get_f0.1

来自「speech signal process tools」· 1 代码 · 共 242 行

.\" Copyright (c) 1996 Entropic Research Laboratory, Inc.; All rights reserved
.\" %W% %G% ERL
.ds ]W (c) 1993 - 1996 Entropic Research Laboratory, Inc.
.TH  GET_F0 1\-ESPS 9/5/96
.SH NAME

.nf
get_f0 \- robust analysis of speech fundamental frequency (pitch tracking)
.fi
.SH SYNOPSIS
.B
get_f0
[
.BI \-P " param_file"
][
.BI \-{pr} "range"
][
.BI \-s " range"
][
.BI \-S " frame_step"
][
.BI \-i " frame_step"
][
.BI \-x " debug_level"
]
.I in_file
.I out_file
.PP
.SH DESCRIPTION
\fIGet_f0\fR implements a fundamental frequency (F0) estimation
algorithm using the normalized cross correlation function and dynamic
programming.  The algorithm implemented here is described exactly in
the reference cited below.
.PP
The input file \fIin_file\fR is a standard FEA_SD sampled-data data
file.  The output file \fIout_file\fR is a FEA file containing 4
fields: \fIF0\fR for fundamental frequency estimate, \fIprob_voice\fR
for "probability of voicing", \fIrms\fR for local root mean squared
measurements, and \fIac_peak\fR for the peak normalized
cross-correlation value that was found to determine the output F0.
The RMS value of each record is computed based on a 30 msec hanning
window with its left edge placed 5 msec before the beginning of the
frame.  In unvoiced regions, \fIac_peak\fI is the largest
cross-correlation value found at any lag.  The \fIprob_voice\fR element
only takes on two values: 0 and 1, whereas in the older \fIformant\fR
program, it was a graded (though error prone) measure.
.PP
Note that the analysis frame onset is well defined as the point where
the reference window used for cross-correlation begins.  However the
effective frame size is really a function of the local F0, except for
the RMS measurement as stated above.
.PP
If \fIin_file\fR is replace by by "-", the standard input is  read.
If \fIout_file\fR is  replaced  by  "-", the standard output is
written.  The processing is truly stream oriented.  There is no limit
on the length of the input sequence.
.PP
\fIGet_f0\fR does not remove the DC component from the input signal.
Large DC offsets will impair the voiced/unvoiced decision and lead to
misleading RMS measurements, especially in low-amplitude regions.  It
is recommended that the DC component be removed by a program such as
\fIrem_dc\fR(1-ESPS) before using \fIget_f0\fR.
.PP
Note that \fIget_f0\fR is designed to replace the pitch tracking
function of the older \fIformant\fR program. fIGet_f0\fR is both
faster and more accurate than \fIformant\fR, and does not have the
batch-processing limitations of the latter.
.PP
.SH OPTIONS
.PP
The following options are supported:
.TP
.BI \-P " param_file \fR[params]\fP"
uses the parameter file \fIparam_file\fR rather than the default, 
which is \fIparams\fR. 
.TP
.BI \-r " first:last"
.TP
.BI \-r " first:+incr"
Determines the range of points from input file.  In the first form, a
pair of unsigned integers gives the first and last points of the
range.  If \fIfirst\fR is omitted, 1 is used.  If \fIlast\fR is
omitted, the last point in the file is used.  The second form is
equivalent to the first with \fIlast = first + incr\fR.  If no range
is specified, the whole input file is processed.
.TP
.BI \-p " "
Same as the \fB-r\fR option.  (Note that this is a change from version
5.0, where \fB-p\fR was used for the frame intertval option.)
.TP
.BI \-s " first:last"
.TP
.BI \-s " first:+incr"
Same function as the \fB-r\fR option, but specifies the range of input data 
in seconds
.TP
.BI \-i " frame_step [0.01]"
Specifies frame step in second, between 0.1 and 1/sampling rate in sec.
.TP
.BI \-S " frame_step [0.01 * sampling frequency]"
Same as the -i option, but specifies frame step in samples
.TP
.BI \-x " debug_level \fR[0]\fP"
If 
.I debug_level
is positive,
.I get_f0
prints debugging messages and other information on the standard error
output.  The messages proliferate as the  
.I debug_level
increases.  If \fIdebug_level\fP is 0 (the default), no messages are
printed.  
.SH ESPS PARAMETERS
\fIGet_f0\fR is designed for use "as is" with little or no need to
change its parameters under most circumstances.  Exceptions include
the frame rate, and the maximum F0 and minimum F0 to track.  If the F0
estimates from \fIget_f0\fR do not appear reasonable, you should check
your signal or signal conditioning before beginning parameter
adjustments.  Common causes of difficulty include a strong periodic
component in the background causing the voicing to stay on, or a
