Improving speech synthesis quality by reducing pitch peaks in the source recordings

Proceedings of NAACL-HLT 2013, pages 502–506, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Improving speech synthesis quality by reducing pitch peaks in the source recordings Luisina Violante, Pablo Rodr´ıguez Zivic and Agust´ın Gravano Departamento de Computaci´on, FCEyN Universidad de Buenos Aires, Argentina {lviolante,prodriguez,gravano}@dc.uba.ar Abstract We present a method for improving the perceived nat- uralness of corpus-based speech synthesizers. It con- sists in removing pronounced pitch peaks in the origi- nal recordings, which typically lead to noticeable dis- continuities in the synthesized speech. We perceptu- ally evaluated this method using two concatenative and two HMM-based synthesis systems, and found that us- ing it on the source recordings managed to improve the naturalness of the synthesizers and had no effect on their intelligibility. 1 Introduction By definition, corpus-based speech synthesizers, such as concatenative and HMM-based systems, rely heavily on the quality of the speech corpus used for building the systems. Creating speech corpora for this purpose is expensive and time consuming, so when the synthesized speech obtained is not as good as expected, it may be desirable to modify or correct the corpus rather than record a new one. Common corrections are limited to discarding mispronounced words or noisy units. In this work we describe a sim- ple method for attenuating pronounced pitch peaks, a frequent problem in recordings made by profes- sional speakers, and evaluate it using four different corpus-based systems. Sections 2 and 3 describe the speech synthesis systems and corpus employed in this work. In Section 4 we present the method for reducing pitch peaks. In Section 5 we describe how we evaluated the effect of our method on intelligibil- ity and naturalness of the synthesizers. 2 Synthesis systems Festival1 is a general framework for building speech synthesis systems, written in C++ and developed by the Center of Speech Technology Research at the University of Edinburgh (Black et al., 2001). provides an implementation of concatenative speechIt synthesis as well as synthesis based on Hidden Markov Models (HMM). In this work we used a Fes- tival module called Clunits unit selection engine to build concatenative synthesizers. The unit size is the phone, although since a percentage of the previous unit is included in the acoustic distance measure, the unit size is rather “phone plus previous phone”, thus similar to a diphone (Black and Lenzo, 2007). Ad- ditionally, we used a second Festival module called Clustergen parametric synthesis engine for building HMM-based speech synthesizers. MARY TTS2 is an open-source synthesis plat- form written in Java, originally jointly developed by the Language Technology Lab at the German Research Center for Artificial Intelligence (DFKI) and the Institute of Phonetics at Saarland Univer- sity, and currently maintained by DFKI. Like Fes- tival, MARY provides toolkits for building unit se- lection and HMM-based synthesis voices (Schr¨oder and Trouvain, 2003). 3 Corpus For building our systems we used the SECYT cor- pus, created by the Laboratorio de Investigacio- nes Sensoriales (Universidad de Buenos Aires) for 1http://festvox.org/festival 2http://mary.dfki.de 502 studying the prosody of Argentine …

Content truncated to 500 words