Skip to content

Latest commit

 

History

History
35 lines (26 loc) · 1.3 KB

readme.md

File metadata and controls

35 lines (26 loc) · 1.3 KB

Dixi Phonetic Vectoriser

A small ML library for audio, and a python file that uses it to create a phonetic vector space.

Dependencies

Install speech synth/ffmpeg for resampling by running sudo apt install ffmpeg espeak mbrola mbrola-en1 mbrola-us1 mbrola-us2 mbrola-us3. If you don't run apt, you have to find these packages. the mbrola voices have restrictions for commercial usage, so if you're an open source purist you can remove the mb voices from synth/voices.json.

tensorflow gpu must also be installed if you want training to finish before the end of days, but for the sake of brevity I will leave that as an exercise to the reader.

Running the model

If you're in a hurry, I included an output log directory, so you can skip straight to Seeing the vectors.

I hope you're using python 2. Clone the repo and generate the training data (it has to make 280,000 small audio clips, so this may take some time);

git clone https://github.com/Threadnaught/dixi
cd dixi
python synthVoices.py

Run the model (this step takes me about 15 hours, this model is very IO-bound);

python phoneticVector.py

Seeing the vectors

See the tensorboard output (open a web browser and go to localhost:6006);

tensorboard --logdir=logs