Version 0.3.0

• Fix GradientAccumulationModel wrapper for tf.keras.Model. Now it can work fine on single GPU (use GA optimizer for multi-GPU according to author's suggestion).
• Add finetune option.

Version 0.2.1

• Update librosa support from 0.10.0
• Support tf.function for BeamRNNT decoder

Super Beta version

This release marks new updates:

• Refactoring multi types of decoders (ALSD, TSD, Beam, and Greedy Search).
• Add time stretch augmentation.

v0.1.0-beta

This marks the first-ever release, more will come in the future. This release:

• Supports Emformer and Conformer as the encoder, RNNT and CTC as the E2E architecture.
• Supports WandB and TensorBoard logging.
• Supports HyperPyYAML for configuration management.
• Supports error analysis with asr evaluation tools.
• Supports multiple augmentation and featurization methods.