voice-emotion-seq2seq

Voice emotion conversion model for DS/ML master's thesis. F0 contour mapping in sequence-to-sequence RNN-LSTM architecture in Tensorflow.

Directories

• audio/ - samples of original and transformed voice samples in WAV format
• data/ - Matlab MAT files produced by AudioSculpt, containing F0 contours and syllable/phoneme alignments
• documentation/ - PDF reports on the state of the art, aims, method, experimental setup, results and discussion
• model/ - python code to build your own seq2seq model
• postprocessing/ - code to process the outputted F0 contours and apply to neutral WAV files
• preprocessing/ - code to process the MAT files to create input data for the model
• utilities/ - code to rename WAV files for use in the online survey, and to process survey results
• visualisation/ - jupyter notebooks to visualise the F0 contours

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.

Prerequisites

Libraries:

• Python 3.6, SciPy, Matplotlib (Anaconda recommended)
• Tensorflow 1.4 or later (1.8 recommended)
• as_pysrc - collection of modules for signal processing, maintained by the AS team on forge.ircam.fr (SSH access needed)

Software:

• AudioSculpt + IrcamAlign to generate .MAT files with F0 values and syllable/phoneme alignments

Pre-processing & Model Training Process

• tf_seq2seq_data_processing_phoneme_to_phoneme.ipynb
  • check phrase/emotions/intensities are set correctly
  • check ratio of train/dev/test are correct (usually 0.8/0.2/0.0)
  • run all blocks to generate the syllable csv files, the combo source/target files and the vocab files
• __tf_seq2seq_data_processing_test_input.ipynb__
  • run first section to generate the test input (all 80 neutral source files)
  • put test_source.txt in the ..out/test directory (delete existing contents)
• upload files to server
  • scp -r /Users/robinson/Dropbox/anasynth/_data/emoVC/Olivia2006/f0_raw_phoneme/out/* robinson@gusli:/data2/anasynth_nonbp/robinson/tf_seq2seq_data3
• tf_seq2seq_config.yml
  • adjust params to fit length of source/target files (attention, max_seq_length, num_units)
• kick off training and inference
  • login to gusli
  • cd ~/code/tf_seq2seq
  • activate gpu environment
  • nohup bash ./tf_seq2seqrun.sh > nohup$(date +"%Y%m%d_%H%M%S").out &

Post-processing Process

• download model directory (all checkpoints and predictions)
  • scp -r robinson@gusli:/data2/anasynth_nonbp/robinson/tf_seq2seq_data/model/20180902_090436 ./20180902_090436
    • change date for latest model directory
• download log
  • scp robinson@gusli:/u/anasynth/robinson/code/tf_seq2seq/nohup_20180827_171528.out .
• move the download
  • make directory /Users/robinson/Downloads/data/pred/20180712_134318/ (change date)
  • make model subdirectory, and move the downloaded model folder into it, so it looks like /Users/robinson/Downloads/data/pred/20180713_153850/model/20180713_153850
• get predictions.txt and process it
  • get from /pred/
  • remove first line of txt file (comp device)
  • remove prefixes if using conditioned model (search and replace a/b/c for nothing)
  • put predictions.txt in directory where model is saved to e.g. /Users/robinson/Downloads/data/pred/20180712_134318/
• tf_seq2seq_data_processing_make_new_contours.ipynb
  • edit datetimestamp in directory path to match where you put predictions.txt
  • run to generate new phoneme files from predictions.txt for conversion operations to follow
• __tf_seq2seq_f0_transform_phoneme_nosyll.ipynb__
  • copy subdirectories of '/Users/robinson/Downloads/data/Olivia2006/Olivia2006_AUDIO’ to previous saved model directory e.g. /Users/robinson/Downloads/data/pred/20180712_134318/
  • set params at the top
    • edit directory path datestamp folder name at the top
    • change emotion to JOY, COL or other emotion tag
    • set interp = True
    • set harmonicity threshold, harm_thresh = 0.7
  • run all blocks
  • listen to the audio in the compare directory, and select the best ones
• tf_seq2seq_data_comparison_phrase_phoneme_nosyll.ipynb
  • run this to generate F0 contour figures in f0compare directory, for visual inspection of results
  • edit syll_input_directory_path datestamp folder name e.g. '/Users/robinson/Downloads/data/pred/20180712_134318/syllables’

Authors

• Carl Robinson - Master's Student - Voice Tech Podcast
• Nicolas Obin - Supervisor - IRCAM - CNRS - Sorbonne Université

License

This project is licensed under the MIT License - see the LICENSE.md file for details

Acknowledgments

This project would not have been possible without the generous help and support of many talented people. I would like to thank the following people for giving me their time, for sharing their knowledge and ideas, and for their continued encouragement:

• Nicolas Obin
• Axel Roebel, Guillaume Doras and Rafael Ferro
• Sylvie Thiria, Cecile Mallet, Sonia Garcia, Nesma Houmani, and Jerome Boudy
• Eric Bolo and the Batvoice team
• Veronique Sieng