This paper discusses techniques for determining the linguistic needs for open-domain synthesis by concatenative methods, and reports on the design and evaluation of a tool for collecting and balancing a speech corpus automatically, in order to ensure optimal coverage of the sounds required for synthesis within a given task-domain. Syntheticallygenerated utterances are used to prompt speakers, and in-line acoustic analysis determines the prosodic as well as phonemic balance of the resulting speech during recording, re-prompting the speaker with textually modi ed versions if necessary, to elicit the desired articulation sequences. The closed-loop process, which incorporates human self-correction and evaluation, allows for more e cient collection of a balanced corpus for concatenative speech synthesis.
[1]
Y. Sagisaka,et al.
Speech synthesis by rule using an optimal selection of non-uniform synthesis units
,
1988,
ICASSP-88., International Conference on Acoustics, Speech, and Signal Processing.
[2]
Nick Campbell.
Talking Machines for Information Access
,
1999
.
[3]
Joseph Olive,et al.
A scheme for concatenating units for speech synthesis
,
1980,
ICASSP.
[4]
Frédéric Bimbot,et al.
Introducing statistical dependencies and structural constraints in variable-length sequence models
,
1996,
ICGI.
[5]
Tomohisa Hirokawa.
Speech synthesis using a waveform dictionary
,
1989,
EUROSPEECH.
[6]
Jonathan Allen,et al.
Text to Speech
,
2015
.