ORTOptimizer for wav2vec2-bert

[aconeil]

What does this PR do?

Add wav2vec2-bert to the list of possible models for optimization

Fixes #2221

Who can review?

@IlyasMoutawwakil

[IlyasMoutawwakil]

please add it to testing as well

[eingrid]

Should'nt we also add
"wav2vec2-bert":NormalizedTextConfig in here:

optimum/optimum/utils/normalized_config.py

Lines 232 to 236 in abd7da3

	# Contribution note: Please add new models in alphabetical order
	_conf = {
	"albert": NormalizedTextConfig,
	"bart": BartLikeNormalizedTextConfig,
	"bert": NormalizedTextConfig,

[IlyasMoutawwakil]

@eingrid yes that as well

[aconeil]

Wav2Vec2-Bert actually uses mel-spectrograms as the input, which is closer to the SpeechT5 Input than the wav2vec inputs.
I tried the suggested changes, with both NormalizedTextConfig and T5LikeNormalizedTextConfig and wasn't successful. I then tried adapting optimum/exporters/onnx/model_configs.py to add a specific class for the model, but receive an error when using the DummyAudioInputGenerator (since the input is different) and the DummySpeechT5InputGenerator (since there aren't speaker embeddings).
Do you have any suggestions?

[aconeil]

I tried with Speech2TextDummyAudioInputGenerator as well today without success. The input for Wav2Vec2Bert is input_features with a size of [1, x, 160].