German and Czech Speech Synthesis Using HMM-Based Speech Segment Database

Abstract

Abstract. This paper presents an experimental German speech synthesis system. As in case of a Czech text-to-speech system ARTIC, statistical approach (using hidden Markov models) was employed to build a speech segment database. This approach was confirmed to be language independent and it was shown to be capable of designing a quality database that led to an intelligible synthetic speech of a high quality. Some experiments with clustering the similar speech contexts were performed to enhance the quality of the synthetic speech. Our results show the superiority of phoneme-level clustering to subphoneme-level one.

Slovak Text-to-Speech Synthesis in ARTIC System

“…A set of three-state left-to-right single-density crossword-triphone hidden Markov models (HMMs) was employed to model context-dependent phone-sized units (triphones) on the basis of the large speech corpus [1,[3][4][5]. The speech corpus consisted of both linguistic and speech data.…”

“…Clustering can be performed at two different levels: model (phone) level or state (sub-phone) level. Listening tests revealed the superiority of the model-level clustering over the state-level one [3]. Viterbi algorithm is then used to align speech waveforms of each sentence from the speech corpus with a sequence of corresponding tied-triphone HMMs.…”

“…German [3]). Phonetic forms of both languages are very similar to their orthographic forms (in fact, this feature is common to all Slavic languages).…”

“…Having been created on the principles mentioned above, it is capable of using different automatically built acoustic unit inventories. An important step towards multilinguality was achieved in [3] where a German voice was successfully designed within the ARTIC system. In this paper another language, Slovak, is modeled within the framework of ARTIC.…”

“…In [3] a German TTS module was successfully created within the ARTIC system. It should be emphasized that all voices (or language modules) were realized fully automatically in the same way as described further in the paper (with respect to the their language-dependent properties).…”