FORMATS.DOC

WAV

.wav files are speech sample files. They have an uncompressed ascii header followed by linear, 2 byte, signed integer samples that have been compressed using the "shorten" speech compression method developed by Tony Robinson at Cambridge University. The headers follow the NIST SPHERE header format (the document header.doc contains a more detailed explanation of this format). The first line specifies the header type and the second line specifies the header length. All of the files in this distribution have a header length of 1024 bytes. Below is a sample header: 

NIST_1A
   1024
sample_byte_format -s2 10
channel_count -i 1
sample_count -i 60111
sample_rate -i 8000
sample_n_bytes -i 2
sample_sig_bits -i 16
sample_max -i 2688
sample_min -i -3248
sample_coding -s26 pcm,embedded-shorten-v1.09
sample_checksum -i 65264
end_head
As per the NIST format definition, the bytes between the "end_head" line and the 1024th byte are undefined.

The "sphere" directory contains the latest release (v2.0Beta2.1) of the NIST SPHERE software package, which includes programs to read, modify and remove the SPHERE headers, and to uncompress the waveform data. These programs are supplied in the form of C source code, together with a UNIX shell script and "makefiles" to compile and install the programs. Please note that a more recent version of this package may be available via anonymous ftp directly from NIST. To obtain the package from NIST: 

	ftp jaguar.ncsl.nist.gov
	Name: anonymous
	Password: <your.email@address>
	cd pub
	dir sphere*
	binary
	get <latest_sphere_version.tar.Z>
	bye
The ftp command "dir sphere*" will list the lastest release available; supply this file name for the following "get" command. (MS-DOS users should note that the Free Software Foundation (FSF/GNU) provides DOS versions for the necessary UNIX environment utilities -- make, cc, tar, uncompress, etc -- to extract and compile the NIST sources on 80386 and 80486 machines.)

SEG

The .seg files are ascii "location and label" files. They are similar to the ".phn" files of the TIMIT database except

  1. the locations are given in a unit of time other than the sample.
  2. there is a short header saying what this unit is
Each file in this distribution has the header 
	MillisecondsPerFrame: 3.0
	END OF HEADER
After that are a series of lines, one per segment, of the form 
	<begin frame> <end frame + 1>  label
For example 
	200  237   ah
	237  289   m
The [ah] segment extends from frame 200 to frame 236 inclusive. The end label is 237 for "historical" reasons.

All the .seg files are in the "seglola" directory within the respective language directories. The total across all languages is 504. The breakup by language is: 

English         55
Farsi           48  
French          52    
German          53
Japanese        47
Korean          47 
Mandarin        51
Spanish         53 
Tamil           46 
Vietnamese      52

TOTAL          504