Low complexity WaveRNN-based speech coding by Jean-Marc Valin
Work in progress software for researching low CPU complexity algorithms for speech compression by applying Linear Prediction techniques to WaveRNN. The goal is to reduce the CPU complexity such that high quality speech can be synthesised on regular CPUs (around 1 GFLOP).
The BSD licensed software is written in C and Keras and currently requires a GPU (e.g. GT1060) to run. For training models, a GTX 1080 Ti or better is recommended.
This software is also a useful resource as an open source starting point for WaveRNN-based speech coding.
Set up a Keras system with GPU.
In the src/ directory, run ./compile.sh to compile the data processing program.
Then, run the resulting executable:
./dump_data input.s16 features.f32 pcm.s16
where the first file contains 16 kHz 16-bit raw PCM audio (no header) and the other files are output files. The input file currently used is 6 hours long, but you may be able to get away with less (and you can always use ±5% or 10% resampling to augment your data).
Now that you have your files, you can do the training with:
./train_lpcnet.py features.f32 pcm.s16
and it will generate a wavenet*.h5 file for each iteration. If it stops with a "Failed to allocate RNN reserve space" message try reducing the batch_size variable in train_wavenet_audio.py.
You can synthesise speech with:
./test_lpcnet.py features.f32 > pcm.txt
The output file pcm.txt contains ASCII PCM samples that need to be converted to WAV for playback
Speech Material for Training
Suitable training material can be obtained from the McGill University Telecommunications & Signal Processing Laboratory. Download the ISO and extract the 16k-LP7 directory, the src/concat.sh script can be used to generate a headerless file of training samples.
cd 16k-LP7 sh ~/CELP/src/concat.sh