TODO

[nglehuy]

• [x] Auto mixed precision and loss scaling
• [x] Fix transducer embedding tflite conversion (tflite conversion raises a bug when using tf.gather in tf.while_loop)
• [x] Fix transducer tflite conversion bug did not get operators, tensors, or buffers in subgraph 1
• [x] Support Word-Pieces (aka Subwords) beside Graphemes
• [x] Support log gammatone spectrogram
• [x] Support Jasper
• [x] Support gradients accumulation
• [x] Support masking in every models
• [x] Support ContextNet
• [ ] Support Streaming Conformer Transducer (https://arxiv.org/pdf/2010.11395.pdf)
• [ ] Re-implement and optimize Transducer Beam Search for tflite
• [ ] Support Semi-Supervised Learning like http://arxiv.org/abs/2010.10504
• [ ] Support NSC Beam Search for Transducer
• [ ] Support other language models than KenLM
• [ ] Support Sequence-to-Sequence Models such as Listen, Attend and Spell
• [ ] Support NovoGrad in Jasper Paper

[andreselizondo-adestech]

@usimarit Hello! I'm interested in seeing support for the Streaming Conformer Transducer. Is there an ETA for this? How can I help? I'm no expert, but I might be able to make a rough implementation to make things easier for you. (P.S. Excellent work on this repo. Love it.)

[nglehuy]

@andreselizondo-adestech I don't know about the ETA but I'll start working on the Streaming Conformer Transducer after refactoring the repo. If you want to help, you can start making some pull requests from your fork (it can be an empty pull request with descriptions so that I can know what you're gonna do :smile: and make some discussions).