significant DC offset causing poor RMS estimates.  The following
parameter file options are supported.
.TP
.I "start - integer"
.IP
The first point in the input sampled data file that is processed.  A
value of 1 denotes the first sample in the file.  This is only read
if the \fB\-p\fP option is not used.  If it is not in the parameter
file, the default value of 1 is used.  
.TP
.I "nan - integer"
.IP
The total number of data points to process.  If 
.I nan
is 0, the whole file is processed.  
.I Nan
is read only if the \fB\-p\fP option is not used.  (See the 
discussion under \fB\-l\fP).
.TP
.I " frame_step - float"
Analysis frame step interval. Computation increases as 1/\fIframe_step\fR. 
Valid value lies in [1/sampling rate,  0.1].
.TP
.I " min_f0 - float"
Minimum F0 to search for. Note that computational cost grows as
1/min_f0.  Valid values are greater than or equal to (Fs/10000) Hz,
where Fs is the sample rate of the input speech signal.  Default is
50.0.  \fImin_f0\fR and \fImax_f0\fR determine the number of
cross-correlation lags to compute for each frame.
.TP
.I " max_f0 - float"
Maximum F0 to search for.  Valid values are greater than \fImin_F0\fR and
smaller than one half the sampling rate of input file.  Default is 550.0.
\fImin_f0\fR and \fImax_f0\fR determine the number of
cross-correlation lags to compute for each frame.
.PP
The default settings of the following parameters were determined by 
exhaustive search of the parameter space using hand-verified
data as the reference.  Twiddling with the parameters is not recommended.
.TP
.I " cand_thresh - float"
Determines cross correlation peak height required for a peak to be
considered a pitch-peak candidate.  Valid value lies in [.01, .99].
Default is 0.3.
.TP
.I " lag_weight - float"
Amount of weight given to the shortness of the proposed pitch interval.
Higher numbers make high F0 estimates more likely. Valid value lies in [0, 1].
Default is 0.3.
.TP
.I " freq_weight - float"
Strength of F0 continuity.  Higher numbers impose smoother contours. 
Valid value lies in [0, 1].  Default is 0.02.
.TP
.I " trans_cost - float"
Fixed cost of making a voicing-state transition.  Higher numbers discourage
state changes. Valid value lies in [0, 1].  Default is 0.005.
.TP
.I " trans_amp - float"
Voicing-state transition cost modulated by the local rate of amplitude
change.  Higher numbers discourage transitions EXCEPT when the rate of
amplitude change is great. Valid values lie in [0, 100].  Default is 0.5.
.TP
.I " trans_spec - float"
Voicing-state transition cost modulated by the local rate of spectral
change.  Higher numbers discourage transitions EXCEPT when the rate of
spectral change is great. Valid values lie in [0, 100].  Default is
0.5.
.TP
.I " voice_bias - float"
Determines fixed preference for voiced or unvoiced state.  Positive
numbers encourage the voiced hypothesis, negative numbers the
unvoiced. Valid values lie in [-1, 1].  Default is 0.0.
.TP
.I " double_cost - float"
The cost of a rapid one-octave (up or down) F0 change.  High numbers
discourage any jumps, low numbers permit octave jumps. Valid values lie in
[0, 10].  Default is 0.35.
.TP
.I " wind_dur - float"
Size of correlation window.  Computation increases directly as wind_dur. 
Valid values lie in [10/sampling rate, .1].  Default is 0.0075.
.TP
.I " n_cands - integer"
The maximum number of correlation peaks considered as possible
F0-peak candidates in any frame.  At most, the top n-cands candidates
are considered in each frame. The computational cost grows approximately
as n_cands SQUARED.  Valid values lie in [3, 100].  Default is 20.
.PP
.SH ESPS COMMON
No ESPS common parameter processing is supported.
.PP
.SH ESPS HEADERS
The usual \fIrecord_freq\fR, \fIstart_time\fR header items, all 
supported parameters are stored as generic header items.  In addition,
the \fIrecord_freq\fR header item of the \fIin_file\fR input file is
saved as the \fIsrc_sf\fR header item.
.PP
.SH FUTURE CHANGES
In a future release DC will be removed prior to RMS comutation.  Also,
an optional element may be added to the output vector to include RMS
computed on the preemphasized speech.
.PP
.SH EXAMPLES
.PP
.SH ERRORS AND DIAGNOSTICS
.PP
.SH BUGS
.PP
None known.
.SH REFERENCES
Talkin, D. (1995). A Robust Algorithm for Pitch Tracking (RAPT). In
Kleijn, W. B. and Paliwal, K. K. (Eds.), \fISpeech Coding and
Synthesis\fR. New York: Elsevier.

.PP
.SH "SEE ALSO"
FEA(5\-ESPS), epochs(1-ESPS), formant(1-ESPS), rem_dc(1-ESPS)
.PP
.SH AUTHORS
David Talkin, Derek Lin