Speech and Language Processing - Science topic

I hope you get it... Best of luck. Commenting for better each for you...

Did you try to use different engines?

Segmentation of naturally spoken speech into words, even when there is a relatively small dictionary of words, is a harder problem than recognizing isolated digits.

People tend to think of spoken words as somehow isolated but "close" in time. This is not the case, unless you have a cooperating speaker (who helps the detection, or at least monitors it and repeats when it misdetects).

You can easily find in the literature the standard end-point detection mechanisms people use (mostly Viterbi based), and then run the isolated word detectors, but they are computationally expensive and don't really work very well for natural speech (the possible exception was flexible endpoint DTW, but I doubt that you are using DTW as a detector).

Y(J)S

Kesavarao,

Lots of links to Praat software. Its a free download.

Try: https://praat.en.lo4d.com

or: https://download.cnet.com/Praat/3000-2279_4-75296009.htm

Jeff Knox

this link is useful

https://books.google.iq/books?isbn=1782830782

regards

There is also this: http://www.oxfordwaveresearch.com

Hi Wy, you may find the RAVDESS helpful in your work. It's a validated multimodal database of emotional speech and song. It contain 7356 recordings in English, with 8 emotions: calm, happy, sad, angry, fearful, surprise, disgust, and neutral, each at two emotional intensities. Downloaded for free - https://zenodo.org/record/1188976

I am not familiar with a prior attempt to collapse the RBANS components and then compare to the Addenbrooke.  I think the "language" and "fluency" domains of the ACE-III could be collapsed to compare directly to the RBANS.  I tend to prefer keeping immediate and delayed memory performances separate as different factors can affect each and if they were to be collapsed, that information would be lost.  However, if you are wanting to have a global mnemonic performance level, it would be interesting to see what you find. 

These are few patents, and you can utilize for your research

there was excellent work in 1987 by Jagodzinski in Sociological Methods on using quasi-simplex models to get estimates of reliability and stability of single items. Tremendous papers and still very relevant when trying to assess psychometric properties of single items

Praat is so useful dear.

Regards

you might be interested in the Aljameel et al (2016) survey article about string similarity for the Arabic language (please see link)

Hi, we wrote a whole series of papers which investigate this issue, looking at disfluent "er", repetition, and silence (in English).  In brief, "er" and silence increase recognition memory for words which follow disfluencies (Corley, MacGregor, and Donaldson, 2007; MacGregor, Corley, and Donaldson, 2010); repetitions do not (MacGregor, Corley, and Donaldson, 2009).  In the case of "er" and silence, there is an associated attenuation of an N400 ERP effect at the target word, suggesting that people's expectations about what they will hear have been affected.  Importantly, Collard, Corley, MacGregor and Donaldson (2008) use a P300 ERP paradigm to show that this altered expectation is accompanied by an attentional modulation (as you suggest above).

A synthetic view of these studies might be that disfluency is detected when the signal becomes "non-linguistic" (hence not repetitions), and acts as a signal that the speaker is unlikely to utter a predictable word (N400 modulation).  This causes the listener to heighten attention to the signal (reliance on bottom-up information; P300 modulation), resulting in a greater recognition memory for the subsequent (target) word.

Hope that helps!

--MC

Dear Fabio Gasparini,

Generally speaking, color naming , color semantics, color  categorization , and  shape of color space across different languages has long been an area of great interest in bilingual/multilingual studies. Prototypicality norms ( Rosch's model) have also been a matter of importance trying to examine how bilinguals tend to categories the tokens belonging to a given language type. I hope the following links shed more light on what you are looking for.

Best of luck,

R. Biria

There are two things that come to my mind.  First, one should be able to see a spike at the onset and a decay at the termination of speech if you are tapping the signals at C3, C4. or F3, F4.  The planning of opening the mouth and the corresponding signal should be evident at F3, F4.  Secondly, one may even get an EMG signal due to the muscle movement.  Strictly speaking, it is an artifact.  However, it may even do the job for your experiments.  Let me know your thoughts.

My simple advice to all non-English native speakers (including me!)  to convert or switch their electronic computers, mobiles, social media accounts, TVs, etc into English mode, I think is the simplest way to gain more vocabularies and be default to communicate by using English language.

Emad

this is speech corpus where is the text?

I'm asking specifically about the tool and algorithms used in producing completed  corpus

you seems added your answer without reading the question and wasted my time

I mean for mexican, I would like one that can recognize speech in natural environment. Anyway, I think I will use nuance.

