Platalea v1.0

Introducing an attention-based encoder-decoder architecture for speech recognition. Multitask training with multiple objectives (e.g. cross-modality retrieval and speech transcription) is also possible now.