What are the best speaker-specific features used in speaker identification?

Hello everybody there,

I need to know about the best extracted features, used in speaker identification systems, that are related to a speaker's voice. In addition, how should the enrollment phase be done?

If someone has done a research before that concerning this field, please help me.

Thank you in advance.

Answer 1

[Walter Roberson]

What you need to know is:

A) the best features depend upon language and context. You should expect that the features would differ between a glottal language such as Hoansi, compared to the Lenurian D Mystery Cult language spoken by sybils of an unknown god by people who have never been identified yet, somewhere near Delos.

B) So, English, eh? Which dialect? Are the best features the same for Welsh English and Punjabi English?

C) Best? As measured on what test? How does it weight the tradeoffs, and how does it justify those choices? What does it omit?

D) No one knows what the best is. The most you could hope for is the best found so far. Remember, unless you qualify, "best" means impossible to improve, which implies theoretical proof.

[Talal Salim]

Walter Roberson,

Thank you for your response.

A) there are 3 languages: Arabic, English, and French.

B) I do not think this issue is as such complicated as you are describing it by your way.

C) I mean by "Best", features that posses low intra-speaker variation and high inter-speaker variation.

D) Ok. what is the best found so far, as you have mentioned?

Thanks in advance.

[Walter Roberson]

The structural features of Arabic are notably different than the structural features of English and French (which are much more similar to each other.) It should not be expected that the same tests would be optimal for both classes of languages.

In order to have a hope of finding the best so far, you will need to read through the more than 70000 papers on speaker recognition, and you will need to find or write source code for all the different methods (except ones that later papers convincingly establish as being inferior.) Then you will need to establish a common test for each of them, a corpus of speakers and words to train them on -- and that corpus should be representative of many many different aspects of society, not just "trained on college students" (those tend to be wealthier and better educated and less likely to be using their local dialects and vernaculars).

You should expect to produce your results in.. hmmmm, let me think... maybe about 60 years from now?

[Talal Salim]

ohh .. you are so helpful, ehh? thank you.

[Walter Roberson]

I don't have 60ish years to spend answering the question. Chances are that you do not have 60 years to answer the question either. This would suggest that you need to spend some time figuring out a question that you can realistically answer in the time you have available.

Speaking of available resources: this resource, MATLAB Answers, helps people learn MATLAB, and helps describe what is triggering a problem in users' code. However, this resource is not designed to do people's research for them. We assist once the person has an algorithm that they have started to program -- that is, we help with the MATLAB part. We do not help much with research and design of algorithms.

You might want to visit Researchgate, as people there do assist with algorithm suggestions... but not so much with implementation. We here help with implementation, but not much with algorithm suggestions.

[Talal Salim]

In all cases, thank you for your time.

Of course, I was not seeking from any one to give me a complete written code for my problem, but I was asking only for theoretical approaches and ideas and suggestions. I have done my researches from many resources. The idea is that I was curious to know if there are any additional information from MATLAB resources.