No matter which combination of pasted code from a Facebook/wav2vec2 blog I use the model outputs are always gibberish and WER does not improve in training. This is especially strange since earlier last week everything seemed to be working (at least when using the pre-trained model for inference), but now without changing anything all I get is gibberish. I even created a whole new python install on a different machine and got the same results. Here is an example of some of the code I pasted and the results I am getting:

import soundfile as sf
import torch
from datasets import load_dataset
from transformers import Wav2Vec2ForCTC, Wav2Vec2Processor

# load pretrained model
processor = Wav2Vec2Processor.from_pretrained("facebook/wav2vec2-base-960h")
model = Wav2Vec2ForCTC.from_pretrained("facebook/wav2vec2-base-960h")


librispeech_samples_ds = load_dataset("patrickvonplaten/librispeech_asr_dummy", "clean", split="validation")

# load audio
audio_input, sample_rate = sf.read(librispeech_samples_ds[0]["file"])

# pad input values and return pt tensor
input_values = processor(audio_input, sampling_rate=sample_rate, return_tensors="pt").input_values

# INFERENCE

# retrieve logits & take argmax
logits = model(input_values).logits
predicted_ids = torch.argmax(logits, dim=-1)

# transcribe
transcription = processor.decode(predicted_ids[0])

print(transcription)

====================
Output below

kuk_aktkckwrk[PAD]klk_dktkcwkrk_arthcrkskqkbaktkdcrkbykrthcru_zkzdcrmkdksakakckarks zr'kckrkswcrfkdkskzkrktkbrk'kckdkmkbkucrkhk_akrkfkbakqkckdkrk