Thank you all!

We might have some posters that could be interesting, on cross-linguistic priming.

I personally prefer Praat scripts

web.stanford.edu/dept/linguistics/corpora/materi PRAAT_workshop_manual_v421.pdf

Good luck

Dear Michal S. Karbownik,

The application of  a questionnaire is believed to be meritorious because of  reliability, comparability, relative ease of administration and  analysis as well as its potential benefit of obtaining  data from  large samples . However, as you have rightly observed, it may also result in certain undesirable  epistemological consequences. In point of fact, the validity of the data depends  on the extent to which the targeted respondents are able to understand the statements or the  items on the questionnaire. Notably, readability of the statements or questions  may have a direct bearing on  the expected outcomes leading to such  problems as low return rates, missing data, and careless responses. Such issues can negatively influence the representativeness of  the expected   outcome and the validity and reliability of the instrument. As such, to measure the validity of the translated questionnaire enlist what Davies calls " specialist opinion" by giving it to several colleagues to evaluate and identify the appropriacy of the items.  To measure the reliability,  you can use pilot testing. By administering the instrument to a small group of participants , you can obtain a clear picture of the efficacy of the items. However, you can also use various readability formulas such as the ones you have referred to . For more information, I refer you to the following link, which can introduce to you a number of more recent readability formulas.

Best regards,

R. Biria 

Set up the calibration tone source at equi-distance from the microphone and the sound level meter. This distance should be the same as the distance from the speaker's mouth to microphone during recording (usually ~15cm). The calibration tone is typically a 1000Hz pure tone that can be generated using Audacity (free software). See a more detailed guide by Anders Asplund here: http://www.clinsci.umu.se/digitalAssets/50/50362_andersa0902.pdf

Thank you all

Johnathan maintains an active listserv for his DMDX program: http://www.u.arizona.edu/~kforster/dmdx/list_serv.htm

Tell him I recommended you to join = )

guess Danila means this one https://catalog.ldc.upenn.edu/ldc93s1

This is the current hottest concept in DL after convolutional neural networks (CNN) and deep belief networks (DBL).

It is one of the old issues in a speech recognition research field. That is on the relationship between any speech enhancement technique and the classification accuracy.

As far as I know, both the MSE and PSNR are frequently used for improving the quality of the input. They are known as useful in reducing WER. However, the relationship with recognition accuracy is not directly proportional. 

Enhancing a noisy signal in terms of MSE or PSNR means that you may have a good quality of the input but there is a risk. Sometimes, unexpected artifacts are produced by the speech enhancement techniques and WER can be increased in the worst case.

So, in phonemic classification task, matched condition is more crucial. And in the case of mismatched condition between train and test, MSE and PSNR are somewhat related to WER, but not directly. It is a case-by-case study.    

Yes,

1. is the phrases in passive or active voice and what is proportion of both

2. What kind of pronouns are used in terms of level of politeness

3. What type of phrases: simple, subjectless, complex; with or without images;

4. Any proverbial references?

5. Plain or idiomatic what is the proportion of both

6. Then establish the features as a defining cluster for sub-genres in songs

thanks

This is a good idea however it needs a lot of work, Firstly, we can employ computer vision to identify objects and in a simple case just convert object names into speech. In summary, the components of such a system are available namely; a module for object identification in images and speaking words module. Just integration is needed if not done somewhere.

Thank you everyone for your instant response. Your solutions are highly appreciated.

I believe you can use the same as for English, since both languages do not differ much on their word length (except monosyllables). They correlate similary among levels.

Jorge

 Hi Pasquale, 

I faced a similar situation - i wanted to compare the phoneme frequency of patient (people with aphasia) productions against normal phoneme frequencies. If you want to look at the overall distributions then you could use a kolmorogov-smirnov 2 test which should tell you whether your phoneme distributions differ generally (i.e. whether the zipfian type distribution we see typically in English phoneme frequency is adhered to in your word list). This will not tell you whether specific phonemes are over represented though. You could also use a correlation analysis. If you wanted to analyse each phoneme separately then maybe a chi- square would be helpful.

Hope this is useful,

Emma.

If you were able to find a database of English words in IPA and import that into SIL's Phonology Assistant program (https://www.sil.org/resources/software_fonts/phonology-assistant, its free) you could easily answer your question and also look at it from multiple angles. 

