Add zipformer recipe for audio tagging

[marcoyang1998]

This PR adds a recipe for audio tagging using Zipformer. We use AudioSet as the training set.

Without using knowledge distillation, the Zipformer-M model achieves mAP value of 45.1. Using knowledge distillation could further improve the results (will add support for this in another PR)

• Add usage example
• Upload the pre-trained model

[marcoyang1998]

You can use the pretrained model to predict the audio events of an audio clip through the following command:

python zipformer/pretrained.py \
    --checkpoint zipformer/exp_at_as_full/pretrained.pt \
    --label-dict downloads/audioset/class_labels_indices.csv \
    downloads/audioset/eval/wav/__p-iA312kg_70.000_80.000.wav

You should see the following output:

2024-03-20 18:40:35,611 INFO [pretrained.py:175] Reading sound files: ['downloads/audioset/eval/wav/__p-iA312kg_70.000_80.000.wav']
2024-03-20 18:40:35,616 INFO [pretrained.py:181] Decoding started
Top 5 predicted labels of the 0 th audio are ['Cat', 'Animal', 'Domestic animals, pets', 'Meow', 'Caterwaul'] with probability of [0.9497291445732117, 0.8830551505088806, 0.8787704110145569, 0.2819702923297882, 0.18947316706180573]

[marcoyang1998]

The pretrained model can be found here: https://huggingface.co/marcoyang/icefall-audio-tagging-audioset-zipformer-2024-03-12#/

[marcoyang1998]

Since there is no official way of downloading AudioSet now, this recipe assumes that you have your own version of AudioSet. The required structure of the dataset will be given in prepare.sh (will add in the next commit).

[marcoyang1998]

Added jit_pretrained.py. The script produces the same results as pretrained.py.

[marcoyang1998]

@csukuangfj Can we merge this now?

[csukuangfj]

Thanks! Just left a minor comment.