Feature files

A data file is the file that contains the feature vectors. It consists of a 16 bytes header followed by the concatenated feature vector. The header is defined by the four following fields:

• data kind (4 bytes), where data kind is an enumeration of the following: OTHER, FBANK, FBCEPSTRA, LPCEPSTRA, LPCOEFF, PARCOR and LAR
• basic feature vector dimension (4 bytes)
• number of feature vectors (4 bytes)
• qualifier (4 bytes)

The qualifier field is a bit field used to describe in more the details the content of the file. The bits have the following meanings:

1. log energy appended (E)
2. long term mean of the data subtracted (Z)
3. absolute log energy suppressed (this bit makes no sense without bit 4) (N)
4. $\Delta$ coefficients appended (D)
5. $\Delta\Delta$ coefficients appended (A)

The letter in parenthesis for each bit is used to refer to a given qualifier (e.g. in the -convert option of scopy). It must be noted that the second header field contains the basic feature vector dimension (i.e. the dimension of the static coefficients regardless of the energy) and not the real feature vector dimension which must be retrieved according to the qualifier field. The feature vectors are stored as float (4 bytes) with the static coefficients first, optionally followed by the log-energy, the $\Delta$ and $\Delta\Delta$ coefficients.