If you can't find a database in IPA, you could import the CMU corpus above, but you would have to define the phonological features of each of the graphemes and digraphemes. If would be a little bit of work, but not too much. You could then easily compare your results against a token frequency list such as found at (http://www.wordfrequency.info/)

Just use word frequency. Better to first get bigrams and tri grams keywords for your data. Then apply VSM model to simply label the clusters. 

If you are focusing on dyslexic learners, I think then see two aspects. The words they can and they words they cannot recognize. These students have diverse reading problems. So keeping 'can' aspect let them attain maximum speed and recognition and calculate the time. Moreover, you have to sort out the extent and variables involved in this process. Otherwise, it will be leading to a generalization. I hope it helps!

What do you mean by noisy reduction?

The dcase is indeed a good reference

depending on what you are looking for you can also have a look to the

sweet-home corpora : http://sweet-home-data.imag.fr/

and

both having sounds captured in a home

Revolutionising communication, tango! the sound of a child's voice by Acapela.

Acapela "childrens voices"

Loquendo TTS Multimedia Package and Voice Creator Now Available

Loquendo emotional TTS

Junichi Yamagishi

 

English speech synthesis of child (under R&D in 2012)

Publication of ongoing research on English speech synthesis of child:

+

McNamara, Lisa. Comparison of Voice Output Types for a Child Using AAC. 2006 ASHA Convention. Missouri State University.

Frequency approach to music, speech and hearing research has never ceased to raise fundamental questions since Ohm's Acoustical Law (1843) and Helmholtz's Resonance Theory (1877). Hussein's question above is just one of millions unanswered questions. . Stephen's reference above to the Nyquist theorem provides adequate answer to the question at hand. If I record musical instruments at say 8 k/Hz sample rate, the sound quality is relatively poor in comparison to the same signal recorded at a higher sample rate. In speech, you might find that different phonemes (particularly fricatives/sibilants) are not well captured at low sampling rates as you'll lose the high frequency components which are critical distinctive features (acoustically speaking).  Thus, if you use a unique sample rate, you might limit yourself to a specific sound source. I adopt Xaver's procedure mentioned above. It is one wise way to discover what sample rate does to the quality of your recordings.

 Thankyou for your kind response

Hi Saket,

You need to train the tagger using Devanagari tagged input data. The following Stack Overflow post explains this:

http://stackoverflow.com/questions/30555426/unicode-tagging-in-python-nltk 

But for convenience and for others I'll repeat the code here. For example using the TNT tagger:

from nltk.corpus import indian

from nltk.tag import tnt

word_to_be_tagged = u"ताजो स्वास आनी चकचकीत दांत तुमचें व्यक्तीमत्व परजळायतात."

train_data = indian.tagged_sents('hindi.pos')

tnt_pos_tagger = tnt.TnT()

tnt_pos_tagger.train(train_data) #Training the tnt Part of speech tagger with hindi data

print tnt_pos_tagger.tag(nltk.word_tokenize(word_to_be_tagged))

When Devanagari (and other Indian languages) were encoded with ISCII, every sign could be directly mapped to a phoneme. Now, with Unicode it may necessitate using context-sensitive rewrite rules equivalent to converting back Unicode to ISCII. I have done the opposite on databases, using the attached table...

Not sure it will help!

Your intuition seems plausible, but it's flawed.

There are phonemes that are considered common, based on their frequency in languages around the world (see http://wals.info/chapter/18 for a discussion of languages that lack common consonants, consonants in the classes of fricatives, bilabials, and nasals - like your examples of /m/ and /n/.). The hypothesis that common phonemes, or at least consonants, are easier to articulate (the Theoretical Issues section in this discussion: http://wals.info/chapter/1).

Vowels are harder to define, and more subject to change, but there are far more types of consonants than vowels (7 is considered a large vowel inventory while the average number of consonants is 22). Languages with small vowel inventories usually include /a/, along with /i/ and /u/ (http://wals.info/chapter/2).

But, if we accept the hypothesis that easier-to-articulate phonemes are less likely to change than phonemes that require complex articulation, we still have to remember that phonemes that are easy to articulate are members of a class of phonemes. As members of a class, what impediments would there be for one easily articulated phoneme to change into another? /m/ and /n/ are both common nasal phonemes with similar articulation, so what would prevent /m/ from changing into /n/?

