Tensorflowasr

Release Notes

- Update dataset config
- Update ASRDataset with option to use preprocessing in pure tf (not `tf.numpy_function`)
- Add Transducer Greedy Decoding V2 from [https://arxiv.org/pdf/1801.00841.pdf](https://arxiv.org/pdf/1801.00841.pdf)
- Update documents and results
- Update examples scripts to use new dataset config

Release Notes

- [x] Implement Slice Dataset that yield data in keras format `(x, y)`
- [x] Implement TFRecord Dataset that yield data in keras format `(x, y)`
- [x] Add new `RnntLoss` subclass of `tf.keras.losses.Loss` to use in `compile` function of `tf.keras.Model`
- [x] Add Transducer with custom `train_step` and `test_step` that uses `RnntLoss`
- [x] Add new `CtcLoss` subclass of `tf.keras.losses.Loss` to use in `compile` function of `tf.keras.Model`
- [x] Add CtcModel with custom `train_step` and `test_step` that uses `CtcLoss`

Release Notes

- Fix tflite conversion and interpretation on Transducer's Greedy Decoding, the reason is somehow tflite drops support for `tf.cond` so we have to use `tf.where` instead. Tested on `tensorflow==2.3.1` and `tf-nightly`
- Small update on `BaseTrainer` class: infinity evaluation after training is done, refactor some utility functions
- Update `SentencePieceFeaturizer`

Release Notes

- Add downloading script for Multilingual LibriSpeech dataset
- Add `SentencePieceFeaturizer`
- Remove `tf.string` from Transducer Beam Search for GPU support
- Add RNNT implementation in pure TF (Ref: [https://github.com/iamjanvijay/rnnt](https://github.com/iamjanvijay/rnnt))
- Add option for tfrecords shards and buffer size

Release Notes

- Update dependencies
- Fix Transducer RNN name

Release Notes

- Add unittests
- Fixed CTC `recognize_tflite` and `recognize_tflite_beam`
- Add `joint_activation` to Transducer