ahclab / turntaking

Multimodal Voice Activity Prediction Model for Turn-taking

Model training for

• Multimodal Voice Activity Prediction: turntaking Events Detection in Expert-Novice Conversation

Installation

• Create conda env: conda create -n turntaking python=3.11
  • source env: conda source turntaking
• PyTorch: conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
• Dependencies:
  • Install Cython: pip install cython
  • Install requirements: pip install -r requirements.txt
• Install turntaking:
  • cd to root directory and run: pip install -e .

Set Up Datasets

• WARNING: Requires NoXi Database data.
• Place the file in any directory as follows

  .
  ├──noxi
  │  ├── Augsburg_01
  │  │   ├── audio_expert.wav
  │  │   ├── audio_mix.wav
  │  │   ├── audio_novice.wav
  │  │   ├── non_varbal_expert.csv
  │  │   ├── non_varbal_novice.csv
  │  │   ├── vad_expert.txt
  │  │   └── vad_novice.txt
  │  ├── Augsburg_02
  │  │   ├── audio_expert.wav
  │  │   ├── audio_mix.wav
  │  │   ├── audio_novice.wav
  │  │   ├── non_varbal_expert.csv
  │  │   ├── non_varbal_novice.csv
  │  │   ├── vad_expert.txt
  │  │   └── vad_novice.txt
    ...

• Rewrite EXTRACTED_PATH in turntaking/dataload/dataset/noxi/noxi.py and Rewrite AUDIO_DIR and MULTIMODAL_DIR in /turntaking/dataload/dataset/noxi/__init__.py.

Train and Test

1. Rewriting Model Training Conditions The training conditions can be changed by rewriting turntaking/conf/config.yaml. The model can be changed by rewriting turntaking/conf/model/model.yaml.

Refer to the following image for the corresponding module name of the model.

2. Model Training

python turntaking/train.py

3. Test

python turntaking/test.py

Citation

TBA