Then there is the question of origin. Why do phonemes with complex articulation exist at all? The number of consonants in a language ranges from 6 to 120 or so; vowels from 2 to somewhere in the 20s. There is a general inverse correlation between vowels and consonants, languages that have few of one tend to have many of the other, but there is a strong tendency toward the "average" amounts or consonants in the 20s and vowels close to 10. But what reason is there for the outliers on the high end of either type of phoneme? If phonemes that are difficult to articulate tend to change more than those easier to articulate, there should be a trend where phonemes that are difficult to articulate disappear. But after many thousands of years of language use, they persist in many languages.

The sense that they "persist" is also a problem. The implication is that, the earlier you go back in human history, the more complex the inventory of phonemes was. What would be the reason for this? The alternative, though, is that languages developed phonemes that are more complex in articulation from an inventory of easier-to-articulate phonemes.

Complexity of articulation is also a difficult to define, as it often depends on what other phonemes exist in a language. If language users are comfortable with an articulation pattern based on one phoneme, minor variations on that pattern should be relatively easy to adopt.

But this isn't necessarily true, and has an impact on the hypothesis that ease of articulation should influence the presence of a phoneme (if a phoneme is more resistant to change, it is more likely to be present at any particular time of measurement). Languages tend to "skip" similar phonemes that other languages use. There is very little difference in articulation of an aspirated stop consonant and the non-aspirated version, so little that languages may contain both sounds but not consider them phonemic, but allophones. English has an aspirated /p/ and a non-aspirated /p/. The /p/ in "pit" is aspirated, the /p/ in "spit" isn't. So why doesn't English take the easy route and use these sounds that already exist as phonemes, and maintain them as comparatively easy to produce? Hindi does. But it has its own quirks.

The issue is complex, and I haven't investigated it, but have thought about it. There are real problems with the hypothesis. There's no reason for one easy-to-articulate phoneme to not change into another easy-to-articulate phoneme. Difficult to articulate phonemes should disappear over time - but they seemingly shouldn't exist in the first place. Languages don't use "easy" sound variations that already exist in the language as phonemes, while other languages do use them.

Attached you will find a rubric I use for speaking proficiency.  If I were to create a similar rubric for listening comprehension, I might include the following sub-areas:

These are some ideas that come to mind for me about the factors influencing listening comprehension.

Hi,

Please find attached some articles that may be of help to you.

Best Wishes!

Dear Manish,

In the following related papers you can find many relevant feature sets:

(1)

Sedaaghi, M. H. (2009). A comparative study of gender and age classification in speech signals. Iranian Journal of Electrical and Electronic Engineering, 5(1), 1-12.‏

(2)

Lingenfelser, F., Wagner, J., Vogt, T., Kim, J., & André, E. (2010). Age and gender classification from speech using decision level fusion and ensemble based techniques. In INTERSPEECH (Vol. 10, pp. 2798-2801).‏

(3)

Chaudhari, S., & Kagalkar, R. (2012). A Review of Automatic Speaker Age Classification, Recognition and Identifying Speaker Emotion Using Voice Signal. International Journal of Science and Research (IJSR).‏

(4)

Metze, F., Ajmera, J., Englert, R., Bub, U., Burkhardt, F., Stegmann, J., ... & Littel, B. (2007, April). Comparison of four approaches to age and gender recognition for telephone applications. In 2007 IEEE International Conference on Acoustics, Speech and Signal Processing-ICASSP'07 (Vol. 4, pp. IV-1089). IEEE.‏

(5)

Brown, W. S., Morris, R. J., Hollien, H., & Howell, E. (1991). Speaking fundamental frequency characteristics as a function of age and professional singing. Journal of Voice, 5(4), 310-315.‏

Best!

Yaakov

You may find these books useful although quite old.

Lucile Charles & Annie-Claude Motron (2001). Phonetique Progressive du Francais avec 600 exercices. Paris: CLE International.

Lhote Elizabeth (1990). Le paysage sonore d'une langue, le francais. Hambourg: Buske Verlag.

Best wishes

You're looking at more of a psychological phenomenon than an acoustic one. It's similar to the "phonetic restoration" effect that's been studied in the past.  

If you think of  the human auditory system as actively seeking evidence for a particular speech event and finding sufficient evidence for it in the sound then you get the observed phantom percept.  A different version of the effect can be observed in "babble", what people will hear in recordings of superimposed voices.

Actually, if you want to pursue this systematically it could get interesting. For example, can you find sounds that, across listeners, appear to be fertile sources of illusion? What are their characteristics?

