Normalisation of parameters by subtraction of mean. More...

Functions
void	spr_meannorm_free (SprSspInfo *Info)

int	spr_meannorm_setup (SprSspInfo Info, const char descript, void aux_info)

int	spr_meannorm_process (SprSspInfo Info, const void frame_in, void *frame_out)

void	spr_meannorm_reset (SprSspInfo Info, SprSspStatus action)

Detailed Description

Normalisation of parameters by subtraction of mean.

This programme subtracts a mean value from the parameters in the input. The initialisation values of the mean are calculated over some input frames of the input file. The mean is updated for each new input frame, using a weight.

[meannorm]
`nfr_init <number>(100)`
Number of frames used for initial mean calculation. Specify -1 if all frames are to be used.
`alpha <number>(0.05)`
Weight for mean-update.
`file <fname>`
Initial means are given in a file; specifying this option has the side-effect of setting the <nfr_init> param to 0.
`no_reset`
Do not reset the average spectrum at the beginning of a new file.
`spdet_in <copy/move> <buf_name>`
Input from a silence speech detector: the mean is only update for speech frames.
`multi_spkr <N> <copy/move> <buf_name>`
Setup meannorm to work in a multi-speaker environment, i.e. cepstral means for the (at least) the N last (leat recent used) speaker id's are calculated. The speaker id's are input from a named buffer.
`max_delay <number>(-1)`
Maximum delay (#frames) one wants to allow in a multi-speaker scenerio.
`min_nfr <number>(-1)`
Minimum number of frames one wants to calculate a mean, if less frames are available, a back-off to silence frames and frames from other speakers is performed.
`adapt init=<yes/no>(yes) window=<exp/dexp>(exp)`
Mean adaptation strategy: do (not) adapt during the first <nfr> frames and use a simple exponential windowing or a double exponential windowing.
`output <xnorm/mean>(xnorm) [mean/nill](nill)`
Output the normalized data (xnorm) optionally followed by the mean values (mean), or the mean values only; other combinations are invalid.
`history <hist_len>`
Limit the history length to <hist_len> (is infinity with any non-zero alpha).

Author: Tom Claes (algorithm); Jacques Duchateau (conversion to routines)

Date: 16 May 1994, 17 October 1994

Revision History:

14/04/98 - KD: Added the silence speech detector option.
17/02/04 - KD: Full rewrite for multi speaker support

Functions

Detailed Description