Design and simulate filter banks. More...

Enumerations
enum	{ SPR_FB_WARP_LIN1, SPR_FB_WARP_LIN2, SPR_FB_NWARPMETHOD }

Functions
void	spr_filter_bank_free (SprSspInfo *Info)

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

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

Detailed Description

Design and simulate filter banks.

designs and evaluates filter banks based on FFT and auditory models. Four types of filterbanks are possible:

rectangular, non-overlapping, centers 1 unit spaced on freqscale
triangular, 50% overlap, centers 1 unit spaced on freqscale (cf. Davis & Mermelstein)
schroeder, approximation of Hermansky to the asymmetric masking-curve of Schroeder (cf. plp)
gammatone, approximation of a gammatone filter (see eqn. (3.7) in A.M. Darling, "Properties and Implementation of the GammaTone Filter: A Tutorial")

These banktypes can be combined with four non-linear frequency-scales :

  - PLP   >  barks as in plp
  - MEL   |
  - ERB   |>  cf. internal report 'Signal Processing for Speech Applications, Formulas and Algorithms', MI2-SPCH-91-5, p. 28 
  - BARK  |
  - MELDM >  linear up to 1000 Hz, logarithmic spacing above (cf. Davis & Mermelstein)
  - MELHTK
  - LINEAR

In addition, warping can be applied on the defining (begin/middle/end) frequencies of the filterbanks before the bank weights are calculated. Currently, only linear warping is implemented.

[filter_bank]
`scale <MEL/BARK/ERB/PLP/MELDM/MELHTK/ERBGM/LINEAR>`
Set scale type. Default MELDM.
`bank <REC/TRI2/SCHROEDER/TRI/GAMMA> [norm/no_norm](no_norm)`
Set bank type. Default TRI.
`output <dB/power>(dB) [dB_low](-379)`
Output in dB (default) or power. When converting to dB, assure that the output is never lower than <dB_low>.
`step <step>(1.0)`
Go through the frequency scale with step greater (<step> > 1.0) or smaller (<step> < 1.0) than the critical bandwidth.
`width <fac>(1.0)`
Make the filterbank wider (<fac>>1.0) or narrower (<fac><1.0) than the default shape (which also depends on <step>). Note: not all filtershapes support this option.
`clip <f0>(0.0) [f1](-1)`
Only use part of the frequence range. Values are either in Herz (positive values) or in fractions of the Nyquist frequency (negative numbers).
`Nfbank <val>`
Specify the number of filterbanks one wants.
`LWfixed <value> [vtl/warp](warp)`
Use fixed value for linear warping. Default (warp), values for male subjects are smaller than one, choose vtl for inverse value meaning.
`LWstream <copy/move> <buf_name> [vtl/warp](warp)`
Use frame-by-frame values in buffer for linear warping. See LWfixed for vtl/warp choice.
`LWhfreq <flat/pwl> [freq] [max_warp]`
Options to solve the missing high frequencies in warping: default <flat> fills in a constant value namely the average power in frequencies from <freq> onwards (default a 2% range is used, so <freq> is 7840Hz for 16kHz), <pwl> is piece wise linear with 2 pieces and corner frequency <freq> (default a 10% range is used, so <freq> is 7200Hz for 16kHz). For a warping factor stream, the input warping values will be restricted to [1/max_warp max_warp], default values: 2.0 for <flat> (in this case, some memory allocations depend on max_warp), and for <pwl> the maximal value possible given <freq> (1.11111 for its default).

Author: Tom Claes (algorithms); Jacques Duchateau (ssp routines)

Date

07/02/1994 - TC: First implementation.
04/11/1994 - JD: Adpated for the on-line signal processing
17/11/2004 - JD: Rewritten, adding vtln warping
02/01/2012 - KD: Rewritten, adding step + renorm options, removed dead weight
30/10/2014 - KD: Added the GammaTone filter

See Also: ssp_master.c

Enumerations

Functions

Detailed Description