Hi, If i calculate the power spectral density of a speech signal using Welchs method i get power/Hz, right ? If i want to find power or energy under 500 Hz i just integrate(or square) from 0-500 Hz? If this is correct would it be the same if i low passed the signal at 500 Hz, and apply Welch? Tommy,

# power spectral density and energy under 500 hz

Started by ●February 28, 2007

Reply by ●March 1, 20072007-03-01

On 1 Mar, 00:41, "stromhau" <strom...@stud.ntnu.no> wrote:> Hi, > > If i calculate the power spectral density of a speech signal using Welchs > method i get power/Hz, right ? > If i want to find power or energy under 500 Hz i just integrate(or square) > from 0-500 Hz?You integrate, without squaring, the spectrum from 0 t0 500 Hz.> If this is correct would it be the same if i low passed the signal at 500 > Hz, and apply Welch?In principle, the two results *could* be the same, yes. In practice, your low pass filter might have a non-unity scaling factor which alters the signal energy, if not accounted for. Welch' method uses certain scaling windows and averages, which also tend to affect the exact numerical value of the PDS. The ususal advice is to be a bit careful using Welch' method if you want exact numerical values for the PDS. Welch' method is better suited to get an idea about general sectrum shape, rather than the fine details. Rune

Reply by ●March 1, 20072007-03-01

>On 1 Mar, 00:41, "stromhau" <strom...@stud.ntnu.no> wrote: >> Hi, >> >> If i calculate the power spectral density of a speech signal usingWelchs>> method i get power/Hz, right ? >> If i want to find power or energy under 500 Hz i just integrate(orsquare)>> from 0-500 Hz? > >You integrate, without squaring, the spectrum from 0 t0 500 Hz. > >> If this is correct would it be the same if i low passed the signal at500>> Hz, and apply Welch? > >In principle, the two results *could* be the same, yes. In practice, >your low pass filter might have a non-unity scaling factor which >alters the signal energy, if not accounted for. Welch' method >uses certain scaling windows and averages, which also tend >to affect the exact numerical value of the PDS. > >The ususal advice is to be a bit careful using Welch' method >if you want exact numerical values for the PDS. Welch' method >is better suited to get an idea about general sectrum shape, >rather than the fine details. > >Rune > > >Thank you for answering, What method would you recommend for this problem? Tommy,

Reply by ●March 1, 20072007-03-01

On 1 Mar, 13:10, "stromhau" <strom...@stud.ntnu.no> wrote:> >On 1 Mar, 00:41, "stromhau" <strom...@stud.ntnu.no> wrote: > >> Hi, > > >> If i calculate the power spectral density of a speech signal using > Welchs > >> method i get power/Hz, right ? > >> If i want to find power or energy under 500 Hz i just integrate(or > square) > >> from 0-500 Hz? > > >You integrate, without squaring, the spectrum from 0 t0 500 Hz. > > >> If this is correct would it be the same if i low passed the signal at > 500 > >> Hz, and apply Welch? > > >In principle, the two results *could* be the same, yes. In practice, > >your low pass filter might have a non-unity scaling factor which > >alters the signal energy, if not accounted for. Welch' method > >uses certain scaling windows and averages, which also tend > >to affect the exact numerical value of the PDS. > > >The ususal advice is to be a bit careful using Welch' method > >if you want exact numerical values for the PDS. Welch' method > >is better suited to get an idea about general sectrum shape, > >rather than the fine details. > > >Rune > > Thank you for answering, > > What method would you recommend for this problem?What question do you try to answer? Do you want relative power contributions from different bands of the spectrum, or do you need exact numbers? Rune

Reply by ●March 1, 20072007-03-01

On Feb 28, 3:41 pm, "stromhau" <strom...@stud.ntnu.no> wrote:> Hi, > > If i calculate the power spectral density of a speech signal using Welchs > method i get power/Hz, right ? > > > Tommy,For correct scaling of welch for your choice of units take a look at: http://www.bksv.com/pdf/bv0031.pdf on page 29 for a discussion of Signals and Units. Dale B. Dalrymple

Reply by ●March 1, 20072007-03-01

>On 1 Mar, 13:10, "stromhau" <strom...@stud.ntnu.no> wrote: >> >On 1 Mar, 00:41, "stromhau" <strom...@stud.ntnu.no> wrote: >> >> Hi, >> >> >> If i calculate the power spectral density of a speech signal using >> Welchs >> >> method i get power/Hz, right ? >> >> If i want to find power or energy under 500 Hz i just integrate(or >> square) >> >> from 0-500 Hz? >> >> >You integrate, without squaring, the spectrum from 0 t0 500 Hz. >> >> >> If this is correct would it be the same if i low passed the signalat>> 500 >> >> Hz, and apply Welch? >> >> >In principle, the two results *could* be the same, yes. In practice, >> >your low pass filter might have a non-unity scaling factor which >> >alters the signal energy, if not accounted for. Welch' method >> >uses certain scaling windows and averages, which also tend >> >to affect the exact numerical value of the PDS. >> >> >The ususal advice is to be a bit careful using Welch' method >> >if you want exact numerical values for the PDS. Welch' method >> >is better suited to get an idea about general sectrum shape, >> >rather than the fine details. >> >> >Rune >> >> Thank you for answering, >> >> What method would you recommend for this problem? > >What question do you try to answer? Do you want relative >power contributions from different bands of the spectrum, >or do you need exact numbers? > >Rune > >Hi, I want to find the amount of energy below 500 Hz, so i guess exact numbers. Tommy,

Reply by ●March 1, 20072007-03-01

This energy or power thing is one of many features i want to implement. When reading paper on the subject(emotional speech) it is only mentioned as energy under a value of hz(often 500Hz). So the question is how can this be done ? Tommy,

Reply by ●March 2, 20072007-03-02

On 1 Mar, 23:03, "stromhau" <strom...@stud.ntnu.no> wrote:> This energy or power thing is one of many features i want to implement. > When reading paper on the subject(emotional speech) it is only mentioned > as energy under a value of hz(often 500Hz). > So the question is how can this be done ?You have to search the literature on speech processing to see what methods they use. There is likely to be some accepted "usual" way of doing things. Post on one of the speech processing newsgroups. Rune