Text-to-speech synthesis is a key component of interactive, speech-based systems. Typically, building a high-quality voice requires collecting dozens of hours of speech from a single professional speaker in an anechoic chamber with a high-quality microphone. There are about 7,000 languages spoken in the world, and most do not enjoy the speech research attention historically paid to such languages as English, Spanish, Mandarin, and Japanese. Speakers of these so-called "low-resource languages" therefore do not equally benefit from these technological advances. While it takes a great deal of time and resources to collect a traditional text-to-speech corpus for a given language, we may instead be able to make use of various sources of "found'' data which may be available. In particular, sources such as radio broadcast news and ASR corpora are available for many languages. While this kind of data does not exactly match what one would collect for a more standard TTS corpus, it may nevertheless contain parts which are usable for producing natural and intelligible parametric TTS voices. In the first part of this thesis, we examine various types of found speech data in comparison with data collected for TTS, in terms of a variety of acoustic and prosodic features. We find that radio broadcast news in particular is a good match. Audiobooks may also be a good match despite their largely more expressive style, and certain speakers in conversational and read ASR corpora also resemble TTS speakers in their manner of speaking and thus their data may be usable for training TTS voices. In the rest of the thesis, we conduct a variety of experiments in training voices on non-traditional sources of data, such as ASR data, radio broadcast news, and audiobooks. We aim to discover which methods produce the most intelligible and natural-sounding voices, focusing on three main approaches: 1) Training data subset selection. In noisy, heterogeneous data sources, we may wish to locate subsets of the data that are well-suited for building voices, based on acoustic and prosodic features that are known to correspond with TTS-style speech, while excluding utterances that introduce noise or other artifacts. We find that choosing subsets of speakers for training data can result in voices that are more intelligible. 2) Augmenting the frontend feature set with new features. In cleaner sources of found data, we may wish to train voices on all of the data, but we may get improvements in naturalness by including acoustic and prosodic features at the frontend and synthesizing in a manner that better matches the TTS style. We find that this approach is promising for creating more natural-sounding voices, regardless of the underlying acoustic model. 3) Adaptation. Another way to make use of high-quality data while also including informative acoustic and prosodic features is to adapt to subsets, rather than to select and train only on subsets. We also experiment with training on mixed high- and low-quality data, and adapting towards the high-quality set, which produces more intelligible voices than training on either type of data by itself. We hope that our findings may serve as guidelines for anyone wishing to build their own TTS voice using non-traditional sources of found data.

This thesis focuses on the problem of using text-to-speech (TTS) to synthesize speech with natural-sounding prosody. I propose a two-step process for approaching this problem. In the first step, I train text-based models to predict the locations of phrase boundaries and pitch accents in an utterance. Because these models use only text features, they can be used to predict the locations of prosodic events in novel utterances. In the second step, I incorporate these prosodic events into a text-to-speech pipeline in order to produce prosodically appropriate speech. I trained models for predicting phrase boundaries and pitch accents on utterances from a corpus of radio news data. I found that the strongest models used a large variety of features, including syntactic features, lexical features, word embeddings, and co-reference features. In particular, using a large variety of syntactic features improved performance on both tasks. These models also performed well when tested on a different corpus of news data. I then trained similar models on two conversational corpora: one a corpus of task-oriented dialogs and one a corpus of open-ended conversations. I again found that I could train strong models by using a wide variety of linguistic features, although performance dropped slightly in cross-corpus applications, and performance was very poor in cross-genre applications. For conversational speech, syntactic features continued to be helpful for both tasks. Additionally, word embedding features were particularly helpful in the conversational domain. Interestingly, while it is generally believed that given information (i.e., terms that have recently been referenced) is often de-accented, for all three corpora, I found that including co-reference features only slightly improved the pitch accent detection model. I then trained a TTS system on the same radio news corpus using Merlin, an open source DNN-based toolkit for TTS. As Merlin includes a linguistic feature extraction step before training, I added two additional features: one for phrase boundaries (distinguishing between sentence boundaries and mid-sentence phrase boundaries) and one for pitch accents. The locations of all breaks and accents for all test and training data were determined using the text-based prosody prediction models. I found that the pipeline using these new features produced speech that slightly outperformed the baseline on objective metrics such as mel-cepstral distortion (MCD) and was greatly preferred by listeners in a subjective listening test. Finally, I trained an end-to-end TTS system on data that included phrase boundaries. The model was trained on a corpus of read speech, with the locations of phrase boundaries predicted based on acoustic features, and tested on radio news stories, with phrase boundaries predicted using the text-based model. I found that including phrase boundaries lowered MCD between the synthesized speech and the original radio broadcast, as compared to the baseline, but the results of a listening test were inconclusive.

No description available.