ِDear Amir 

No I did not find the answer for NIST 2004

I have not NIST 2008, If you have it, we may have beneficial negotiations.

Please let me know of your decision. 

Dear Volker! I hope to address are very useful articles on the subject of my electronic library.

Vladimir

These models treat observations independently across time, so there should be no problem with train and testing on utterances of different length.

There are few concerns that you should take into account:

1) It is better to make sure you have enough observations (per speaker) for training. If you have several short utterances per speaker, this should be fine

2) UBM training can be severely affected by silences. When you have long utterances in test, they likely contain a lot of silences. A common practice is to use at least energy-based voice activity detector and to score using only voiced frames.

You may find useful the SRE10 Kaldi recipe at least for having some general ideas about data pre-processing.

Hi Ayoub, 

Language mismatch between train and test data can surely affect the performance of the SR system (the error rate can be doubled or even worse if there is a mismatch between languages used to train the GMM, TV matrix and the PLDA model), I think that this paper can be a good starting point [1], it also proposes a phoneme histogram normalization technique to match the phonetic spaces of train and test languages. A possible solution is to use many languages to train your system, i.e. GMM, TV matrix and PLDA (using NIST SRE data) or other databases. Some systems [2] use up to 11 languages. 

--- 

[1] Abhinav Misra, John H. L. Hansen , Spoken language mismatch in speaker verification: An investigation with NIST-SRE and CRSS Bi-Ling corpora

[2] Pavel Matejka et al, Analysis of DNN approaches to speaker identification. 

You don't have to perform i-vector reduction using PLDA/LDA, in that case you can use Cosine Distance Scoring for comparison with your model i-vectors.

TO whom shall i communicate for the Nemours and Torgo database.

Is it available to PhD scholars

Dear Muhammad,

Indubitably, translation loss is inevitable even in the translation of the simplest texts. The main culprit may be said to be related to  the message planned for the source language, original readership. Very often, the message intended by the SL author  may be radically different for the recipients in the target context. Consequently, translator's attempt to create equivalence may miss the original and something may go amiss during the process. In point of fact, translation loss may occur by different causes such as change of register, cultural protection, literal translation of puns, lexical translation leading to ambiguity, use of words lacking direct translation.

Best regards,

R. Biria

The way I did it was inspired by an algorithm I found in a book titled DAFX by Zölzer (2002). Basically you calculate the cross-correlation between the end of the first phoneme and the beginning of the second and find the point where they correlate the most. This is where you want to concatenate the phonemes/diphones/units. Based on this point, you multiply the end of the first phoneme by a decreasing ramp and the beginning of the second phoneme by an increasing ramp and then just add them up. I hope you understand the idea.

Dear Rashmi,

In addition I send to you some papers to play with this subject. To resolve the problem of signal noise, I recommend a band-pass Butterworth filter of order 8, formed by

a high-pass filter of 30 Hz and a low pass filter of 450 Hz; the information can be recorded at a sampling frequency of 50 kHz in vectors of 2 seconds and 50 signals per word. I think that this filter will can help you.

Sincerely yours.

Check out the prosodylab aligner by Kyle Gorman and Michael Wagner

Info and tutorial here:

code on github:

Communicative Competence is the ability to communicate your intended message in socio-cultural context. The choice of grammatical structure is clearly related to the circumstances and is chosen for its appropriacy in those circumstances. Learners need to know how to express a variety of functions and which choices are appropriate in different circumstances

Dear Kanchana Prapphal,

As a new catchword in education, constructivism tends to address how people learn. Accordingly, it suggests that learners construct knowledge for themselves. As a consequence, constructing meaning materializes  learning. The idea is deeply rooted in the philosophies practiced by Piaget and Vygostsky. The latter, in particular, maintains that meaningful  negotiation between the expert and the novice can result in learning. Therefore, to actualize the full potentiality of a learner's linguistic competence, it is  necessary to adopt  an interactive approach in which the affective support provided by the teacher/expert can establish a meaningful rapport helping the individual's progress and learning.

Best regards,

R. Biria 

The keyword in your query appears to be 'specifically' and addressing this I would have to say Constructivism. However, the notion of noticing as proposed by Schmidt (2010) and the Interactional Hypothesis by Long (1981) are key to eventually internalizing the L2 input being acquired and successfully putting the acquired knowledge to use.

