Mercurial > hg > emotion-detection-top-level
diff Code/Descriptors/yin/yin.html @ 4:92ca03a8fa99 tip
Update to ICASSP 2013 benchmark
| author | Dawn Black |
|---|---|
| date | Wed, 13 Feb 2013 11:02:39 +0000 |
| parents | |
| children |
line wrap: on
line diff
--- /dev/null Thu Jan 01 00:00:00 1970 +0000 +++ b/Code/Descriptors/yin/yin.html Wed Feb 13 11:02:39 2013 +0000 @@ -0,0 +1,84 @@ +<html> +<head> +<title> YIN </title> +</head> + +<body> + +See README for copyright information. + +<hr> +<h2> YIN: fundamental frequency estimator </h2> +<hr> + +YIN estimates the fundamental frequency (F0) of an audio signal. +Features are: +<ul> +<li> Reliability (based on tests, see reference below). +<li> Accuracy (subsample resolution). +<li> Wide search range (default is 30 Hz - sr/4). +<li> Good temporal resolution. +<li> Ease of use. +</ul> + +YIN operates on vectors or files. YIN outputs a structure containing a set +of four vectors: F0 vs time, two estimates of aperiodic/total power +(one gross estimate, one fine estimate), and a period-smoothed estimate +of instantaneous power. +<p> +If no output argment is specified, YIN +plots F0 as a function of time (in octaves re: 440 Hz), aperiodicity, and power. +<p> +In the F0 plot, samples in blue are reckoned reliable (aperiodicity<threshold), +green are intermediate (aperiodicity<2*threshold), and +yellow unreliable (aperiodicity>2*threshold). + +<p> +Type 'help yin' for a description of the parameters. Read the reference below and +the code to understand their meaning. In brief: +<ul> +<li> To increase speed: increase 'hop' or 'minf0'. +<li> To reduce memory needs: reduce 'bufsize', or increase 'hop' or 'minf0'. +<li> To slightly increase reliability: reduce 'hop'. +<li> To slightly increase precision: upsample before processing. +<li> To improve temporal resolution: increase 'minf0', decrease 'hop'. +<li> To process lower F0s: reduce 'minf0'. Higher F0s: upsample and increase 'maxf0'. +<li> To avoid subharmonic errors: increase 'thresh'. +<li> To avoid harmonic/formant errors: reduce 'thresh'. +<li> Make sure that the range [minf0 maxf0] includes the expected f0. +</ul> +Parameter 'thresh' sets the proportion of aperiodic power that is +tolerated within a "periodic" signal. This may vary according to the application. +<p> +For speech or musical instruments a value of 0.1 is usually adequate. Singing voice +may require a smaller value (as low as 0.001) if a harmonic is reinforced by a +sharp formant. +<p> +Some signals are inherently ambiguous. For example +the response of a high-Q resonator excited by a pulse train may be seen either as a complex +tone with an F0 equal to that of the pulse train, or as an +amplitude modulated pure tone with an F0 equal to the resonant frequency. +Neither is more "correct" than the other. To obtain the result +that you expect, you must set the threshold to an appropriate value: small for the +fundamental periodicity, large for the resonance periodicity. + +<p> +YIN is described in: +<br></tt> +de Cheveigné, A., and Kawahara, H. (2002). "YIN, a fundamental frequency estimator +for speech and music," J. Acoust. Soc. Am., 111, 1917-1930. (<a href= + "http://www.ircam.fr/pcm/cheveign/ps/yin.pdf">pdf</a>) + </tt> + + <hr> +[Code is <a href="http://www.ircam.fr/pcm/cheveign/sw/yin.zip">here</a>] +[<a href="http://www.ircam.fr/pcm/cheveign">Alain de Cheveigné</a>] + + + + + + </body> + </html> + +