scholarly journals End-to-End Text-to-Speech for Low-Resource Languages by Cross-Lingual Transfer Learning

2019 ◽  
Author(s):  
Yuan-Jui Chen ◽  
Tao Tu ◽  
Cheng-chieh Yeh ◽  
Hung-Yi Lee
Keyword(s):  
Transfer Learning ◽  
Text To Speech ◽  
Low Resource ◽  
End To End ◽  
Cross Lingual

scholarly journals Text-to-speech system for low-resource language using cross-lingual transfer learning and data augmentation

2021 ◽  
Vol 2021 (1) ◽  
Author(s):  
Zolzaya Byambadorj ◽  
Ryota Nishimura ◽  
Altangerel Ayush ◽  
Kengo Ohta ◽  
Norihide Kitaoka
Keyword(s):  
Transfer Learning ◽  
Data Augmentation ◽  
Prediction Models ◽  
Target Language ◽  
Text To Speech ◽  
Paired Data ◽  
Low Resource ◽  
Language Data ◽  
Single Speaker ◽  
Cross Lingual

AbstractDeep learning techniques are currently being applied in automated text-to-speech (TTS) systems, resulting in significant improvements in performance. However, these methods require large amounts of text-speech paired data for model training, and collecting this data is costly. Therefore, in this paper, we propose a single-speaker TTS system containing both a spectrogram prediction network and a neural vocoder for the target language, using only 30 min of target language text-speech paired data for training. We evaluate three approaches for training the spectrogram prediction models of our TTS system, which produce mel-spectrograms from the input phoneme sequence: (1) cross-lingual transfer learning, (2) data augmentation, and (3) a combination of the previous two methods. In the cross-lingual transfer learning method, we used two high-resource language datasets, English (24 h) and Japanese (10 h). We also used 30 min of target language data for training in all three approaches, and for generating the augmented data used for training in methods 2 and 3. We found that using both cross-lingual transfer learning and augmented data during training resulted in the most natural synthesized target speech output. We also compare single-speaker and multi-speaker training methods, using sequential and simultaneous training, respectively. The multi-speaker models were found to be more effective for constructing a single-speaker, low-resource TTS model. In addition, we trained two Parallel WaveGAN (PWG) neural vocoders, one using 13 h of our augmented data with 30 min of target language data and one using the entire 12 h of the original target language dataset. Our subjective AB preference test indicated that the neural vocoder trained with augmented data achieved almost the same perceived speech quality as the vocoder trained with the entire target language dataset. Overall, we found that our proposed TTS system consisting of a spectrogram prediction network and a PWG neural vocoder was able to achieve reasonable performance using only 30 min of target language training data. We also found that by using 3 h of target language data, for training the model and for generating augmented data, our proposed TTS model was able to achieve performance very similar to that of the baseline model, which was trained with 12 h of target language data.

scholarly journals Cross-Language End-to-End Speech Recognition Research Based on Transfer Learning for the Low-Resource Tujia Language

Symmetry ◽  
2019 ◽  
Vol 11 (2) ◽  
pp. 179 ◽  
Author(s):  
Chongchong Yu ◽  
Yunbing Chen ◽  
Yueqiao Li ◽  
Meng Kang ◽  
Shixuan Xu ◽  
...  
Keyword(s):  
Speech Recognition ◽  
Transfer Learning ◽  
Short Term Memory ◽  
Recognition System ◽  
Language Recognition ◽  
Low Resource ◽  
End To End ◽  
The Cross ◽  
Hidden Layer ◽  
Cross Language

To rescue and preserve an endangered language, this paper studied an end-to-end speech recognition model based on sample transfer learning for the low-resource Tujia language. From the perspective of the Tujia language international phonetic alphabet (IPA) label layer, using Chinese corpus as an extension of the Tujia language can effectively solve the problem of an insufficient corpus in the Tujia language, constructing a cross-language corpus and an IPA dictionary that is unified between the Chinese and Tujia languages. The convolutional neural network (CNN) and bi-directional long short-term memory (BiLSTM) network were used to extract the cross-language acoustic features and train shared hidden layer weights for the Tujia language and Chinese phonetic corpus. In addition, the automatic speech recognition function of the Tujia language was realized using the end-to-end method that consists of symmetric encoding and decoding. Furthermore, transfer learning was used to establish the model of the cross-language end-to-end Tujia language recognition system. The experimental results showed that the recognition error rate of the proposed model is 46.19%, which is 2.11% lower than the that of the model that only used the Tujia language data for training. Therefore, this approach is feasible and effective.

scholarly journals ManyModalQA: Modality Disambiguation and QA over Diverse Inputs

