Details will be shared tomorrow, but from what I have read they have distilled the large model decoder into this turbo that only has 4 layers instead of 32, the

Details will be shared tomorrow, but from what I have read they have distilled the large model decoder into this turbo that only has 4 layers instead of 32, the encoder should remain the same size. Similar to https://github.com/huggingface/distil-whisper https://github.com/huggingface/distil-whisper but the model is distilled using multilingual data instead of just English, and the decoder is 4 layers instead of 2.