TTS for Low Resource Languages: A Bangla Synthesizer

TTS for Low Resource Languages: A Bangla Synthesizer Alexander Gutkin, Linne Ha, Martin Jansche, Knot Pipatsrisawat, Richard Sproat Google, Inc. 1600 Amphitheatre Parkway, Mountain View, CA {agutkin,linne,mjansche,thammaknot,rws}@google.com Abstract We present a text-to-speech (TTS) system designed for the dialect of Bengali spoken in Bangladesh. This work is part of an ongoing effort to address the needs of under-resourced languages. We propose a process for streamlining the bootstrapping of TTS systems for under-resourced languages. First, we use crowdsourcing to collect the data from multiple ordinary speakers, each speaker recording small amount of sentences. Second, we leverage an existing text normalization system for a related language (Hindi) to bootstrap a linguistic front-end for Bangla. Third, we employ statistical techniques to construct multi-speaker acoustic models using Long Short-Term Memory Recurrent Neural Network (LSTM-RNN) and Hidden Markov Model (HMM) approaches. We then describe our experiments that show that the resulting TTS voices score well in terms of their perceived quality as measured by Mean Opinion Score (MOS) evaluations. Keywords: Text-to-Speech, Under-resourced languages, Bangla 1. Introduction Developing a text-to-speech (TTS) system is a major in- vestment of effort. For the best concatenative unit-selection systems (Hunt and Black, 1996), many hours of recording are typical, and one needs to invest in careful lexicon devel- opment, and complex rules for text normalization, among other things. All of this requires resources, as well as cura- tion from native-speaker linguists. For low-resource languages it is often hard to find relevant resources, so there has been much recent work on methods for developing systems using minimal data (Sitaram et al., 2013). The downside of these approaches is that the quality of the resulting systems can be low and it is doubtful people would want to use them. We are therefore interested in approaches that minimize ef- fort, but still produce systems that are acceptable to users. This paper describes our development of a system for Bangla, the main language of Bangladesh and a major lan- guage of India, and in particular the speech, lexicon and text normalization resources, all of which we are planning to release, under a liberal open-source license. A core idea is the use of multiple ordinary speakers, rather than a single professional speaker (the normal approach). There are two main justifications. First, voice talents are expensive, so it is more cost-effective to record ordinary people; but these quickly get tired reading aloud, limiting how much they can read. We thus need multiple speakers for an adequate database. Second, there is an added benefit of privacy: we can create a natural-sounding voice that is not identifiable as a specific individual. Unit selection (Hunt and Black, 1996) is a dominant ap- proach to speech synthesis, but it is not suitable when work- ing with multiple speakers, one obvious reason being that the system will often adjoin units from different speakers, resulting in very unnatural output. Instead we adopt a sta- tistical parametric approach (Zen et al., 2009). In statis- tical parametric synthesis the training stage uses multiple speaker data …

Content truncated to 500 words