[BUG] KaldiFatalError for inputs of particular sizes

[scottbreyfogle]

Debugging checklist

[ ] Have you updated to latest MFA version?
I may be two minor versions behind, but I'm 90% sure this bug should be in the current version.
[ ] Have you tried rerunning the command with the --clean flag?
I'm using MFA library functions directly

Describe the issue
Technically, this is a bug with Kaldi. I plan to file a bug with them as well, but I think there is an easy mitigation for Kalpy. When an input audio file is of particular size[1], MFCC and Pitch values produced are of different lengths. When Kalpy attempts to use paste_feats to put the two together, the tolerance of 0 causes an empty array to be generated. When those features are then used, it causes a KaldiFatalError because the input is empty.

1. The size appears to be any size where (length - 400) % 160 > 156. This is related to the frame size (400) and stride (160) of the feature calculation at 16khz.

Mitigation
I think that setting the second argument of this command to 1 to account for a mismatch in frame length of 1 would avoid the error. I'm sorry that I don't currently have the time to fully test and validate this fix.

For Reproducing your issue
Here's a minimal repro of the issue:

import tempfile

import kalpy.data
import kalpy.utterance
import numpy
import soundfile
import montreal_forced_aligner.models


acoustic_model = montreal_forced_aligner.models.AcousticModel('local_pl_model.zip')
with tempfile.NamedTemporaryFile(suffix='.wav') as audio_file:
    soundfile.write(audio_file, numpy.array([0.0] * 1038), samplerate=16_000)
    duration = soundfile.info(audio_file.name).duration
    segment = kalpy.data.Segment(audio_file.name, begin=0.0, end=duration)
    utterance = kalpy.utterance.Utterance(segment, '')
    feats = utterance.generate_features(
      acoustic_model.mfcc_computer,
      acoustic_model.pitch_computer,
      lda_mat=acoustic_model.lda_mat)

Please fill out the following:

1. Corpus structure
   • What language is the corpus in?
     This happens for Polish, but not English. I have not triangulated exactly what options trigger this code path.
   • How many files/speakers?
     It's reproducible with a single file.
   • Are you using lab files or TextGrid files for input?
     Wave and text files
2. Dictionary
   • Are you using a dictionary from MFA? If so, which one?
     Yes. I believe this occurs in acoustic feature computation before the dictionary is relevant, but we're using https://github.com/MontrealCorpusTools/mfa-models/releases/download/dictionary-polish_mfa-v2.0.0a/polish_mfa.dict
3. Acoustic model
   • If you're using an acoustic model, is it one download through MFA? If so, which one?
     Yes, we're using https://github.com/MontrealCorpusTools/mfa-models/releases/download/acoustic-polish_mfa-v2.0.0a/polish_mfa.zip

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

Desktop (please complete the following information):

• OS: [e.g. Windows, OSX, Linux] Linux
• Version [e.g. MacOSX 10.15, Ubuntu 20.04, Windows 10, etc] Ubuntu 22.04.1
• Any other details about the setup (Cloud, Docker, etc)

Additional context
Add any other context about the problem here.