Speech Representations

Models and code for deep learning representations developed by the AWS AI Speech team:

• DeCoAR (self-supervised contextual representations for speech recognition)
• BERTphone (phonetically-aware acoustic BERT for speaker and language recognition)
• DeCoAR 2.0 (deep contextualized acoustic representation with vector quantization)

Installation

We provide a library and CLI to featurize speech utterances. We hope to release training/fine-tuning code in the future.

Kaldi should be installed to kaldi/, or $KALDI_ROOT should be set.

We expect Python 3.6+. Our bertphone models are defined in MXNet and DeCoAR model is defined in Pytorch. Clone this repository, then:

pip install -e .
pip install mxnet-mkl~=1.6.0   # ...or mxnet-cu102mkl for GPU w/ CUDA 10.2, etc.
pip install gluonnlp # optional; for featurizing with bertphone
pip install torch fairseq

Pre-trained models

First, download the model weights:

mkdir artifacts
# For DeCoAR trained on LibriSpeech (257M)
wget https://speech-representation.s3.us-west-2.amazonaws.com/checkpoint_decoar.pt
# For BertPhone_8KHz(λ=0.2) trained on Fisher
wget -qO- https://apache-mxnet.s3-us-west-2.amazonaws.com/gluon/models/bertphone_fisher_02-87159543.zip | zcat > artifacts/bertphone_fisher_02-87159543.params
# For Decoar 2.0:
wget https://speech-representation.s3.us-west-2.amazonaws.com/checkpoint_decoar2.pt

We support featurizing individual files with the CLI:

speech-reps featurize --model {decoar,bertphone,decoar2} --in-wav <input_file>.wav --out-npy <output_file>.npy
# --params <file>: load custom weights (otherwise use `artifacts/`)
# --gpu <int>:     use GPU (otherwise use CPU)

or in code:

from speech_reps.featurize import DeCoARFeaturizer
# Load the model on GPU 0
featurizer = DeCoARFeaturizer('artifacts/decoar-encoder-29b8e2ac.params', gpu=0)
# Returns a (time, feature) NumPy array
data = featurizer.file_to_feats('my_wav_file.wav')

We plan to support Kaldi .scp and .ark files soon. For now, batches can be processed with the underlying featurizer._model.

References

If you found our package or pre-trained models useful, please cite the relevant work:

DeCoAR

@inproceedings{decoar,
  author    = {Shaoshi Ling and Yuzong Liu and Julian Salazar and Katrin Kirchhoff},
  title     = {Deep Contextualized Acoustic Representations For Semi-Supervised Speech Recognition},
  booktitle = {{ICASSP}},
  pages     = {6429--6433},
  publisher = {{IEEE}},
  year      = {2020}
}

BERTphone

@inproceedings{bertphone,
  author    = {Shaoshi Ling and Julian Salazar and Yuzong Liu and Katrin Kirchhoff},
  title     = {BERTphone: Phonetically-aware Encoder Representations for Speaker and Language Recognition},
  booktitle = {{Speaker Odyssey}},
  publisher = {{ISCA}},
  year      = {2020}
}

DeCoAR 2.0

@misc{ling2020decoar,
      title={DeCoAR 2.0: Deep Contextualized Acoustic Representations with Vector Quantization}, 
      author={Shaoshi Ling and Yuzong Liu},
      year={2020},
      eprint={2012.06659},
      archivePrefix={arXiv},
      primaryClass={eess.AS}
}