VoiceActivityProjection

Voice Activity Projection is a Self-supervised objective for Turn-taking Events. This is an extended version which trains a stereo model (mono is still possible) that does not require any VAD information as input BUT do require separate channels for both speakers. Overbleed between the channels is fine as long as you have access to the VAD information (used as label during training). The stereo model greatly simplifies inference where the only input is a stereo waveform. The model is trained on a multitask loss defined by the original VAP-objective and a VAD-objective (predict the current voice activity over each frame for the two separate channels).

Installation

• Create conda env: conda create -n vap python=3.10
  • source env: conda source vap
  • Working with python 3.10 but I don't think it matters too much...
• PyTorch: conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
  • Have not tested all versions but should work from torch 2.0.1 as of time of writing...
• Install VAP (this repo):
  • cd to root directory and run:
    • pip install -r requirements.txt
    • pip install -e .

VapStereo

1. Audio encoder
   • (Default) CPC facebookresearch)
   • Causal / incremental
2. Self attention
   • (Default) 1 layer
   • Shared weights
3. Cross attention between speakers
   • (Default) 3 layers
   • Shared weights
4. Combine the output from the two channels
   • The model is channel agnostic and used for both channels
5. Linear projection layer
   • Map to VAP vocabulary/states
6. Voice activity detection
   • Predict the voice activity for both channels independently
   • Good for "bleed-over", background noise, etc
7. VAP objective
   • 256 states
   • Representing the next 2s of dialog

The model is a GPT-like transformer model, using AliBI attention, which operates on pretrained speech representation (extract by submodule defined/trained/provided by CPC facebookresearch). A state-dict tensor is included in the examples/ folder:

• example/VAP_3mmz3t0u_50Hz_ad20s_134-epoch9-val_2.56.pt

Training

Training is done with the main.py script that uses hydra, read more about the data (csv-files) in vap/data/README.md. Look into the default config in vap/conf/default_config.yaml to get started. Entries that exists in the config can be changed by python vap/main.py datamodule.num_workers but other entries you could add requires a '+' sign python vap/main.py +trainer.limit_val_batches=0.5.

• SEE /scripts/finetune.bash or /scripts/train.bash or /scripts/eval.bash

As an example:

python vap/main.py \ 
    datamodule.train_path=example/data/sliding_train.csv \ 
    datamodule.val_path=example/data/sliding_val.csv \ 
    datamodule.num_workers=12 \ 
    datamodule.batch_size=10

Load checkpoint

Both the state_dict.pt and checkpoint.ckpt are 'loadable' with torch.load() the extensions are arbitrary.

from vap.modules.lightning_module import VAPModule

# Load entire module (includes metrics, optimizers, epochs, ...)
module = VAPModule.load_from_checkpoint("/PATH/TO/checkpoint.ckpt")
# type(module) -> vap.modules.lightning_module.VAPModule

# Load only the `VAP` model
model = VAPModule.load_model("/PATH/TO/checkpoint.ckpt")
# type(model) -> vap.modules.VAP.VAP

# Load raw state dict (model)
from vap.modules.VAP import VAP
from vap.modules.encoder import EncoderCPC
from vap.modules.modules import TransformerStereo

# WARNING: The model requires `encoder` and `transformer` modules as input
# The correct parameters sizes are required.
encoder = EncoderCPC()
transformer = TransformerStereo()
model = VAP(encoder, transformer)  # the barebones model
model.load_state_dict("/PATH/TO/state_dict.pt")

Barebones parameters

• SEE code in /scripts/checkpoint_to_state_dict.py

python scripts/checkpoint_to_state_dict.py \ 
    --checkpoint runs_new/VAP2/lh55ur9z/checkpoints/epoch=5-step=16085.ckpt \ 
    --state_dict /OUTPUTP/TO/state_dict.pt # default model_state_dict.py

Citation

@inproceedings{ekstedt22_interspeech,
  author={Erik Ekstedt and Gabriel Skantze},
  title={{Voice Activity Projection: Self-supervised Learning of Turn-taking Events}},
  year={2022},
  booktitle={Proc. Interspeech 2022},
  pages={5190--5194},
  doi={10.21437/Interspeech.2022-10955}
}

@inproceedings{ekstedt-skantze-2022-much,
    title = {How Much Does Prosody Help Turn-taking? Investigations using Voice Activity Projection Models},
    author={Erik Ekstedt and Gabriel Skantze},
    booktitle = {Proceedings of the 23rd Annual Meeting of the Special Interest Group on Discourse and Dialogue},
    month = {9},
    year = {2022},
    address = {Edinburgh, UK},
    publisher = {Association for Computational Linguistics},
    url = {https://aclanthology.org/2022.sigdial-1.51},
    pages = {541--551},
}