Hindi-English speech-to-speech translation system for travel expressions

Speech-to-speech translation system enables in translation of speech signals in a source language A to target language B. A good speech-to-speech translation (S2ST) system can be characterized by its ability to keep intact the fluency and meaning of the original speech input. An S2ST system to enable translation between Hindi and English is the main idea of the proposed work. A preliminary dataset concentrating on basic travel expressions in both the languages considered is used for this work. In order to develop a successful S2ST system three subsystems are required namely, automatic speech recognition (ASR) system, machine translation (MT) system and text-to-speech synthesis (TTS) system. Hidden Markov models based ASR system is developed for both the languages and their performances are analyzed based on the word error rate (WER). The MT subsystem makes use of the statistical machine translation (SMT) approach for the purpose of translating the text between the two languages involved. The SMT makes use of IBM alignment models and language models to enable proper translation. The performance of MT is analyzed based on translated edit rate (TER) and analysis of the translation table. HMM-based speech synthesis system (HTS) is used to synthesize the translated text. Performance of the synthesizer is analyzed based on mean opinion score (MOS) from a group of listeners.