## OpenL3
OpenL3 is an open-source Python library for computing deep audio and image embeddings.
Using pysoundfile library for audio, scikit-image library for image, and moviepy for video.

### Audio Embeddings Computation

import openl3
import soundfile as sf

audio, sr = sf.read('/path/to/file.wav')
emb, ts = openl3.get_audio_embedding(audio, sr)

`emb` is a T-by-D numpy array:
>T is the number of embedding frames
D is the dimensionality of the embedding

`ts` is a length-T numpy array:
>timestamps for each embedding frame

### Get a list of embeddings and timestamp arrays for each of the input arrays.

import openl3
import soundfile as sf

audio1, sr1 = sf.read('/path/to/file1.wav')
audio2, sr2 = sf.read('/path/to/file2.wav')
audio3, sr3 = sf.read('/path/to/file3.wav')
audio_list = [audio1, audio2, audio3]
sr_list = [sr1, sr2, sr3]

# Pass in a list of audio arrays and sample rates
emb_list, ts_list = openl3.get_audio_embedding(audio_list, sr_list, batch_size=32)
# If all arrays use sample rate, can just pass in one sample rate
emb_list, ts_list = openl3.get_audio_embedding(audio_list, sr1, batch_size=32)

### OpenL3怎麼get embeddings?
1. Uses a mel-spectrogram time-frequency representation with 128 bands
2. Returns an embedding of dimensionality 6144 for each embedding frame

If you want to change the parameter:

emb, ts = openl3.get_audio_embedding(audio, sr, content_type="env",# <-model trained on environmental videos
                               input_repr="linear", embedding_size=512)
#spectrogram with linear frequency axis

### To compute embeddings for an audio file and directly save them to disk you can use `process_audio_file`

audio_filepath = '/path/to/file.wav'

# Save the embedding to '/path/to/file.npz'
openl3.process_audio_file(audio_filepath)

Load embeddings from disk

data = np.load('/path/to/file.npz')
emb, ts = data['embedding'], data['timestamps']

### Choosing an Audio Frontend (CPU / GPU)
OpenL3 provides two different audio frontends to choose from.
1. Kapre (GPU)
2. Librosa (CPU)