This code implements the Contrast Predictive Coding algorithm on audio data, as described in the paper Unsupervised Pretraining Transfers well Across Languages. This is an unsupervised method to train audio features directly from the raw waveform.
Moreover, this code also implements all the evaluation metrics used in the paper:
- ABX discriminability
- Phone and speaker linear separability
- Transfer learning on other languages, using the common voices datasets
The installation is a tiny bit involved due to the torch-audio dependency.
0/ Clone the repo:
git clone [email protected]:facebookresearch/CPC_audio.git && cd CPC_audio
1/ Install libraries which would be required for torch-audio https://github.com/pytorch/audio :
- MacOS:
brew install sox
- Linux:
sudo apt-get install sox libsox-dev libsox-fmt-all
2/ conda env create -f environment.yml && conda activate cpc37
3/ Run setup.py
python setup.py develop
You can test your installation with:
nosetests -d
This setup is given for CUDA 9.2 if you use a different version of CUDA then please change the version of cudatoolkit in environment.yml. For more information on the cudatoolkit version to use, please check https://pytorch.org/
We suggest to train the model either on Librispeech or libri-light.
To run a new training session, use:
python cpc/train.py --pathDB $PATH_AUDIO_FILES --pathCheckpoint $PATH_CHECKPOINT_DIR --pathTrain $TRAINING_SET --pathVal $VAL_SET --file_extension $EXTENSION
Where:
- $PATH_AUDIO_FILES is the directory containing the audio files. The files should be arranged as below:
PATH_AUDIO_FILES
│
└───speaker1
│ └───...
│ │ seq_11.{$EXTENSION}
│ │ seq_12.{$EXTENSION}
│ │ ...
│
└───speaker2
└───...
│ seq_21.{$EXTENSION}
│ seq_22.{$EXTENSION}
Please note that each speaker directory can contain an arbitrary number of subdirectories: the speaker label will always be retrieved from the top one. The name of the files isn't relevant. For a concrete example, you can look at the organization of the Librispeech dataset.
- $PATH_CHECKPOINT_DIR in the directory where the checkpoints will be saved
- $TRAINING_SET is a path to a .txt file containing the list of the training sequences (see here for example)
- $VALIDATION_SET is a path to a .txt file containing the list of the validation sequences
- $EXTENSION is the extension of each audio file
The code allows you to train a wide range of architectures. For example, to train the CPC method as described in Van Den Oord's paper just run:
python cpc/train.py --pathDB $PATH_AUDIO_FILES --pathCheckpoint $PATH_CHECKPOINT_DIR --pathTrain $TRAINING_SET --pathVal $VAL_SET --file_extension $EXTENSION --normMode batchNorm --rnnMode linear
Or if you want to train a model with a FFD prediction network instead of a transformer:
python cpc/train.py --pathDB $PATH_AUDIO_FILES --pathCheckpoint $PATH_CHECKPOINT_DIR --pathTrain $TRAINING_SET --pathVal $VAL_SET --file_extension $EXTENSION --rnnMode ffd --schedulerRamp 10
The --schedulerRamp option add a learning rate ramp at the beginning of the training: it barely affects the performance of a model with a transformer predictor but is necessary with other models.
Launch cpc/train.py -h to see all the possible options.
To restart a session from the last saved checkpoint just run
python cpc/train.py --pathCheckpoint $PATH_CHECKPOINT_DIR
All evaluation scripts can be found in cpc/eval/.
After training, the CPC model can output high level features for a variety of tasks. For an input audio file sampled at 16kHz, the provided baseline model will output 256 dimensional output features every 10ms. We provide two linear separability tests one for speaker, one for phonemes, in which a linear classifier is trained on top of the CPC features with aligned labels, and evaluated on a held-out test set.
Train / Val splits as well as phone alignments for librispeech-100h can be found here.
Speaker separability:
python cpc/eval/linear_separability.py $PATH_DB $TRAINING_SET $VAL_SET $CHECKPOINT_TO_LOAD --pathCheckpoint $PATH_CHECKPOINT
Phone separability:
python cpc/eval/linear_separability.py $PATH_DB $TRAINING_SET $VAL_SET $CHECKPOINT_TO_LOAD --pathCheckpoint $PATH_CHECKPOINT --pathPhone $PATH_TO_PHONE_LABELS
You can also concatenate the output features of several model by providing several checkpoint to the --load option. For example the following command line:
python cpc/eval/linear_separability.py -$PATH_DB $TRAINING_SET $VAL_SET model1.pt model2.pt --pathCheckpoint $PATH_CHECKPOINT
Will evaluate the speaker separability of the concatenation of the features from model1 and model2.
You can run the ABX score on the Zerospeech2017 dataset. To begin, download the dataset here. Then run the ABX evaluation on a given checkpoint with:
python ABX.py from_checkpoint $PATH_CHECKPOINT $PATH_ITEM_FILE $DATASET_PATH --seq_norm --strict --file_extension .wav --out $PATH_OUT
Where:
- $PATH_CHECKPOINT is the path pointing to the checkpoint to evaluate
- $PATH_ITEM_FILE is the path to the .item file containing the triplet annotations
- $DATASET_PATH path to the directory containing the audio files
- $PATH_OUT path to the directory into which the results should be dumped
- --seq_norm normalize each batch of features across the time channel before computing ABX
- --strict forces each batch of features to contain exactly the same number of frames.
