Wavfile, wavefile or soundfiles are names for the files containing the original waveform of the speech or sound data. At some level, this is just a series of numbers corresponding to the voltage measured (or 'sampled') across a microphone at a regular, short interval (the sampling period). When converted back to a voltage and fed to a speaker, the original sound is played back. The majority of soundfiles we deal with store samples as 16 bit signed integers, sampled at either 8 or 16 kHz (i.e. a sampling interval of 125us or 62.5us). The differences between the different file formats are merely in the header, i.e. the extra information that may be included to describe the format of the bulk of the file.
For information about how to view and listen to wavfiles, see sndplay, xwaves, and wavesurfer.tcl under the "What are the principal programs used to do speech recognition research at ICSI?" FAQ question.
Soundfiles can vary across the following dimensions. Often, this information (or part of it) is stored in the header (depending on the format):
Here are the most common soundfile formats encountered at ICSI;
| Format | SNDF tag | Filename | Description |
| NIST/Sphere | NIST | *.wav, *.nist, *.sph | Headered soundfile format defined by NIST and used in many of the LDC and other standard speech databases. Can contain compressed data, for instance by holding embedded shorten data. Header format is mainly ascii and thus easily extensible, with the result that previously-unknown formats sometimes appear and break the sndutils. |
| Microsoft WAVE | MSWAVE | *.wav | Microsoft's basic sampled sound format, used in Windows. |
| AIFF | AIFF | *.aif, *.aiff | Audio Interchange File Format, a fully-fledged file format able to hold sampled waveforms plus ancilliary data. Popular on the Mac and SGI platforms. |
| Raw PCM | PCM | *.raw, *.pcm | Headerless sound samples, usually 16 bit shorts, byte order unspecified. |
| ESPS | ESPS | *.sd | The wavefile format used by the xwaves/ESPS packages. This is a pretty complicated header that I haven't fully decoded, and is probably best avoided (xwaves now accepts NIST format as an alternative). Unfortunately, some of the databases at ICSI (e.g. BeRP) have wavfiles still in this format. |
| Shorten | - | *.shn | Shorten is not a wavfile format per se, but a compression algorithm specifically designed for audio data. It is implemented by Tony Robinson's shorten program, which is invoked transparently by the sndutil programs (sndcat, sndplay, etc.) for any filename ending .shn. Shorten compression is usually lossless (this is an option at compress time) and typically saves 50-80% of the original file size, depending on the amount of energy in the original file. Shorten compression can be applied to data in any wavfile format; since it is lossless, it doesn't have to specifically avoid the header data (which will be exactly reconstructed), although the compression will be very poor if the format, alignment or endianness of the actual sound data is not correctly matched. Alternatively, shorten can skip a specified number of bytes at the beginning of the file to leave an initial header unmodified (shorten -a 1024). Compressed wavfiles conventionally have two extensions e.g. chan1.wav.shn is a shorten-compressed NIST wavfile. |
The SNDF tag in the table above is the shorthand used to refer to that format by the various sndutils tools ( sndcat, sndplay etc. - any program based on the sndf library) as well as some scripts that use them.
Use sndcat's -k, -K, -d, -D, -e, or -E options.
The sndcat tool allows you to convert between many different file formats by specifying input format with the -S option and output format with the -T option. To convert between different encoding types use the -f option. The NIST w_decode and w_encode tools can also be useful for converting between encoding types. They may be able to deal with NIST-format files whose headers confuse sndcat (this may happen for example with the embedded-shorten encoding type).
There are also the sox and sph2pipe tools: