GuitarLSTM Browser Experiment

The goal of the experiment was to enable loading GuitarML's LSTM models in browsers or similiar environments (eg. WebViews)

The code is capable of predicting neural network's output based on a choosen pre-trained model and a user-specified tensor (imported as json) and allows users to listen to the result by utilizing Web Audio API

A smaller (50-input) model of Ibanez Tube Screamer TS9 was used for better perfomance and smaller tensor file-size

Example tensor (tensor.json), input audio (in.wav) and prediction result (out_pred.wav) are available in the samples directory

Loading your own tensors

First, install GuitarLSTM and all the dependencies

git clone https://github.com/GuitarML/GuitarLSTM

pip install -r requirements.txt

After training a model, run savetensor.py script, specifying paths to your input audio, output tensor file and your model

savetensor.py data/in.wav test.json models/t9sm/t9sm.h5

Once finished, download the tensor file (test.json in our example) and load it in the web application using the "Upload" button

Loading your own models

Install tensorflowjs pip package

pip install tensorflowjs

Locate and convert your model using tensorflowjs_converter utility

tensorflowjs_converter --input_format keras models/t9sm/t9sm.h5 models/t9sm/t9sm

Then download the resulting files (.json, .bin)

Specify the path to your .json file by changing this line in index.html

const model = await tf.loadLayersModel('/models/ts9sm/model.json')

Areas of improvement / Future work

• Make possible to work with .wav files natively instead of tensors
• Make possible to upload/specify custom models
• Add effects (with projects like Tone.js)