Re: [AUDITORY] Tool for automatic syllable segmentation

[Sarah Hawkins <0000033950e3b3ae-dmarc-request@xxxxxxxxxxxxxxx>]

    I can't advise on current software, but I can suggest things that
    are worth considering in corpus design, and that may help the
    community to give you better answers if you could provide them with
    the information.
    The distinctions I think matter fall into two types of information:
    the phonetic composition of the syllables, and the type of speech.
    Though the language/dialect also matters. Where it matters, I've
    assumed French.
     I've also assumed little knowledge of phonetics and
    acoustic-phonetics; forgive me if I'm wrong.
    
    Phonetic composition:  What do you mean by 'consonant'?  
                    Consonants cover 4 main classes of articulation type
    (plus several other types):  
       stops ( e.g. p t k    b d g );      fricatives ( e.g. f s ʃ x χ     v z ʒ ɣ
    ʁ
    );     nasals ( e.g. m n );      approximants ( e.g. w j l and some
    types of r )
    
    Many people tend to think only of stops and maybe fricatives, and
    neglect the others, but functionally they are all consonants, and
    segmenting them involves very different issues. So it's helpful to
    give a bit more information.
     If you ONLY mean stops and fricatives, then a single word to use is
    'obstruent'.
     Obstruents are usually the easiest to segment.
    
    Next, each member of each class may be so-called voiced or
    voiceless.
    In the above lists of obstruents, the first items in the parentheses
    are classed as voiceless and the second set as voiced.
    The nasals and approximants I've listed above are all classed as
    voiced. Voiceless ones exist.
    
    Importantly, how voiced and voiceless consonants are pronounced in
    any given language can be very different, and involve a great deal
    more than whether or not the larynx is vibrating. That is, used as
    above, these terms voiced and voiceless are phonological: a general
    descriptive term with no necessary basis in acoustic reality
    for a particular language or speaking style. 
    Example: French voiced obstruents typically contain significant
    vocal fold vibration, whereas British English voiced obstruents can
    include it, but often contain very little or none at all.
    
    Phonation is the term indicating existence of vocal fold vibration.
         So French voiced obstruents are typically phonated; English
    ones are not, or not phonated throughout their duration.
    Vowels are normally phonated (voiced) but not always; whisper is the
    common word for speech whose vowels are all unphonated. 
    
    Syllables that alternate voicelessness and voicing (non-phonation
    and phonation) tend to be easiest to segment: you seek silence or
    aperiodic noise, vs. periodicity (or an f0). Examples: /ki, sa/ (as
    in French qui, comme ça).
    
    Type of speech:  How natural is your speech data, and how
    much variation is there in it?
             Is it synthetic or spoken by a human or humans?    If
    human, were the speakers trained professional speakers or untrained?
             Were their instructions clear, or where they just asked to
    speak fast? (e.g. were care or clarity mentioned in the
    instructions?)
             How was the speed achieved? E.g. by speeding up careful
    speech algorithmically, or by making it fast in the first place?
    
    For your CV syllables, slow, clear, speech containing only voiceless
    obstruents and (phonated) vowels, with little or no variation,
    should be easy to segment both automatically and by
    minimally-trained humans. 
    The same, containing voiced obstruents (in French) or any nasals and
    approximants, will likely be less accurate compared with the human
    ear.
    
    But when spoken fast by untrained speakers with no instruction to
    maintain clarity, even voiceless obstruents are likely to be hard to
    segment automatically, for two main reasons: clearly-alternating
    phonation and silence/frication (aperiodic noise) may be lost, and
    clear phonetic distinctions may be lost because pronunciation
    changes -- stops become more like fricatives, and both stops and
    fricatives may become more like approximants, which are basically
    just rather short vowels.
    
    Finally, though I've limited experience of segmenting French, if
    your algorithms include relative amplitude as a criterion, then
    syllables like /di, vi, gu, vu/ (French dit, vi(sa), goȗt, vous) may be mis-segmented
    because the amplitude of the periodicity may be greater in the
    consonant than in the vowel.  I have seen this in some French, but I
    don't know if it's the norm. I'd be delighted if my French
    phonetician colleagues could tell us!
    
    In summary: you're likely to be best off if you can use
    syllables that are easy to segment. But if your spoken syllables are
    inherently hard to segment, for whatever reason, then you will need
    a more sophisticated algorithm that might use, for example, spectral
    shape as well as relative amplitude and/or existence of an f0. Or an
    algorithm that you can train yourself.
    
    Hope this helps, and apologies it doesn't!
    Sarah Hawkins

On 2024-09-18 16:52, Rémy MASSON wrote:

Hello AUDITORY list,

 

We are attempting to do automatic syllable segmentation on a collection of sound files that we use in an experiment. Our stimuli are a rapid sequence of syllables (all beginning with a consonant and ending with a vowel) with no underlying semantic meaning and with no pauses. We would like to automatically extract the syllable/speech rate and obtain the timestamps for each syllable onset.

 

We are a bit lost on which tool to use. We tried PRAAT with the Syllable Nuclei v3 script, the software VoiceLab and the website WebMaus. Unfortunately, for each of them their estimation of the total number of syllables did not consistently match what we were able to count manually, despite toggling with the parameters.  

 

Do you have any advice on how to go further? Do you have any experience in syllable onset extraction?

 

Thank you for your understanding,

 

Rémy MASSON

Research Engineer

Laboratory "Neural coding and neuroengineering of human speech functions" (NeuroSpeech)

Institut de l’Audition – Institut Pasteur (Paris)