parselmouth.praat.call stuck in "Change gender"

[auspicious3000]

Hello,

I have multiple processes of a function calling the parselmouth.praat.call(...). After certain number of iterations, the program hangs forever. Is this because the praat.call() also spawns another subprocess? What would be the recommended way to use praat.call in multiprocessing?

Thanks!

[YannickJadoul]

Hi, @auspicious3000!

parselmouth.praat.call is not spawning a new subprocess. It's really just calling into Praat, so I'm surprised this is happening.

But it definitely sounds like a bug, and should not happen. Do you perhaps have a small script that reproduces the problem and that you can share? This way, I could have a hands-on look at the problem.
(If you don't like sharing it publicly on GitHub but do want to share it with me, you could also send me a private email.)

[YannickJadoul]

What would be the recommended way to use praat.call in multiprocessing?

Just to be extra clear: in principle, as far as I know, nothing should stop multiprocessing use of praat.call, and I don't know of anything you should do differently that usual.

[auspicious3000]

After more exploration, I find it does hang forever in some cases, but the cause is not multiprocessing. I will let you know as soon as I find more details. Sorry for the confusion.
So far, the situation is the following:
I was iterating a very large speech dataset using praat.call("change geneder") for each utterance. However, sometimes the praat function raises errors. So I used try and catch to skip those utterances. After a certain number of failures, the function call hangs.

[YannickJadoul]

Thank you for the update!

I perfectly understand it's hard to figure out the actual problem, whenever multiprocessing is/was involved :-)
But indeed, you are right that this is still something that should not happen. If you would manage to figure out a (minimal) reproducing example (or anything else that gives more information), I'd be very interested to dig into it and figure out why it hangs. Again, it should not hang, even after a bunch of failures in try-catch statements, so you did find some sort of bug, I think.

[auspicious3000]

@YannickJadoul I just managed to make a minimal reproducing example. I send it to your email because there is data. Thanks for your help!

[YannickJadoul]

@auspicious3000, that's great! Thank you very much; I did gt your email :-) I will see if I can figure out where things go wrong!

[YannickJadoul]

So, I already sent @auspicious3000 an email about this, but quickly a summary to get something in the public record of the repository:

• The issue seems to be in Praat's source code, as running "Change gender" on the file @auspicious3000 sent me, with the same parameters in Praat's GUI also results in an infinite loop.
• According to my debugger, the infinite loop happens at https://github.com/praat/praat/blob/586bc8b05a9d3d3dcff45a3da63882516773c13c/fon/Manipulation.cpp#L418, as endOfTargetNoise is +inf.
• endOfTargetNoise is +inf, because of a division by zero at https://github.com/praat/praat/blob/586bc8b05a9d3d3dcff45a3da63882516773c13c/fon/Manipulation.cpp#L399.

[YannickJadoul]

This will be fixed in Praat 6.2.02, according to the answer here: praat/praat#1926

I've looked up the fix, and it's in this commit: praat/praat@f8b74b5
Maybe looking into this fix will teach us how to avoid the issue upfront?

I should also start working on Parselmouth 0.5.0, I guess, which will include a new version of Praat and this fix.

[auspicious3000]

@YannickJadoul That sounds reasonable! Looking forward to the newer version of parselmouth.

[YannickJadoul]

I got confirmation from @auspicious3000 (per email) that the solution proposed here works: praat/praat#1926 (comment); many thanks to @PaulBoersma!

Minor note to add: the Pitch object should be calculated with parselmouth.Sound.to_pitch with time_step argument set to 0.8 / pitch_floor, to exactly match what Praat does inside of the "Change gender" action.

[YannickJadoul]

I'll keep this open until we have a next minor Parselmouth release, which will include an update of Praat and the fixes in 6.2.02.

[auspicious3000]

Should the argument time_step be set to 0.8/pitch_floor ?

[YannickJadoul]

Should the argument time_step be set to 0.8/pitch_floor ?

Agh, you're right. I'm very sorry. Too hasty in typing that reply.

I edited the post. Thanks a lot!

[maitycyrus]

This bug still exists in version 0.4.3. "Change gender" does not throw an exception, it still is caught in an infinite loop.

[YannickJadoul]

@maitycyrus, yeah, I still haven't had the time to finish up 0.5.0.
However, Parselmouth's internal Praat version on the latest development version has been updated. So if you urgently need this and don't mind trying out the development version, you can download the wheels from the latest build: https://github.com/YannickJadoul/Parselmouth/actions/runs/3782438748