I am new to Praat and trying to read the spectrograms of different sound files using these videos. I am learning to read the specrograms using the UI of Praat but I am unable to find any way through which I can read the slices using Praat script.

Someone please guide me about any way through which I read and analyse the spectrograms of sound files.


If I understand your goal (unassisted speech to text conversion, using Praat), this is not doable out of the box, but you might eventually develop the tools for doing this. An alternative interpretation is simply "how do I find phonetic landmarks, using Praat and my ears?", which is actually quite doable, if you know basic phonetics.

Before you get to using Praat, you need to decide what the scope of your project is: any language, or a specific language? If you can't do it for a specific language (and then for 6 specific languages), you can't do it for all languages. If you know what the language is, then you can use that information to rule out possible interpretations of the signal. Within Praat, you have basically two problems, segmentation and identification (the latter assumes the former). The easiest case for segmentation in picking segments in a VCV sequence with voiceless stops, then fricatives, then voices stops, voiced fricatives, nasals, laterals, rhotics and glides. The reason why VtV is easiest is because a big change in the signal matches segment boundaries. The endpoint of the process is parsing similar vowel sequences such as [eɛ, ɨi], where it's impossible to find a crisp segment boundary (and the usual approach is to divide the contested zone in two parts).

| improve this answer | |
  • Thanks you so much it helped a bit but what exactly I want to do is to divide the spectrogram into phoneme segments and then process those segments for phoneme assessment Is there any way of doing this ? – Itban Saeed Oct 25 '16 at 6:36
  • No, the point is that that you're talking about unlikely future technology. In 5 years you might be able to implement what I suggested. Then in 60 years you might be able to do phonemic analysis as you want, and that assumes unlikely progress in the cross-linguistic phonetic database. The way to go is pick a specific language, and use statistical methods with a pre-analyzed corpus. – user6726 Oct 25 '16 at 15:36

From the comments I conclude that the CLARIN service WebMAUS hosted by Bayerisches Archiv für Sprachsignale, Munich, is what you want to use. This service takes a spectogram and a transcription as inputs and outpus a segmentation.

On the linked website, you will also find tutorials on how to use WebMAUS.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.