Speaker Adaptation Insistence

[BeckettFrey]

version: 8db3afd

Problem:

When use_speaker_adaption=False for align_dirs()

This error propagates:

Alignment failed: Error(s) in loading state_dict for Linear:
size mismatch for weight: copying a param with shape torch.Size([42, 1280]) from checkpoint, the shape in current model is torch.Size([42, 768])
Works fine for that same dataset when use_speaker_adaption=True

Proposed Solution

Not sure right now.

[pkadambi]

I think this may be an issue with an updated model, I'll take a look at this.

You can try and add redownload_model=True as an argument

Wav2TextGrid/src/Wav2TextGrid/wav2textgrid.py

Line 82 in 8db3afd

aligner = xVecSAT_forced_aligner(args.aligner_model, satvector_size=sat_size)

Wav2TextGrid/src/Wav2TextGrid/wav2textgrid.py

Line 110 in 8db3afd

aligner = xVecSAT_forced_aligner(aligner_model, satvector_size=sat_size)

[BeckettFrey]

@pkadambi I tried redownloading the model and it was still the same error, in the meantime I added an argument check which requires use_speaker_adaptation to be True when using the default model (your model). This enforces speaker adaptation by default by raising a ValueError ->

ValueError: The default aligner model requires speaker adaptation. Please enable speaker adaptation or provide a different aligner model.

And since no compatble model exists, a value error is raised regardless of the chosen model until verification ->

ValueError: Disabling speaker adaptation is not supported as of now.

Here's the commit -> 90576d7

I suspect this is not super urgent to add as a feature since speaker adaptation will always be True for our use case. Let me know if you want me to pursue this further.