[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: Robust method of fundamental frequency estimation.

Dear members of the list,

I forgot the most important part in my last e-mail. I am taking a Grant
Writing class at the University of Florida and, as part of my coursework,
I need to write a grant (of course) and have it reviewed by 5 experts in
the area. If you are willing to serve as a reviewer of my proposal I
really would appreciate it. I have already heard from some of my
classmates that they have had serious problems finding people who are
willing to take some of their time for doing this.

The proposal is about a pitch estimation algorithm (that will improve the
one we will present in ISCAS), a model for production of combination
tones, and a model for pitch perception (using the combination tones
model). The format is the required by NSF. Its length is about 22 pages
including everything (i.e., summary, description, biographies, facilities
and budget). The proposal will be on-line on February 12th for you to
download. You will have 25 days to review it and fill a form in which you
have to evaluate: grammar and style, appropriateness of objectives,
relevance of literature cited, experimental designs, appropriateness of
proposal analyses, budget, and qualifications of the principal
investigator. Then you will have to send it to the professor by e-mail.

If you are interested in being a reviewer please send me an e-mail to give
you the details.



> Dear members,
> I just want to add two points to what Yi-Wen said:
>> Dear list,
>> Just want to draw your attention to a good summary on various
>> auto-correlation based pitch determination methods,
>> Arturo Camacho and John G. Harris, "A biological inspired pitch
>> determination algorithm", Fourth Joint Meeting of ASA and ASJ, Honolulu,
>>  Nov. 2006.
>> Contact arturo@xxxxxxxxxxxx if interested.
>> Best regards,
>> Yi-Wen
> First, in that presentation we not only did a summary of pitch estimation
>  algorithms (PEA) but also pointed out some pitfalls they have. Second,
> we did it not only for autocorrelation based algorithms, but also for many
>  other algorithms we considered to be ?classical?. Although some of these
>  algorithms were initially proposed using a time-domain approach, all of
> them can also be formulated using the spectrum of the signal, and that is
>  the approach we took. We expressed those algorithms as the selection of
> the pitch candidate (PC) that maximizes an integral transform of a
> function of the spectrum.
> Below  is a summary of our findings. For each algorithm, we give a short
> DESCRIPTION, then the FUNCTION applied to the spectrum, the KERNEL of the
> integral transform, and finally a PROBLEM of the algorithm. Sometimes you
> will find that the algorithm also have problems presented before or
> problems that will be presented later. Notice that the order we present
> the algorithms is such that each subsequent algorithm does not exhibit
> the problem mentioned for the previous algorithm. A final note about
> semantics, to make the writing short in the descriptions, when we say
> spectrum we mean MAGNITUDE of the spectrum.
> -------------------------------
> DESCRIPTION: multiplies the spectrum at multiples of the PC, or
> equivalently, adds the log of the spectrum at multiples of the PC.
> KERNEL: periodic sum of pulses
> PROBLEM: If any harmonic of the pitch is missing, the log is minus
> infinity and therefore the integral is also minus infinity.
> ----------------------------
> DESCRIPTION: adds the spectrum at multiples of the PC.
> FUNCTION: none
> KERNEL: periodic sum of pulses
> PROBLEM: Any subharmonic of the pitch has the same score as the pitch.
> ---------------------------------
> DESCRIPTION: Same as SHS but uses a decaying factor to give less weight to
>  high order harmonics. FUNCTION: none
> KERNEL: decaying periodic sum of pulses
> PROBLEM: The same score it produces for a pulse train at the pitch is
> produced for white noise at each PC. Therefore, not only it produces an
> infinite number of pitch estimates for white noise but also they have the
>  same strength as a pulse train.
> -----------------------------------
> DESCRIPTION: Same as SHS but subtracts the spectrum at the middle points
> between harmonics. Uses log spectrum, though. FUNCTION: log
> KERNEL: periodic sum of positive pulses plus half-period-shifted sum of
> negative pulses PROBLEM: Like all the algorithms presented above, it does
> not work for inharmonic signals
> -------------------
> DESCRIPTION: Same as SHS but instead of pulses it uses rectangles
> FUNCTION: none
> KERNEL: sum of rectangles
> PROBLEM: weighting applied to spectrum is too sharp. A slight shift in a
> component may take it in or out of the rectangle, possibly changing the
> estimated pitch drastically.
> -------------
> DESCRIPTION: Same as SHR but instead of pulses uses a cosine to transition
>  from 1 to -1. FUNCTION: log
> KERNEL: cosine
> PROBLEM: uses the log (see HPS)
> ------------------------------
> DESCRIPTION: Same as CEP but squares the spectrum
> FUNCTION: square
> KERNEL: cosine
> PROBLEM: If signal is periodic then UAC is also periodic. Therefore there
> are infinite number of maximums. Taking the first local maximum (excluding
>  maximum at zero) does not work either. Try a signal with first four
> harmonics with magnitudes 1,6,1,1. At high enough levels its pitch
> corresponds to the fundamental frequency, however, the first maximum in
> the UAC corresponds to the second harmonic.
> ------------------------------
> DESCRIPTION: Same as UAC but a bias is applied such that a weight of one
> is applied to a period of 0 and decays linearly to zero for a period T,
> where T is the size of the window. FUNCTION: square
> KERNEL: cosine
> PROBLEM: Like UAC, the squaring of the spectrum gives to much emphasis to
> salient harmonics. This feature combined with the bias may cause problems.
>  For example, for the 1,6,1,1 signal, the bias can make the score of the
> second harmonic higher than the score of the fundamental (take for example
>  the fundamental period as T/4)
> =========In ISCAS 2007 we will be presenting an algorithm that avoids the
> problems presented here. It will be published in the proceedings of the
> conference. From the order we presented here the algorithms it is easy to
> infer what the algorithm looks like.
> Arturo
> --
> __________________________________________________
> Arturo Camacho
> PhD Student
> Computer and Information Science and Engineering
> University of Florida
> E-mail: acamacho@xxxxxxxxxxxx
> Web page: www.cise.ufl.edu/~acamacho
> __________________________________________________


Arturo Camacho
PhD Student
Computer and Information Science and Engineering
University of Florida

E-mail: acamacho@xxxxxxxxxxxx
Web page: www.cise.ufl.edu/~acamacho