Is it (theoretically) possible to provide of phonetic analysis of a spectrogram

[zaba]

.... without knowing the language?

I have a friend who claims that she can do this for any language, but I am suspicious. How can I prove her wrong? Well, I could use my own data and try to show her how tough it is, obviously -- but how could I prove her wrong on an intellectual level?

[MalFet]

What do you mean by "phonetic analysis"? A spectrogram *is* a phonetic analysis, or at least a schematization of phonetic data.

Is she claiming that she can "hear" spectrograms?

[lx]

It's pretty normal to be able to look at a spectogram and be able to see what is going on phonetically. You can follow strong vowel formants and hypothesise about tongue height/position, separate voiced from unvoiced sounds, spot VOT bursts to hint at plosives, use relative energy to determine where in the mouth it came from. If you mean being able to get back to the complete word, that's really rare, but you can get close. The first time we were introduced to spectograms in university, by the end of the third class it would take us 2-3 goes in order to get back to the short word that was recorded, by looking at the spectogram and reading what it was telling us about the phonetics.

I don't really get your question though, because as MalFet says, a spectogram is a phonetic analysis.

[Daniel]

Phonetics yes. Phonology no. And since the two interact it can be limiting to only have acoustic information. Broad categories are easy enough (like vowel vs consonant) but getting an exact transcription would be hard because of no phonological/grammatical information. Plus, spectrogram a have limited info in them, low resolution. There are programs that can play a spectrogram as sound and if you try one you'll see that it works and conveys some info but is far from clearly intelligible.

[Pramod Kumar Agrawal]

I wanted to check my knowledge on spectrograms.
1. The front vowels and back vowels can be represented by F2-F1. More of the value of F2-F1, the vowel will be more towards front vowel /i/.
2. Devanagari vowels /ɻ/ and /Ɩ/ can be represented by minimum F0 and maximum F0 respectively.
3. The unvoiced consonants have blank ‘stop period’, and voiced consonants have filled ‘stop period’.
4. We cannot differentiate two different consonants if both consonants are unvoiced or both are voiced.
Please correct me.

[lx]

I wanted to check my knowledge on spectrograms.
1. The front vowels and back vowels can be represented by F2-F1. More of the value of F2-F1, the vowel will be more towards front vowel /i/.
2. Devanagari vowels /ɻ/ and /Ɩ/ can be represented by minimum F0 and maximum F0 respectively.
3. The unvoiced consonants have blank ‘stop period’, and voiced consonants have filled ‘stop period’.
4. We cannot differentiate two different consonants if both consonants are unvoiced or both are voiced.
Please correct me.

1. What you wrote there isn't really correct. There is a better way to look at it. F1 is negatively correlated with vowel height (low F1 = high vowel / high F1 = low vowel) and F2 is positively correlated with vowel frontness/backness (high F2 = front, low F2 = back). So, if you want /i/ (a high front vowel) then you know a) that it is high, and therefore that it will have a low F1, and you know that it is front, so it will have a high F2. So you can take the features and build back up what you know about the vowel.

2. I've got no idea about Devanagari vowels so I will skip this.

3. 'Plosives' you mean, not all consonants. There is no period of silence in, for example, /s/. For voiced plosives, there is still a period of silence, it's not filled, but you see a very low string of energy low in the spectrum (it's called a voice bar) while all the higher regions can be considered a period of (relative) silence.

4. You can differentiate. Having a spectrum of /afa/ will look a lot different from /asa/. There is less energy released with non-sibilant fricatives than there is for sibilant fricatives and in a comparison of two spectra it should be easy to tell which is which.

[Daniel]

What lx said. And:

2. Devanagari vowels /ɻ/ and /Ɩ/ can be represented by minimum F0 and maximum F0 respectively.

I believe you might need to also include F3 values, because I know that in other languages (like English) F3 is important for distinguishing R sounds. (This is said to be the reason that, for example, speakers of Japanese do not hear a difference between R and L, because they are not listening for F3 variation, only F1 and F2.)

