Open jadestorm opened 1 year ago
Just FYI I am out tomorrow so I won't be able to reply to anything until Monday. =) But if there's more debugging you'd like me to do I can certainly do that.
I can fill in these details:
Corpus structure
What language is the corpus in? English
How many files/speakers? 1/1
Are you using lab files or TextGrid files for input? TextGrid
Dictionary
Are you using a dictionary from MFA? If so, which one? no
If it's a custom dictionary, what is the phoneset? arpabet, based on CMU dictionary
Acoustic model
If you're using an acoustic model, is it one download through MFA? If so, which one? english_us_arpa
If it's a model you've trained, what data was it trained on? n/a
Oh, deleting my previous comment, thought the issue was something else.
Right, so validate by default attempts to train a model on the corpus specified, so the acoustic model path is an optional argument specified via: mfa validate ... --acoustic_model_path english_us_arpa
The file is 167 seconds long. I've confirmed that in 2.2.6 I can align a larger set of recordings but this individual file intermittently gives the same error. I have aligned this file with previous MFA versions. It's used in a tutorial for teaching our students how to use P2FA and MFA. With --beam 1000 it took 2039 seconds to validate but ultimately produced the same error on alignment. Since then it aligned successfully without a larger beam and I don't think anything was different from the previous times it gave the error.
It would be great to be able to align single recordings. I have a web interface for students to plot their own vowel spaces based on a reading passage, and I would like to switch the forced alignment part of it from P2FA to MFA. When I have previously aligned these short wav files with both aligners the MFA alignments have been a lot better than the P2FA ones.
thanks Jeff
Jeff Mielke Professor Linguistics program Department of English North Carolina State University
On Sat, Mar 18, 2023 at 8:56 PM Michael McAuliffe @.***> wrote:
How long is the file? You can try bumping the beam size higher to see if it aligns mfa align ... --beam 1000, but that's usually the solution for longer files as long as the transcripts are accurate.
If it's possible, doing a larger batch than a single file is usually much more accurate (as feature transforms like CMVN and speaker adaptation benefit immensely from them). It does seem like more people are using it for single file alignment, so I'll try to figure out a better solution for the one-off files soon.
— Reply to this email directly, view it on GitHub https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/issues/587#issuecomment-1475054047, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH3Q3BKHVGZW2KGFAAWE73DW4ZKVBANCNFSM6AAAAAAVVTOSP4 . You are receiving this because you commented.Message ID: @.*** com>
Hi Jeff, I realized after posting that the issue in this particular case is not (necessarily) related to the length of the audio file, but rather due to a change in the argument specification for the mfa validate
command. When using argparse
previously, I had acoustic_model_path
as an optional argument, but the new CLI uses click, which takes the opinionated route that all arguments should be non-optional, and optional arguments should be explicitly flagged via options. So the new command would be something like mfa validate ... --acoustic_model_path english_us_arpa
.
Still possible that 167 second files might not align, but it should with a beam of 1000, but the error here is caused by MFA trying to do a test monophone training, which crashes when there's no alignments generated on the second iteration.
Hi Michael.
Thanks. I'm confused about this. mfa validate works for me when I add --acoustic_model_path. But itt seems like the dictionary and the input directory don't need to be explicitly flagged, and when I add --dictionary_path it doesn't work. And when I try to align the same way I validated (explicitly flagging the acoustic model):
mfa align ../files/jeff_vowelplot english_us_arpa --acoustic_model_path english_us_arpa jeff_vowelplot_output/
I get this error message:
╭─ Error ─────────────────────────────────────────────────────────────────────╮ │ Invalid value for 'ACOUSTIC_MODEL_PATH': PretrainedModelNotFoundError: │ │ │ │ Could not find a model named "--acoustic_model_path" for acoustic. │ │ Available: english_us_arpa. │ ╰─────────────────────────────────────────────────────────────────────────────╯
When I take --acoustic_model_path back out, it aligns fine. This is 2.2.6.
Jeff
On Wed, Mar 22, 2023 at 10:03 PM Michael McAuliffe @.***> wrote:
Hi Jeff, I realized after posting that the issue in this particular case is not (necessarily) related to the length of the audio file, but rather due to a change in the argument specification for the mfa validate command. When using argparse previously, I had acoustic_model_path as an optional argument, but the new CLI uses click https://click.palletsprojects.com/en/8.1.x/, which takes the opinionated route that all arguments should be non-optional, and optional arguments should be explicitly flagged via options. So the new command would be something like mfa validate ... --acoustic_model_path english_us_arpa.
Still possible that 167 second files might not align, but it should with a beam of 1000, but the error here is caused by MFA trying to do a test monophone training, which crashes when there's no alignments generated on the second iteration.
— Reply to this email directly, view it on GitHub https://github.com/MontrealCorpusTools/Montreal-Forced-Aligner/issues/587#issuecomment-1480478774, or unsubscribe https://github.com/notifications/unsubscribe-auth/AH3Q3BOBSWI7F77LINXNQZTW5OVP5ANCNFSM6AAAAAAVVTOSP4 . You are receiving this because you commented.Message ID: @.*** com>
Debugging checklist
[x] Have you updated to latest MFA version? yes, 2.2.4 [x] Have you tried rerunning the command with the
--clean
flag? yesDescribe the issue When running a validation, following a set of steps one of our researchers uses for his class, but he (who is far more familiar with all of this) and I (IT staff) are running into the same error and neither of us are sure where to go from here.
For Reproducing your issue TBH I have no idea what the answers to these questions are. -- skipping for now. I CAN link you the lesson: https://phon.wordpress.ncsu.edu/workshops/eng-523-tutorial/part-1/ I did the mfa server init part before this -- and also where I'm at is the very first mfa validate attempt.
Please fill out the following:
Log file
Desktop (please complete the following information):
Additional context Note: we also had problems with the PostgreSQL db setup that we discussed in a different issue. After upgrading to 2.2.4, and after deleting ~/Documents/MFA, I was able to launch the server with mfa server init. I doubt that's relevant but I thought I would mention it. I'm not entirely sure why the auto-start/stop is not working.