No abstract

Building text-to-speech (TTS) synthesisers is a difficult task, especially for low resource languages. Language-specific modules need to be developed for system building. End-to-end speech synthesis has become a popular paradigm as a TTS can be trained using only pairs. However, end-to-end speech synthesis is not scalable in a multilanguage scenario, as the vocabulary increases with the number of different scripts. In this paper, TTSes are trained for Indian languages using two text representations-character-based and phone-based. For the character-based approach, a multi-language character map (MLCM) is proposed to easily train Indic speech synthesisers. The phone-based approach uses the common label set (CLS) representation for Indian languages. Both approaches leverage the similarities that exist among the languages. The advantage is a compact representation across multiple languages. Experiments are conducted by building TTSes using monolingual data and by pooling data across two languages. The ability to synthesise code-mixed text using the phone-based approach is also assessed. Subjective evaluations indicate that reasonably good quality Indic TTSes can be developed using both approaches. This emphasises the need to incorporate multilingual text processing in the end-to-end framework.

Most Indians are inherently bilingual or multilingual owing to the diverse linguistic culture in India. As a result, code-switching is quite common in conversational speech. The objective of this work is to train good quality text-to-speech (TTS) synthesisers that can seamlessly handle code-switching. To achieve this, bilingual TTSes that are capable of handling phonotactic variations across languages are trained using combinations of monolingual data in a unified framework. In addition to segmenting Indic speech data using signal processing cues in tandem with hidden Markov model-deep neural network (HMM-DNN), we propose to segment Indian English data using the same approach after NIST syllabification. Then, bilingual HTS-STRAIGHT based systems are trained by randomizing the order of data so that the systematic interactions between the two languages are captured better. Experiments are conducted by considering three language pairs: Hindi+English, Tamil+English and Hindi+Tamil. The code-switched systems are evaluated on monolingual, code-mixed and code-switched texts. Degradation mean opinion score (DMOS) for monolingual sentences shows marginal degradation over that of an equivalent monolingual TTS system, while the DMOS for bilingual sentences is significantly better than that of the corresponding monolingual TTS systems.

No abstract