The YOHO Speaker Verification Corpus supports development, training and testing of speaker verification systems that use limited vocabulary, free-text input. The particular vocabulary employed in this collection consists of two-digit numbers ("thirty-four", "sixty-one", etc), spoken continuously in sets of three (e.g. "36-45-89").
This release of YOHO has been designed, with regard to the quantity and organization of data, to answer the following question: does a speaker verification system perform at 0.01% False-Accept and 0.1% False-Reject at 75% confidence with a 50% probability of passing the test? There are 138 speakers (108 male, 30 female); for each speaker, there are 4 enrollment sessions of 24 utterances each, and 10 verification sessions of four utterences each, for a total of 136 utterances in 14 sessions per speaker.
The sample rate for the speech files is 8 kHz, and the sample coding is 12-bit linear (stored as 16-bit words). For publication on CD-ROM, each waveform file has been given an appropriate SPHERE file header, and has been compressed using the "shorten" waveform compression algorithm developed by Tony Robinson at Cambridge University. The SPHERE implementation of the "shorten" decompression process (see the program "w_decode" in the SPHERE package) allows the output to be written in either "big-endian" or "little-endian" format -- that is, either high-byte or low-byte first in the 16-bit word.
Use and availability of SPHERE
The SPHERE software package, developed by the speech group at the National Institute of Standards and Technology (NIST) was used to incorporate the SPHERE headers and apply waveform compression. It can be used to access and modify header information, remove headers from the files, and uncompress (and recompress) the waveform data. For convenience, we have included on this disc (in the "sphere" directory) the most recent release of the SPHERE package as of the time of this publication (version 2.1 Beta; this was the release used in preparing the corpus for publication.) Users are advised that NIST is maintaining the package, and updating or correcting the software as needed; when a more recent release becomes available, it can be obtained for free directly from NIST, by means of anonymous ftp, via the ftp host "jaguar.ncsl.nist.gov", in the "pub" directory.
Organization of the corpus
The corpus is divided into "enrollment" and "verification" segments (in the "enroll" and "verify" diorectories"); each segment contains data from all 138 speakers, who are designated by 3-digit numbers (these are the directory names contained under "enroll" and "verify"). There are four enrollement sessions per speaker, numbered 1 through 4, and each session contains 24 utterances; thus, each speaker directory under "enroll" contains subdirectories "1", "2", "3" and "4", and each of these contains 24 waveform files. The name of each waveform file indicates the "text" of the utterance in that file; for instance, "62_31_53.wav" contains the phrase "sixty-two thirty-one fifty-three". Each speaker directory under "verify" contains 10 sessions, numbered 1 through 10, and each session contains four utterances; again, the file name for each utterance indicates the prompting text.
Other documentation
The file "yoho_db.doc" contains portions of larger reports by Joseph Campbell, Al Higgins, and others, that describe the creation of YOHO in greater detail.
The file "speaker.doc" contains a list of the 138 speaker designations (the 3-digit numbers), together with the speaker's gender ("M" or "F") and, in most cases, the speaker's geographic origin.