Here is the link of one paper which is about vowel recognition in Assamese language...This can be implemented in case of Arabic language as well, however you will have to create your own Arabic database for all the vowel phonemes, hope it will be useful

Thanks once again. I am already checking the sources out. I truly appreciate your interesting in my question.

Best wishes,

Onwu

Dear Jose,

You might have a look the article on the /r/ sound by following the link below:

Hope you find it useful! 

Ali. 

Hi Khaled,

I am looking for free speech databases for speaker recognition (at least more than 80 speakers) ... Can you help me ?

Dear Suciati Anandes,

First, you have to make an inventory of certain targeted statements that carefully pose the suprasegmental features you want to investigate . Subsequently, ask the participants (the selected sample) to read them. Naturally, you record their voices in this particular stage. Finally, ask three native speakers of English to rate them based on the accuracy and appropriacy of their production of suprasegmentals. Good luck with your research.

Best regards,

R.Biria

The L1 synapses / neural net for the L1 are stronger than the neural nets formed for the L2. The L1 knowledge is more entrenched. That's why the L1 plays a toll upon the L2. So, the fact that the L1 is transferring to the L2 (interference is not a good term if you are actually working with language in a cognitive perspective) is an evidence that the L1 synapses are actually not changing. The extensive use of the L2, on the other hand, such as in immigration to an l2-dominant community where the L1 is not frequently used, for example, may lead the L2 to become more robust, making the L1 change.  If thats what you are interested in, I would suggest reading about Language Attrition... Look for Monika Schmid and Kees De Bot paper on it..

Speech Imagery is a rather fresh topic in Brain-Computer Interfaces to my best knowledge, there are few papers published on EEG based Speech Imagery classification, here's a list and links to the ones I'de read:

1- DaSalla et al. "Single-trial classification of vowel speech imagery using common spatial patterns"

2- Matsumuto & Hori "Classification of silent speech using support vector machine and relevance vector machine"

In India, we are working on the Indian language (Tamil & Hindi) speech recognizer  based on the KALDI toolkit.

I would say that the most common use of the word "leaf" is leaves on a tree. However with respect to books, "leaf" is usually used as a verb, i.e. to leaf through a book. Leaf can also mean "hinged flap on the side of a table" is from 1550s" because pages in a book are also hinged.  

So you are correct that "leaf" meaning a page of a book is a secondary meaning and a is metaphor. Metaphors do not always use the most common meanng

Children may be grouped ; familiar contexts or scenarios are to be planned; guided conversation for dialogue delivery in frequently experienced situations will make it really natural and easy for communication. A graded pattern of this kind will ensure quick acquisition of communication-both verbal and nonverbal.

To identify a language spoken by any human being, first of all, we need to verify its specific as well as common linguistic features such as pronunciation including stress, intonation, rhythm and so on.   These common linguistic features are attributed to many languages with little or much variations.  So all the languages are considered as different and unique in themselves.  So,  In order to identify any particular language or languages using language identification programs such as through i-vector or whatever, one should need to incorporate each and every feature of a target language or language for identification or understanding into application (computer programme). Further, To solve the problem of dual language or multiple languages in identification; incorporating most frequently used words and their features from target languages into the software  can be helpful.

Thank you 

Communication Matrix focuses early communicative skills and has been used in studies Before. https://www.communicationmatrix.org/

Have you checked fx Solomon-Rice&Soto "Facilitating Vocabulary in Toddlers Using AAC A Preliminary Study Comparing Focused Stimulation and Augmented Input" for inspiration?

Good luck

In fact that, when we talk about "Mông" in Vietnam, it's presented for Hmong people. In other country, they have another name such as Miáo (Chinese) or Maew (Thai). I'm not sure whether Hmong Do is a subgroup of H'Mong or refer to H'mong. However, if you are know that they are live in Ha Giang, Lao Cai and Bac Ha province, Dong Van and Meo Vae district (which province?), then you can look up on the file I gave you. It presented the H'mong people depend on their location. If you have the trouble with Vietnamese, I can help you sort this data.

I don't get what you mean, I think you misunderstood the question somehow!

Hello. Here are some references you might be able to use. The article "An evaluation of HMM-based Techniques for the Recognition of Screen rendered text" describes the basic idea of applying a HMM to a text recognition problem. It treats the horizontal axis (direction of writing) as the "time axis" that is usually used in HMMs for speech and other time-series based signal processing. You could also try random fields, which also use the HMM formalism.