Existing singing voice synthesis models (SVS) are usually trained on singing data and depend on either error-prone time-alignment and duration features or explicit music score information. In this paper, we propose Karaoker, a multispeaker Tacotron-based model conditioned on voice characteristic features that is trained exclusively on spoken data without requiring time-alignments. Karaoker synthesizes singing voice following a multi-dimensional template extracted from a source waveform of an unseen speaker/singer. The model is jointly conditioned with a single deep convolutional encoder on continuous data including pitch, intensity, harmonicity, formants, cepstral peak prominence and octaves. We extend the text-to-speech training objective with feature reconstruction, classification and speaker identification tasks that guide the model to an accurate result. Except for multi-tasking, we also employ a Wasserstein GAN training scheme as well as new losses on the acoustic model's output to further refine the quality of the model.
Existing Chinese text error detection mainly focuses on spelling and simple grammatical errors. These errors have been studied extensively and are relatively simple for humans. On the contrary, Chinese semantic errors are understudied and more complex that humans cannot easily recognize. The task of this paper is Chinese Semantic Error Recognition (CSER), a binary classification task to determine whether a sentence contains semantic errors. The current research has no effective method to solve this task. In this paper, we inherit the model structure of BERT and design several syntax-related pre-training tasks so that the model can learn syntactic knowledge. Our pre-training tasks consider both the directionality of the dependency structure and the diversity of the dependency relationship. Due to the lack of a published dataset for CSER, we build a high-quality dataset for CSER for the first time named Corpus of Chinese Linguistic Semantic Acceptability (CoCLSA). The experimental results on the CoCLSA show that our methods outperform universal pre-trained models and syntax-infused models.
Causal inference is the process of capturing cause-effect relationship among variables. Most existing works focus on dealing with structured data, while mining causal relationship among factors from unstructured data, like text, has been less examined, but is of great importance, especially in the legal domain. In this paper, we propose a novel Graph-based Causal Inference (GCI) framework, which builds causal graphs from fact descriptions without much human involvement and enables causal inference to facilitate legal practitioners to make proper decisions. We evaluate the framework on a challenging similar charge disambiguation task. Experimental results show that GCI can capture the nuance from fact descriptions among multiple confusing charges and provide explainable discrimination, especially in few-shot settings. We also observe that the causal knowledge contained in GCI can be effectively injected into powerful neural networks for better performance and interpretability.
The Latent Dirichlet Allocation (LDA) model is a popular method for creating mixed-membership clusters. Despite having been originally developed for text analysis, LDA has been used for a wide range of other applications. We propose a new formulation for the LDA model which incorporates covariates. In this model, a negative binomial regression is embedded within LDA, enabling straight-forward interpretation of the regression coefficients and the analysis of the quantity of cluster-specific elements in each sampling units (instead of the analysis being focused on modeling the proportion of each cluster, as in Structural Topic Models). We use slice sampling within a Gibbs sampling algorithm to estimate model parameters. We rely on simulations to show how our algorithm is able to successfully retrieve the true parameter values and the ability to make predictions for the abundance matrix using the information given by the covariates. The model is illustrated using real data sets from three different areas: text-mining of Coronavirus articles, analysis of grocery shopping baskets, and ecology of tree species on Barro Colorado Island (Panama). This model allows the identification of mixed-membership clusters in discrete data and provides inference on the relationship between covariates and the abundance of these clusters.
Speech synthesis and music audio generation from symbolic input differ in many aspects but share some similarities. In this study, we investigate how text-to-speech synthesis techniques can be used for piano MIDI-to-audio synthesis tasks. Our investigation includes Tacotron and neural source-filter waveform models as the basic components, with which we build MIDI-to-audio synthesis systems in similar ways to TTS frameworks. We also include reference systems using conventional sound modeling techniques such as sample-based and physical-modeling-based methods. The subjective experimental results demonstrate that the investigated TTS components can be applied to piano MIDI-to-audio synthesis with minor modifications. The results also reveal the performance bottleneck -- while the waveform model can synthesize high quality piano sound given natural acoustic features, the conversion from MIDI to acoustic features is challenging. The full MIDI-to-audio synthesis system is still inferior to the sample-based or physical-modeling-based approaches, but we encourage TTS researchers to test their TTS models for this new task and improve the performance.
Recent advances in sophisticated synthetic speech generated from text-to-speech (TTS) or voice conversion (VC) systems cause threats to the existing automatic speaker verification (ASV) systems. Since such synthetic speech is generated from diverse algorithms, generalization ability with using limited training data is indispensable for a robust anti-spoofing system. In this work, we propose a transfer learning scheme based on the wav2vec 2.0 pretrained model with variational information bottleneck (VIB) for speech anti-spoofing task. Evaluation on the ASVspoof 2019 logical access (LA) database shows that our method improves the performance of distinguishing unseen spoofed and genuine speech, outperforming current state-of-the-art anti-spoofing systems. Furthermore, we show that the proposed system improves performance in low-resource and cross-dataset settings of anti-spoofing task significantly, demonstrating that our system is also robust in terms of data size and data distribution.
The goal of voice conversion (VC) is to convert input voice to match the target speaker's voice while keeping text and prosody intact. VC is usually used in entertainment and speaking-aid systems, as well as applied for speech data generation and augmentation. The development of any-to-any VC systems, which are capable of generating voices unseen during model training, is of particular interest to both researchers and the industry. Despite recent progress, any-to-any conversion quality is still inferior to natural speech. In this work, we propose a new any-to-any voice conversion pipeline. Our approach uses automated speech recognition (ASR) features, pitch tracking, and a state-of-the-art waveform prediction model. According to multiple subjective and objective evaluations, our method outperforms modern baselines in terms of voice quality, similarity and consistency.
We present a detector for curved text in natural images. We model scene text instances as tubes around their medial axes and introduce a parametrization-invariant loss function. We train a two-stage curved text detector, and evaluate it on the curved text benchmarks CTW-1500 and Total-Text. Our approach achieves state-of-the-art results or improves upon them, notably for CTW-1500 by over 8 percentage points in F-score.
The mental disorder of online users is determined using social media posts. The major challenge in this domain is to avail the ethical clearance for using the user generated text on social media platforms. Academic re searchers identified the problem of insufficient and unlabeled data for mental health classification. To handle this issue, we have studied the effect of data augmentation techniques on domain specific user generated text for mental health classification. Among the existing well established data augmentation techniques, we have identified Easy Data Augmentation (EDA), conditional BERT, and Back Translation (BT) as the potential techniques for generating additional text to improve the performance of classifiers. Further, three different classifiers Random Forest (RF), Support Vector Machine (SVM) and Logistic Regression (LR) are employed for analyzing the impact of data augmentation on two publicly available social media datasets. The experiments mental results show significant improvements in classifiers performance when trained on the augmented data.
This paper presents our proposed methods to ICDAR 2021 Robust Reading Challenge - Integrated Circuit Text Spotting and Aesthetic Assessment (ICDAR RRC-ICTEXT 2021). For the text spotting task, we detect the characters on integrated circuit and classify them based on yolov5 detection model. We balance the lowercase and non-lowercase by using SynthText, generated data and data sampler. We adopt semi-supervised algorithm and distillation to furtherly improve the model's accuracy. For the aesthetic assessment task, we add a classification branch of 3 classes to differentiate the aesthetic classes of each character. Finally, we make model deployment to accelerate inference speed and reduce memory consumption based on NVIDIA Tensorrt. Our methods achieve 59.1 mAP on task 3.1 with 31 FPS and 306M memory (rank 1), 78.7\% F2 score on task 3.2 with 30 FPS and 306M memory (rank 1).