To begin download the common voices datasets here, you will also need to download our phonem annotations and our train / val / test splits for each language here. Then unzip your data at PATH_COMMON_VOICES. Unfortunately, the audio files in common voices don't have the same sampling rate as in Librispeech. Thus you'll need to convert them into 16kH audio using the command:
DIR_CC=$PATH_COMMON_VOICES
for x in fr zh it ru nl sv es tr tt ky; do python cpc/eval/utils/adjust_sample_rate.py ${DIR_CC}/${x}/clips ${DIR_CC}/${x}/validated_phones_reduced.txt ${DIR_CC}/${x}/clips_16k; done
You can now run the experiments described in the paper. To begin, you must train the linear classifier. You will find below the instructions for the Spanish dataset: you can run the experiments on any other dataset in the same fashion.
To run the training on frozen features with the one hour dataset, just run:
python cpc/eval/common_voices_eval.py train $PATH_COMMON_VOICES/es/clips_16k $PATH_COMMON_VOICES/es/validated_phones_reduced.txt $CHECKPOINT_TO_TEST --pathTrain $PATH_COMMON_VOICES/es/trainSeqs_1.0_uniform_new_version.txt --pathVal $PATH_COMMON_VOICES/es/trainSeqs_1.0_uniform_new_version.txt --freeze -o $OUTPUT_DIR
The command is quite similar to run the fine-tuning experiments on the 5 hours dataset. For example in French you need to run:
python cpc/eval/common_voices_eval.py train $PATH_COMMON_VOICES/es/clips_16k $PATH_COMMON_VOICES/es/validated_phones_reduced.txt $CHECKPOINT_TO_TEST --pathTrain $PATH_COMMON_VOICES/es/trainSeqs_5.0_uniform_new_version.txt --pathVal $PATH_COMMON_VOICES/es/trainSeqs_5.0_uniform_new_version.txt --freeze -o $OUTPUT_DIR
Once the training is done, you can compute the associated phone error rate (PER) on the test subset. To do so, just run:
python cpc/eval/common_voices_eval.py per $OUTPUT_DIR --pathVal $PATH_COMMON_VOICES/es/testSeqs_uniform_new_version.txt --pathPhone $PATH_COMMON_VOICES/es/validated_phones_reduced.txt
To begin download the common voices datasets here, you will also need to download our phonem annotations and our train / val / test splits for each language here. Then unzip your data at PATH_COMMON_VOICES. Unfortunately, the audio files in common voices don't have the same sampling rate as in Librispeech. Thus you'll need to convert them into 16kH audio using the command:
DIR_CC=$PATH_COMMON_VOICES
for x in fr zh it ru nl sv es tr tt ky; do python cpc/eval/utils/adjust_sample_rate.py ${DIR_CC}/${x}/clips ${DIR_CC}/${x}/validated_phones_reduced.txt ${DIR_CC}/${x}/clips_16k; done
You can now run the experiments described in the paper. To begin, you must train the linear classifier. You will find below the instructions for the Spanish dataset: you can run the experiments on any other dataset in the same fashion.
To run the training on frozen features with the one hour dataset, just run:
python cpc/eval/common_voices_eval.py train $PATH_COMMON_VOICES/es/clips_16k $PATH_COMMON_VOICES/es/validated_phones_reduced.txt $CHECKPOINT_TO_TEST --pathTrain $PATH_COMMON_VOICES/es/trainSeqs_1.0_uniform_new_version.txt --pathVal $PATH_COMMON_VOICES/es/trainSeqs_1.0_uniform_new_version.txt --freeze -o $OUTPUT_DIR
The command is quite similar to run the fine-tuning experiments on the 5 hours dataset. For example in French you need to run:
python cpc/eval/common_voices_eval.py train $PATH_COMMON_VOICES/es/clips_16k $PATH_COMMON_VOICES/es/validated_phones_reduced.txt $CHECKPOINT_TO_TEST --pathTrain $PATH_COMMON_VOICES/es/trainSeqs_5.0_uniform_new_version.txt --pathVal $PATH_COMMON_VOICES/es/trainSeqs_5.0_uniform_new_version.txt --freeze -o $OUTPUT_DIR
Once the training is done, you can compute the associated phone error rate (PER) on the test subset. To do so, just run:
python cpc/eval/common_voices_eval.py per $OUTPUT_DIR --pathVal $PATH_COMMON_VOICES/es/testSeqs_uniform_new_version.txt --pathPhone $PATH_COMMON_VOICES/es/validated_phones_reduced.txt
This model is also available via torch.hub. For more details, have a look at hubconf.py.
Please consider citing this project in your publications if it helps your research.
@misc{rivire2020unsupervised,
title={Unsupervised pretraining transfers well across languages},
author={Morgane Rivière and Armand Joulin and Pierre-Emmanuel Mazaré and Emmanuel Dupoux},
year={2020},
eprint={2002.02848},
archivePrefix={arXiv},
primaryClass={eess.AS}
}
CPC_audio is MIT licensed, as found in the LICENSE file.