Updating model card for wav2vec2

[AshAnand34]

What does this PR do?

This pull request updates the wav2vec2 documentation to improve readability, enhance usability with examples, and reorganize the structure for better navigation. Key changes include the addition of usage examples, restructuring of the API reference, and updates to the model overview.

Documentation Enhancements:

• Replaced the abstract with a concise explanation of Wav2Vec2 and added links to Hugging Face Hub for checkpoints and examples.
• Added Python code examples demonstrating how to use the Wav2Vec2 model for automatic speech recognition and audio classification.
• Included guidance on using Flash Attention 2 for faster inference and quantization techniques for memory optimization.

Structural Improvements:

• Reorganized the API reference section by converting model components (Wav2Vec2Config, Wav2Vec2CTCTokenizer, etc.) into subsections for better navigation.

These changes make the documentation more user-friendly and accessible, especially for developers new to Wav2Vec2.

#36979

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@stevhliu

[stevhliu]

Thanks!

Please apply some of these same changes (such as changing the API reference) to your other PRs 🤗

[stevhliu]

Suggested change

	[Wav2Vec2](https://huggingface.co/papers/2006.11477) is a self-supervised learning framework for speech representations that masks speech input in the latent space and solves a contrastive task over quantized latent representations. It's like having a speech expert that learns the patterns of human speech from raw audio alone, then fine-tunes on transcribed speech to achieve remarkable accuracy.
	[Wav2Vec2](https://huggingface.co/papers/2006.11477) is a self-supervised learning framework that makes pretraining audio models easier and more efficient. It is pretrained on large amounts of unlabeled data where certain chunks of audio are masked and the model must predict it. The model is fine-tuned on a much smaller labeled dataset and demonstrates it can still achieve competitive results.

[stevhliu]

Suggested change

	You can find all the original [Wav2Vec2](https://huggingface.co/models?search=wav2vec2) checkpoints on the Hugging Face Hub.
	You can find all the original Wav2Vec2 checkpoints in the [Wav2Vec2.0](https://huggingface.co/collections/facebook/wav2vec-20-651e865258e3dee2586c89f5) collection.

[stevhliu]

Suggested change

	> [!TIP]
	> Click on the Wav2Vec2 models in the right sidebar for more examples of how to apply Wav2Vec2 to different speech recognition and audio classification tasks.
	> [!TIP]
	> This model was contributed by [patrickvonplaten](https://huggingface.co/patrickvonplaten).
	>
	> Click on the Wav2Vec2 models in the right sidebar for more examples of how to apply Wav2Vec2 to different speech recognition and audio classification tasks.

[stevhliu]

import torch
from transformers import pipeline

pipeline = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h", device=0, torch_dtype="auto")
pipeline("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/1.flac")

[stevhliu]

import torch
from datasets import load_dataset
from transformers import AutoProcessor, AutoModelForCTC

dataset = load_dataset("hf-internal-testing/librispeech_asr_demo", "clean", split="validation")
dataset = dataset.sort("id")
sampling_rate = dataset.features["audio"].sampling_rate

processor = AutoProcessor.from_pretrained("facebook/wav2vec2-base-960h")
model = AutoModelForCTC.from_pretrained("facebook/wav2vec2-base-960h", torch_dtype="auto", device_map="auto", attn_implementation="sdpa")

inputs = processor(dataset[0]["audio"]["array"], sampling_rate=sampling_rate, return_tensors="pt").to("cuda")

with torch.no_grad():
    logits = model(**inputs).logits
predicted_ids = torch.argmax(logits, dim=-1)

transcription = processor.batch_decode(predicted_ids)
transcription[0]

[stevhliu]

No need to include a quantization example since the model isn't that large

[stevhliu]

You can remove all the notes except for:

- Wav2Vec2 is a speech model that accepts a float array corresponding to the raw waveform of the speech signal.
- If you have a `head_mask`, you must use the eager attention implementation for it to work.

   ```py
   from transformers import AutoModelForCTC

   model = AutoModelForCTC.from_pretrained("facebook/wav2vec2-base-960h", attn_implementation="eager")

[stevhliu]

Remove all the resources except for links to external ones (ie, not guides or scripts you can find in the docs)

[stevhliu]

You don't need to change any of the API references

[AshAnand34]

Thanks!

Please apply some of these same changes (such as changing the API reference) to your other PRs 🤗

I am out of town at the moment. I will make the changes when I am back.