text-to-speech (TTS) system that can generate speech to suit specific acoustic environments. It is also motivated by the desire to leverage massive data of speech audio from heterogeneous sources in TTS system development. The key idea is to model the acoustic environment in speech audio as a factor of data variability and incorporate it as a condition in the process of neural network based speech synthesis. Two embedding extractors are trained with two purposely constructed datasets for characterization and disentanglement of speaker and environment factors in speech. A neural network model is trained to generate speech from extracted speaker and environment embeddings. Objective and subjective evaluation results demonstrate that the proposed TTS system is able to effectively disentangle speaker and environment factors and synthesize speech audio that carries designated speaker characteristics and environment attribute. Audio samples are available online for demonstration.
End-to-end speech synthesis models directly convert the input characters into an audio representation (e.g., spectrograms). Despite their impressive performance, such models have difficulty disambiguating the pronunciations of identically spelled words. To mitigate this issue, a separate Grapheme-to-Phoneme (G2P) model can be employed to convert the characters into phonemes before synthesizing the audio. This paper proposes SoundChoice, a novel G2P architecture that processes entire sentences rather than operating at the word level. The proposed architecture takes advantage of a weighted homograph loss (that improves disambiguation), exploits curriculum learning (that gradually switches from wordlevel to sentence-level G2P), and integrates word embeddings from BERT (for further performance improvement). Moreover, the model inherits the best practices in speech recognition, including multi-task learning with Connectionist Temporal Classification (CTC) and beam search with an embedded language model. As a result, SoundChoice achieves a Phoneme Error Rate (PER) of 2.65% on whole-sentence transcription using data from LibriSpeech and Wikipedia.
Text normalization (TN) systems in production are largely rule-based using weighted finite-state transducers (WFST). However, WFST-based systems struggle with ambiguous input when the normalized form is context-dependent. On the other hand, neural text normalization systems can take context into account but they suffer from unrecoverable errors and require labeled normalization datasets, which are hard to collect. We propose a new hybrid approach that combines the benefits of rule-based and neural systems. First, a non-deterministic WFST outputs all normalization candidates, and then a neural language model picks the best one -- similar to shallow fusion for automatic speech recognition. While the WFST prevents unrecoverable errors, the language model resolves contextual ambiguity. We show for English that the approach is effective and easy to extend. It achieves comparable or better results than existing state-of-the-art TN models.
The goal of automatic dubbing is to perform speech-to-speech translation while achieving audiovisual coherence. This entails isochrony, i.e., translating the original speech by also matching its prosodic structure into phrases and pauses, especially when the speaker's mouth is visible. In previous work, we introduced a prosodic alignment model to address isochrone or on-screen dubbing. In this work, we extend the prosodic alignment model to also address off-screen dubbing that requires less stringent synchronization constraints. We conduct experiments on four dubbing directions – English to French, Italian, German and Spanish – on a publicly available collection of TED Talks and on publicly available YouTube videos. Empirical results show that compared to our previous work the extended prosodic alignment model provides significantly better subjective viewing experience on videos in which on-screen and off-screen automatic dubbing is applied for sentences with speakers mouth visible and not visible, respectively.
In human speech, the attitude of a speaker cannot be fully expressed only by the textual content. It has to come along with the intonation. Declarative questions are commonly used in daily Cantonese conversations, and they are usually uttered with rising intonation. Vanilla neural text-to-speech (TTS) systems are not capable of synthesizing rising intonation for these sentences due to the loss of semantic information. Though it has become more common to complement the systems with extra language models, their performance in modeling rising intonation is not well studied. In this paper, we propose to complement the Cantonese TTS model with a BERT-based statement/question classifier. We design different training strategies and compare their performance. We conduct our experiments on a Cantonese corpus named CanTTS. Empirical results show that the separate training approach obtains the best generalization performance and feasibility.
This paper proposes a task and method for estimating a sequence of facial action units (AUs) solely from speech. AUs were introduced in the facial action coding system to objectively describe facial muscle activations. Our motivation is that AUs can be useful continuous quantities for representing speaker's subtle emotional states, attitudes, and moods in a variety of applications such as expressive speech synthesis and emotional voice conversion. We hypothesize that the information about the speaker's facial muscle movements is expressed in the generated speech and can somehow be predicted from speech alone. To verify this, we devise a neural network model that predicts an AU sequence from the mel-spectrogram of input speech and train it using a large-scale audio-visual dataset consisting of many speaking face-tracks. We call our method and model ``crossmodal AU sequence estimation/estimator (CAUSE)''. We implemented several of the most basic architectures for CAUSE, and quantitatively confirmed that the fully convolutional architecture performed best. Furthermore, by combining CAUSE with an AU-conditioned image-to-image translation method, we implemented a system that animates a given still face image from speech. Using this system, we confirmed the potential usefulness of AUs as a representation of non-linguistic features via subjective evaluations.
With the recent developments in speech synthesis via machine learning, this study explores incorporating linguistics knowledge to visualise and evaluate synthetic speech model training. If changes to the first and second formant (in turn, the vowel space) can be seen and heard in synthetic speech, this knowledge can inform speech synthesis technology developers. A speech synthesis model trained on a large General American English database was fine-tuned into a New Zealand English voice to identify if the changes in the vowel space of synthetic speech could be seen and heard. The vowel spaces at different intervals during the fine-tuning were analysed to determine if the model learned the New Zealand English vowel space. Our findings based on vowel space analysis show that we can visualise how a speech synthesis model learns the vowel space of the database it is trained on. Perception tests confirmed that humans could perceive when a speech synthesis model has learned characteristics of the speech database it is training on. Using the vowel space as an intermediary evaluation helps understand what sounds are to be added to the training database and build speech synthesis models based on linguistics knowledge.
Training a text-to-speech (TTS) model requires a large-scale text labeled speech corpus, which is troublesome to collect. In this paper, we propose a transfer learning framework for TTS that utilizes a large amount of unlabeled speech dataset for pre-training. By leveraging wav2vec2.0 representation, unlabeled speech can highly improve performance, especially in the lack of labeled speech. We also extend the proposed method to zero-shot multi-speaker TTS (ZS-TTS). The experimental results verify the effectiveness of the proposed method in terms of naturalness, intelligibility, and speaker generalization. We highlight that the single speaker TTS model fine-tuned on only 10 minutes of labeled dataset outperforms the other baselines, and the ZS-TTS model fine-tuned on only 30 minutes of single speaker dataset can generate the voice of the arbitrary speaker, by pre-training on an unlabeled multi-speaker speech corpus.