The CMU_SIN database was constructed at the Language Technologies Institute at Carnegie Mellon University as a database of speech in noise designed for use in unit selection speech synthesis research. This work builds on the CMU Arctic database. This work was carried out as part of the Let's Go Project improving spoken dialog systems for non-natives and the elderly.

The distributions include a 500 utterance subset from the CMU Arctic database, recorded from one male US English speaker. There are two versions of these recordings: one where the speaking style is speech in noise (sin) and one with normal speaking style (swn). We also include two complete Festival voices built from these recordings. It is possible to build a voice from both sets of recordings that can speak both normally and in noise, though we are not distributing such a voice at this time. The distribution also contains the modifications to the FestVox voice building script prompt_them to provide a method of obtaining clean recordings of speech in noise, as well as the noise source used for these recordings.

Full phonetic labeling was performed with CMU Sphinx using the FestVox based labeling scripts. No hand-correction of the automatic labels has been done.

A description of this work is published in Creating a Database of Speech in Noise for Unit Selection Synthesis by Brian Langner and Alan W Black, at the 5th ISCA Speech Synthesis Workshop

CMU SIN Database

• US English bdl (male) sin/swn

Acknowledgements

This work is supported by the US National Science Foundation under grant number 0208835, "LET'S GO: improved speech interfaces for the general public". Any opinions, findings, and conclusions or recommendations expressed in this material are those of the authors and do not necessarily reflect the views of the National Science Foundation.