When rendered as part of a graph's axis label or title, Devanagari text renders incorrectly: the "matras" get affixed to the wrong letters (see the attached image, and note the difference between the string literal in the code and the rendered text in the title of the graph).
I'm going to start by apologizing that I know nothing more about Devanagari rendering than this issue and a very quick skimming of it's wikipedia entry.
It looks like the matras is a vowel diacritic that combines with the letter to its left, where as the other vowels appear below or to the right (ref https://en.wikipedia.org/wiki/Devanagari#Vowel_diacritics). Do these render correctly or are all vowels broken?
It also looks like there are also ligatures between some consonants, do those render correctly?
I am not sure off the top of my head if we are doing the detailed text layout here or freetype is, but it is clear that the diacritic is not being handled correctly.
A couple questions:
does it work with other fonts? We have seen issues where the fonts have errors in their metadata about the size of the glyphs which results in incorrect layouts
does it work if you use https://github.com/matplotlib/mplcairo ? That under the hood uses a different text layout engine which does RTL correctly so it is likely it will also handle this case correctly.
No apology necessary, thank you for looking into this issue. :)
Your understanding is correct. All other vowels are fine. 'क' + 'ि' is "special" in the sense that the 'ि' needs to be rendered to the left of the 'क'.
The problem persists when I use a different (Devanagari-compatible) font.
I don't know how to use mplcairo, but you should be able to run the experiment using the code snippet I shared above. On Macs that font is built-in.
The consonant-consonant ligatures "work" in the sense that they are not wrong, but not ideal. The Devanagari script has special ligatures for each pair of consonant-ligatures, but there is also a generic way to write them by adding the "halanth" diacritic ([्]) to the first consonant. Matplotlib seems to just use the generic form everywhere. See below (compare to the respective string literals in Python):
Looking at the last example , it looks like the shaping on 'ि' is such that it goes over a the whole ligature?
Unfortunately, I do not think this is going to be an easy fix (it is possible that we are clipping something to be positive that should be negative to move the diacritic to the left), but I suspect that our text system is not currently up to this.
You're right, the 'ि' should loop around the entire consonant cluster, though there are exceptions. For example, in मूर्ति the little arc above the त that makes it र्त should be "outside" the matra.
I do need to produce graphs with (correct) Devanagari labels on them. Do you have recommendations on how to proceed?
importmatplotlibmatplotlib.use('module://mplcairo.qt')
matplotlib.rc('font', family='Noto Sans Devanagari') # I already had NoTo installedimportmatplotlib.pyplotaspltplt.title('किसान')
I am not sure how hard it would be to get mplcairo to play nice with inline, but if you need to save the output you are good to go!
If you definitly need images in a notebook, I think the way to go is module://mplcairo.base and then rely on the repr of a Figure is a png of it it.
Activity
tacaswell commented on May 20, 2022
I'm going to start by apologizing that I know nothing more about Devanagari rendering than this issue and a very quick skimming of it's wikipedia entry.
It looks like the matras is a vowel diacritic that combines with the letter to its left, where as the other vowels appear below or to the right (ref https://en.wikipedia.org/wiki/Devanagari#Vowel_diacritics). Do these render correctly or are all vowels broken?
It also looks like there are also ligatures between some consonants, do those render correctly?
I am not sure off the top of my head if we are doing the detailed text layout here or freetype is, but it is clear that the diacritic is not being handled correctly.
A couple questions:
tacaswell commented on May 20, 2022
One more detail: we are rendering the text in the order it comes in in the string:
In [1]: list('किसान') Out[1]: ['क', 'ि', 'स', 'ा', 'न']kach commented on May 20, 2022
Hi Thomas,
No apology necessary, thank you for looking into this issue. :)
Your understanding is correct. All other vowels are fine. 'क' + 'ि' is "special" in the sense that the 'ि' needs to be rendered to the left of the 'क'.
The consonant-consonant ligatures "work" in the sense that they are not wrong, but not ideal. The Devanagari script has special ligatures for each pair of consonant-ligatures, but there is also a generic way to write them by adding the "halanth" diacritic ([्]) to the first consonant. Matplotlib seems to just use the generic form everywhere. See below (compare to the respective string literals in Python):
tacaswell commented on May 20, 2022
Looking at the last example , it looks like the shaping on 'ि' is such that it goes over a the whole ligature?
Unfortunately, I do not think this is going to be an easy fix (it is possible that we are clipping something to be positive that should be negative to move the diacritic to the left), but I suspect that our text system is not currently up to this.
kach commented on May 20, 2022
Hi Thomas,
You're right, the 'ि' should loop around the entire consonant cluster, though there are exceptions. For example, in मूर्ति the little arc above the त that makes it र्त should be "outside" the matra.
I do need to produce graphs with (correct) Devanagari labels on them. Do you have recommendations on how to proceed?
anntzer commented on May 20, 2022
Unless I am mistaken, this is basically a duplicate of #8765.
tacaswell commented on May 20, 2022
The fastest thing is probably to use mplcairo
I am not sure how hard it would be to get mplcairo to play nice with inline, but if you need to save the output you are good to go!
If you definitly need images in a notebook, I think the way to go is module://mplcairo.base and then rely on the repr of a Figure is a png of it it.
QuLogic commented on Apr 10, 2026
The text-overhaul branch is now merged to main.