4. We cannot differentiate two different consonants if both consonants are unvoiced or both are voiced.

Certainly hearers of the language can differentiate sounds. Therefore, there is enough acoustic information to do so. But it is harder than distinguishing other things like vowel height/backness or voicing.
As lx said, amplitude of frication at different frequencies can help determine different places of articulation for fricatives, and manner can be distinguished by the general shape of the acoustic pattern (for example, stops vs. fricatives).
Distinguishing different places of articulation for stops may be the most difficult. This is because the specific vowels affect the stops differently, due to interaction of higher/lower frequencies. The direction of change in F1 and F2 is an indication, but there are complicated co-articulation effects, so it's not simple to read. Here's an example:

[Pramod Kumar Agrawal]

During an experiment I found a specific pattern in F1, F2, and F3. I took a series of vowels /i e æ ǝ ɒ o u/. In my experiment, I found that value of F1 does not change significantly. It increases till ‘central vowel’ and again goes down in the case of ‘back vowels’. The value of F2 decreases continuously. And value of F3 remains within limitations. The value of F2-F1 (F2 minus F1) is in decreasing order. And the value of F3-F2 (F3 minus F2) is in increasing order.
By repeating same experiment with IPA and Devanagari vowels, I found that individually F1, F2 or F3 does not say anything much, but their successive differences explain many things. (I have taken sufficient precaution during the experiments, but still there are chances of mistakes)

[Daniel]

This topic has been researched a lot.
https://www.google.com/search?q=vowel+formants

Both the relative distance and the absolute position of the formants matters.

[Pramod Kumar Agrawal]

I found that absolute positions of F1, F2, and F3 differ with person to person. Female and children voices have higher values. But the patterns of  F2-F1 and F3-F2 do not change.

[lx]

I found that absolute positions of F1, F2, and F3 differ with person to person. Female and children voices have higher values. But the patterns of  F2-F1 and F3-F2 do not change.

But you would have expected that without even doing any experiment, right? It's the type of thing a student would expect to find with only minimal familiarity with the field of phonetics. No shockers here. It is important to separate out what the brain processes and what physically exists in the environment as they can be quite different. I could play you a few frequencies at the same time, with  the lowest one being 600Hz and I could order them every 200Hz (600Hz, 800Hz, 1000Hz) and it'd force you to think you were listening to a 200Hz fundamental frequency when that wasn't even played. You always have to be aware of the way the ear builds sounds back up and the ways in which it can be tricked.

[jkpate]

The task you are discussing here is speaker normalization. This is actually quite tricky, and taking the differences of formants doesn't quite work. For example, Adank et al. (2004) compared several speaker normalization algorithms. One of these, the Syrdal and Gopal (S&G) approach, is a difference of formants (in barks). S&G performed worst at preserving phonemic identity and talker regional dialect. The algorithms that worked best generally focused on aggregating information across vowels to squish and stretch talker vowel spaces, rather than simply transforming measures from individual tokens in isolation.

[Daniel]

I should clarify what I said earlier:

Everything is relative to the speaker. Nothing is absolute. What I meant is that the individual formants are important information ("absolute=individual positions"), as well as the distance between them ("relative positions"). But that's all relative to the speaker.

And again, what you're finding has been documented by hundreds of studies as is very well understood.



Edit:
Here's a paper from 1952 that covers the basics. This information has been known for a long time and hasn't changed much. Details in descriptions of various languages are of course important, but the overall information about how vowels work across speakers is clear.
http://www.ling.upenn.edu/courses/ling520/PetersonBarney52.pdf
http://dx.doi.org/10.1121/1.1906875

[Pramod Kumar Agrawal]

Thanks! My knowledge regarding spectrogram is not up to the mark. I am incorporating the fact in my next book. It is an important circumstantial evidences in favour of the theory of phonosemantics.
There is one more question. As regards consonants, can we draw a detailed spectrogram showing something just before the ‘stop’? My theory suggests that there should be a little nasal sound just before the ‘stop’.

[Daniel]

As regards consonants, can we draw a detailed spectrogram showing something just before the ‘stop’?

Yes. You can look this up in any introductory phonetics textbook, or online.

My theory suggests that there should be a little nasal sound just before the ‘stop’.

No, there's no reason there would be a nasal sound unless you are talking specifically about prenasalized consonants.


This is also a very uncomfortable approach for me because I believe that theories should be built from data, rather than determining a theory and then trying to fit the data to it. You can, of course, use the method of falsification, but then you must be prepared to abandon your theory if you find the test fails.