Montreal-Forced-Aligner icon indicating copy to clipboard operation
Montreal-Forced-Aligner copied to clipboard

MFA folder not created/beam too low [BUG]

Open kaitowen opened this issue 2 years ago • 4 comments

I have been trying to run the forced aligner using mfa validate (and later mfa align just to see if it would work) because I'd like to see if I could force align my corpus of music files. Any time I run mfa validate or mfa align I get a message that the beam is too low (beam=10 retry beam=40) with the suggestion to add --beam=100 and --retry_beam=400. I add these in as well as --clean and I still repeatedly get that the beam is too low (beam=10 retry beam=40). I have also realized that I don't have MFA as a folder that was generated anywhere on my laptop and think that could somehow be part of the issue.

  1. Corpus structure My corpus is in French (specifically Québec French) and is specifically of singing. I currently am trying to test if forced alignment would even work on my song corpus so I've been testing with just one file with one speaker (3-4 minutes in length). I am using TextGrid.
  2. Dictionary I am using french_prosodylab as my dictionary but have adjusted the phone set to exclude x and G.
  3. Acoustic model I am using french_prosodylab as well for my acoustic model but have also tried french_qc.

Log file I don't have a log file because I don't have an MFA folder in my files.

Desktop

  • OS: Windows (with access to Linux functionality, specifically Ubuntu, if needed)
  • Version Windows 10

kaitowen avatar Jul 05 '22 00:07 kaitowen

I also get this error for my own custom aligner. I tried to align a single 13second audio file. At first I try to align without specifying a beam but it fails:

%  mfa align CORPUS DICT MODEL OUTPUT_DIRECT
... ERROR - There was an error in the run, please see the log.
NoAlignmentsError:
  There were no successful alignments for 1 utterances.
  The current set up used a beam of 10 and a retry beam of 40.
  You can try rerunning with a larger beam (i.e. "mfa align ... --beam=100 --retry_beam=400").

But then I specify a larger beam, but it seems that MFA just doesn't use my beam specification.

%  mfa align CORPUS DICT MODEL OUTPUT_DIRECT --beam=100  --retry_beam=400 --clean
... 
ERROR - There was an error in the run, please see the log.
NoAlignmentsError:

  There were no successful alignments for 1 utterances.
  The current set up used a beam of 10 and a retry beam of 40.
  You can try rerunning with a larger beam (i.e. "mfa align ... --beam=100 --retry_beam=400").

Linked are the wav file, transcript, model, and dictionary.

It seems that mfa align just ignores the beam settings that I manually provide.

jhdeov avatar Nov 13 '22 03:11 jhdeov

I was able to solve this issue by not including the "=" despite the fact that it is written in what the error message says to add.

So, I specified the beam with ...--beam 100 --retry_beam 400

Hope that helps!

kaitowen avatar Nov 13 '22 03:11 kaitowen

You're right! Thanks :D That's an odd glitch.

jhdeov avatar Nov 13 '22 03:11 jhdeov

@kaitowen Hi, I removed the "=" sign as you suggested but unfortunately, it still gives the error that my beam is 10 and I need to retry with 100. Any suggestions? Thanks

skanda1005 avatar Jan 06 '23 05:01 skanda1005