[BUG] --speaker_characters results in KeyError

[thealk]

Debugging checklist

[ X] Have you updated to latest MFA version? [ X] Have you tried rerunning the command with the --clean flag?

Describe the issue When running mfa align with the --speaker_characters (-s) flag for speaker adaptation, alignment fails. The KeyError is the id for a single speaker. This error also occurs for mfa validate.

(aligner) thea MFA % mfa align -s 4 --clean /Users/thea/Documents/0_test english_us_arpa english_us_arpa /Users/thea/Documents/1_test_out

File name formats are, for example: oc01_[bla...].wav/.TextGrid, where the first 4 characters denote the speaker ID.

• mfa align and validate work as expected when --speaker_characters flag is omitted.

Command: mfa align --speaker_characters 4 --clean /Users/thea/Documents/0_test english_us_arpa english_us_arpa /Users/thea/Documents/1_test_out

Output:

 INFO     Setting up corpus information...                                      
 INFO     Loading corpus from source files...                                   
   1% ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━ 1/100  [ 0:00:02 < -:--:-- , ? it/s ]
 INFO     Stopped parsing early (0.0847590000000018 seconds)                    
 ERROR    There was an error in the run, please see the log.                    
Exception ignored in atexit callback: <bound method ExitHooks.history_save_handler of <montreal_forced_aligner.command_line.mfa.ExitHooks object at 0x18bfbbbd0>>
Traceback (most recent call last):
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/command_line/mfa.py", line 98, in history_save_handler
    raise self.exception
  File "/Users/thea/miniconda3/envs/aligner/bin/mfa", line 10, in <module>
    sys.exit(mfa_cli())
             ^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/core.py", line 1157, in __call__
    return self.main(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/rich_click/rich_group.py", line 21, in main
    rv = super().main(*args, standalone_mode=False, **kwargs)
         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/core.py", line 1078, in main
    rv = self.invoke(ctx)
         ^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/core.py", line 1688, in invoke
    return _process_result(sub_ctx.command.invoke(sub_ctx))
                           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/core.py", line 1434, in invoke
    return ctx.invoke(self.callback, **ctx.params)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/core.py", line 783, in invoke
    return __callback(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/click/decorators.py", line 33, in new_func
    return f(get_current_context(), *args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/command_line/align.py", line 113, in align_corpus_cli
    aligner.align()
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/alignment/pretrained.py", line 412, in align
    self.setup()
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/alignment/pretrained.py", line 205, in setup
    self.load_corpus()
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/corpus/acoustic_corpus.py", line 1209, in load_corpus
    self._load_corpus()
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/corpus/base.py", line 1288, in _load_corpus
    self._load_corpus_from_source_mp()
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/corpus/acoustic_corpus.py", line 1023, in _load_corpus_from_source_mp
    import_data.add_objects(self.generate_import_objects(file))
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/thea/miniconda3/envs/aligner/lib/python3.11/site-packages/montreal_forced_aligner/corpus/base.py", line 1018, in generate_import_objects
    "speaker_id": self._speaker_ids[u.speaker_name],
                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^
KeyError: 'oc01'

For Reproducing your issue Please fill out the following:

1. Corpus structure
   • What language is the corpus in? English
   • How many files/speakers? Test case has 2 speakers and 5 files each (10 files in total)
   • Are you using lab files or TextGrid files for input? TextGrids
2. Dictionary
   • Are you using a dictionary from MFA? If so, which one? Yes, english_us_arpa
   • If it's a custom dictionary, what is the phoneset?
3. Acoustic model
   • If you're using an acoustic model, is it one download through MFA? If so, which one? Yes, english_us_arpa
   • If it's a model you've trained, what data was it trained on?

Log file Please attach the log file for the run that encountered an error (by default these will be stored in ~/Documents/MFA). 0_test.log

Desktop (please complete the following information):

• OS: [e.g. Windows, OSX, Linux] OSX
• Version [e.g. MacOSX 10.15, Ubuntu 20.04, Windows 10, etc] MacOSX 12.6.1
• Any other details about the setup (Cloud, Docker, etc)

Additional context

• I was messing around with python installs earlier today, but miniconda3 seems to be properly installed and mfa align works as expected, which seem to rule that out. - - When I ran mfa align on the whole data set (all files in a single directory) without specifying speaker adaptation it automatically detected 2 speakers. There are actually ~30 speakers (oc01, oc02, pd01, pd02, etc), but 2 speaker "groups", oc## and pd##. Not sure if this is relevant. A test folder with a couple oc01 and pd01 files only detected a single speaker, so it must have detected a different string for speaker id by default...
• Note that the error was identical on a set of files with longer names too (for example: original file names were caspd_v1_oc01..., where the first 13 characters denoted the unique speaker ID, but also contained project and version info in the name. I simplified the filenames thinking perhaps extra underscores were throwing an error).
• I ran into the same error on another small corpus with 20 speakers, where each file was just the 4 character speaker id (e.g., DM91.wav).
• I have attempted this on two computers; my current laptop (MacOSX 12.6.1) is an M1 Pro, which I know can sometimes generate problems, but I get the exact same output on an older laptop with updated miniconda3 and mfa installs (MacOSX 12.4 with Intel Core i5)
• Thank you so much as always for your insights!

[thealk]

Some additional things I've tried that result in the same errors now also include:

• renaming the transcription tier to match each of the speaker IDs
• -s flag with more mfa commands: mfa align, validate, adapt all have the same issue
• tested on another computer running Mac OSX 13.0 with an Apple M2 Pro chip with a brand new installation of MFA (and after running mfa validate)
• tested on previous mfa version 2.2.5

[thealk]

Another update: Another issue apparently was that speaker adaptation was just not detecting my distinct speakers correctly at all (or at least not in a way I understand). That is, when I ran mfa align (without specifying -s), it would detect 2 speakers consistently, no matter if there were 10 different filenames, or if the files were in distinct directories.

Revising my previous updated because the original solution I came up with (concatenate files and assign one tier per speaker, named with speaker ID) was more complicated than necessary. I erroneously thought my simpler solution, below, wasn't working (just rename your tiers in your original transcript textgrids to match the speaker IDs), but I must have been doing something wrong in my initial tests because now it's working.

The ULTIMATE solution, which I am now happy with, is to make sure your transcription tier in EACH FILE has a name corresponding to the speaker. For a larger corpus with many files per speaker, this means every file for a given speaker has the same tier name, but no two speakers have the same tier name. For a smaller corpus with one file per speaker, each file just contains a tier name that matches the file name.

For anyone working in Praat, I wrote a tiny script to rename tiers according to speaker_chars in the filenames: https://github.com/thealk/PraatScripts/blob/master/mfa_prep/rename_tiers_to_speaker.praat

I am leaving this issue open because the original problem, namely that the -s flag consistently results in an error, is still an unresolved issue. However, this workaround is easy enough to implement.