2020 ◽  
Vol 34 (05) ◽  
pp. 7879-7886
Author(s):  
Darryl Hannan ◽  
Akshay Jain ◽  
Mohit Bansal
Keyword(s):  
Transfer Learning ◽  
Human Performance ◽  
Question Answering ◽  
State Of The Art ◽  
Fine Tuning ◽  
Low Resource ◽  
Lower Accuracy ◽  
End To End ◽  
Text Images ◽  
Training Sets

We present a new multimodal question answering challenge, ManyModalQA, in which an agent must answer a question by considering three distinct modalities: text, images, and tables. We collect our data by scraping Wikipedia and then utilize crowdsourcing to collect question-answer pairs. Our questions are ambiguous, in that the modality that contains the answer is not easily determined based solely upon the question. To demonstrate this ambiguity, we construct a modality selector (or disambiguator) network, and this model gets substantially lower accuracy on our challenge set, compared to existing datasets, indicating that our questions are more ambiguous. By analyzing this model, we investigate which words in the question are indicative of the modality. Next, we construct a simple baseline ManyModalQA model, which, based on the prediction from the modality selector, fires a corresponding pre-trained state-of-the-art unimodal QA model. We focus on providing the community with a new manymodal evaluation set and only provide a fine-tuning set, with the expectation that existing datasets and approaches will be transferred for most of the training, to encourage low-resource generalization without large, monolithic training sets for each new task. There is a significant gap between our baseline models and human performance; therefore, we hope that this challenge encourages research in end-to-end modality disambiguation and multimodal QA models, as well as transfer learning.

scholarly journals Multilingual Offensive Language Identification for Low-resource Languages

2022 ◽  
Vol 21 (1) ◽  
pp. 1-13
Author(s):  
Tharindu Ranasinghe ◽  
Marcos Zampieri
Keyword(s):  
Transfer Learning ◽  
Hate Speech ◽  
Training And Development ◽  
Language Identification ◽  
Shared Task ◽  
Low Resource ◽  
Government Organizations ◽  
Cross Lingual ◽  
Offensive Language ◽  
Clear Majority

Offensive content is pervasive in social media and a reason for concern to companies and government organizations. Several studies have been recently published investigating methods to detect the various forms of such content (e.g., hate speech, cyberbullying, and cyberaggression). The clear majority of these studies deal with English partially because most annotated datasets available contain English data. In this article, we take advantage of available English datasets by applying cross-lingual contextual word embeddings and transfer learning to make predictions in low-resource languages. We project predictions on comparable data in Arabic, Bengali, Danish, Greek, Hindi, Spanish, and Turkish. We report results of 0.8415 F1 macro for Bengali in TRAC-2 shared task [23], 0.8532 F1 macro for Danish and 0.8701 F1 macro for Greek in OffensEval 2020 [58], 0.8568 F1 macro for Hindi in HASOC 2019 shared task [27], and 0.7513 F1 macro for Spanish in in SemEval-2019 Task 5 (HatEval) [7], showing that our approach compares favorably to the best systems submitted to recent shared tasks on these three languages. Additionally, we report competitive performance on Arabic and Turkish using the training and development sets of OffensEval 2020 shared task. The results for all languages confirm the robustness of cross-lingual contextual embeddings and transfer learning for this task.

scholarly journals Machine Translation in Low-Resource Languages by an Adversarial Neural Network

2021 ◽  
Vol 11 (22) ◽  
pp. 10860
Author(s):  
Mengtao Sun ◽  
Hao Wang ◽  
Mark Pasquine ◽  
Ibrahim A. Hameed
Keyword(s):  
Machine Translation ◽  
Transfer Learning ◽  
Grammatical Structure ◽  
Neural Machine Translation ◽  
Low Resource ◽  
High Resource ◽  
Learning Techniques ◽  
Good Potential ◽  
Target Languages ◽  
Cross Lingual

Existing Sequence-to-Sequence (Seq2Seq) Neural Machine Translation (NMT) shows strong capability with High-Resource Languages (HRLs). However, this approach poses serious challenges when processing Low-Resource Languages (LRLs), because the model expression is limited by the training scale of parallel sentence pairs. This study utilizes adversary and transfer learning techniques to mitigate the lack of sentence pairs in LRL corpora. We propose a new Low resource, Adversarial, Cross-lingual (LAC) model for NMT. In terms of the adversary technique, LAC model consists of a generator and discriminator. The generator is a Seq2Seq model that produces the translations from source to target languages, while the discriminator measures the gap between machine and human translations. In addition, we introduce transfer learning on LAC model to help capture the features in rare resources because some languages share the same subject-verb-object grammatical structure. Rather than using the entire pretrained LAC model, we separately utilize the pretrained generator and discriminator. The pretrained discriminator exhibited better performance in all experiments. Experimental results demonstrate that the LAC model achieves higher Bilingual Evaluation Understudy (BLEU) scores and has good potential to augment LRL translations.

Sign in / Sign up

Export Citation Format

Share Document