Version 0.1.0

• Excluded corpus creation from this project and moved it to its own (speech-corpus-dl)
• Restructured the bucket boundary calculations
• Updated documentation

Version 0.0.7

Update to TensorFlow version 1.12.0+. Use the tf.data function.

• The model now uses TensorFlow's current tf.data API.
• The model now uses an Estimator.
• Corpus
  • The corpus metadata is now stored in CSV files.
  • LibriSpeach loader can now use multiple threads.
  • Improved corpus generation performance.
• Fixed the GPU logging hook.
• Major code refactoring.

Version 0.0.6

• Added a downloader script that downloads and creates the used training corpus.
• Unified python/model.py for the different DS1 and DS2 architectures.

Version 0.0.5

First Batch of Extended Goals

• Removed overlong samples
• Extended plotting (pyplot without display)
• Added new dataset: Tatoeba
• Rewritten large parts of the load dataset functionality
• Added random seed
• Unified relative paths based on params.BASE_PATH
• Bug fixes

Version 0.0.4

Performance Optimization and Evaluation

• After every epoch, a validation run on the dev dataset is being run. #68
• Fixed SortaGrad bug with switching from the first epoch. #108, #121
• Increased number of buckets, to reduce the amount of padding used. #106
• Storing more checkpoints, also spaced out the steps at which checkpoints are written. #122
• Combined [DS1] style model into the modified [DS2] style one. Switching between them can be done via the params.py options. #119, #110
• Added threading to most /loader/ functions. #112
• Optimized loader methods and datasets further. #112, #99, #113
• See Milestone v0.0.4 for further information's.

Version 0.0.3

Deep Speech 2 Inspired Features

• Added SortaGrad ordering of training data for the first epoch. [Deep Speech 2]
• Replaced the first 3 dense layers with 2D convolutions over time and frequency [DS2]
• Added support for cuDNN RNN cell usage.
• Datasets
  • Added Mozilla Common Voice data.
  • Removed audio files with only 1 character long labels.
  • Removed files from TEDLIUM that had labels shorter than 5 words.
  • Removed audio files that had feature vectors longer than 1750.
• Added documentation about tensor shapes back in.
• Changed the input features to an 80 element logarithmic filter bank, with an option to switch to 80 element MFCC features.
• See Milestone v0.0.3 for further information.

Version 0.0.2

Improvements & Preprocessing

• Added Dropout functionality.
• Evaluation can now be run on the validation or test dataset.
• Preprocessing and input optimization:
  • Using python_speech_features to load audio, instead of the old librosa, which had problems running in parallel.
  • Added three new feature normalization methods.
  • Skipped every 2nd input feature frame (on the time axis).
  • Audiofiles can now be normalized by power level.
• Added support for Baidu's WarpCTC, which should be faster.
  • WarpCTC usage is bugged in evaluation runs.
• Updated project README.md.
• Updated generate_txt.py for new datasets (TEDLIUMv2, LibriSpeech, TIMIT).
• For more details please reference the Milestone v0.0.2

Version 0.0.1

First Prototype

DeepSpeech (v1)

• Trains on LibriSpeech ASR Coprus.
• Uses GRUCells in the bidirectional layer.
• Layout: 3 dense, 1 bidirectional RNN, 1 dense, 1 dense/logits layers.
• Evaluation isn't polished.
• No single input inference implemented.
• See Milestone v0.0.1 for further infos.