Natural-sounding speech synthesis using variable-length units

Abstract

The goal of this work was to develop a speech synthesis system which concatenates variable-length units to create natural sounding speech. Our initial work in this area showed that by careful design of system responses to ensure consistent intonation contours, natural-sounding speech synthesis was achievable with wordand phraselevel concatenation. In order to extend the flexibility of this framework, subsequent work focused on the problem of generating novel words from a pre-recorded corpus of sub-word units. The design of the sub-word units was motivated by perceptual experiments that investigated where speech could be spliced with minimal distortion and what contextual constraints were necessary to maintain in order to produce natural sounding speech. This sub-word corpus is then searched at synthesis time with a Viterbi search which selects a sequence of units based on how well they individually match the input specification and on how well they sound as an ensemble. This concatenative speech synthesis system, ENVOICE, has been used in a conversational information retrieval system in two application domains to convert meaning representations into speech waveforms.

Automatic construction of a prosodically rich text corpus for speech synthesis systems

“…The past research revealed that more natural-sounding speech is obtained if prosodic information is included in unit selection [7]. However, signal processing techniques employed to do prosodic modifications were reported to reduce the quality of synthesized speech [1,8,29].…”

“…Past research has shown that natural-sounding synthetic speech can be produced by selecting non-uniform units (i.e. units of variable length) from large speech databases [1,2,6,29,17,20]. Studies [2,6,29,25] indicate that the naturalness of synthetic speech can be improved by excising longer sequences of recorded speech from the database, this reduces the number of concatenation points in a synthesized utterance.…”

“…units of variable length) from large speech databases [1,2,6,29,17,20]. Studies [2,6,29,25] indicate that the naturalness of synthetic speech can be improved by excising longer sequences of recorded speech from the database, this reduces the number of concatenation points in a synthesized utterance. It was argued by [26] that longer chunks of recorded speech preserve natural rhythm and prosody better than shorter sequences.…”

Voice transformations: from speech synthesis to mammalian vocalizations

“…The original synthesis corpus [6] was converted by lowering the fundamental frequency by 30% and the spectrum (formants) by 25%. Thus, the spectral envelope was interpolated by a factor of ¦ ¦ )( 10 )2 ¤ ( 3 .…”

CU VOCAL: corpus-based syllable concatenation for Chinese speech synthesis across domains and dialects

“…We focus on Cantonese, a major Chinese dialect predominant in Hong Kong, South China and many overseas Chinese communities. The corpus-based concatenation technique has been gaining popularity in speech synthesis [2][3][4][5][6] due to its ability to achieve a high degree of naturalness. The use of corpus-based syllable concatenation is particularly suitable for Chinese, since the language is monosyllabic in nature.…”