EntropyBasedCLDMetrics

Github repository accompanying the paper Measuring language development from child-centered recordings

Repository organization

• configs/ contains all the necessary options to train/test the models.
• src/ contains all the python source code.
• analysis/ contains all the R notebook for analysing the results.
• plots contains the graphics reported in the paper.

For reproducing all the experiments, you will need to:

1. Prepare the data to train and test the models
2. Run the training
3. Run the testing
4. Reproduce the anlaysis (plots and statistical analysis of the results)

Corpus

Librispeech : https://www.openslr.org/12
Thomas : https://gin.g-node.org/LAAC-LSCP/thomas
Providence: https://gin.g-node.org/LAAC-LSCP/providence

Working environment

Clone this github repository and move to it:

git clone https://github.com/yaya-sy/EntropyBasedCLDMetrics.git
cd EntropyBasedCLDMetrics

Create the Python environment:

conda env create -f environment.yml

and activate it:

conda activate ent_cldm

Data preparation

Prepare the librispeech data for phones n-gram language model

The phonemizer requires the espeak backend, it can be installed with this command line: apt-get install espeak-ng

python src/librispeech_for_ngram_lm.py -i [LIBRISPEECH_TRAIN-CLEAN-360_FOLDER] -o data/ngram_lm/

This will create two files in data/ngram_lm. The one with the *.orthographic extension contains the orthographic utterances and the one with *.phonemized extension contains the phonemized utterances.

n-gram language model training

We need first to train the n-gram language model in order to prepare the data for the other experiments.

For training the n-gram language model, you will need to install KenLM:

conda install -c anaconda cmake
git clone https://github.com/kpu/kenlm.git
cd kenlm
python setup.py develop
mkdir -p build
cd build
cmake ..
make -j 4

Once installed in the current directory, you can run the training:

mkdir checkpoints
kenlm/build/bin/lmplz --discount_fallback -o 5 < data/ngram_lm/librispeech.phonemized > checkpoints/librispeech_360.arpa

The trained model will be stored in the checkpoints folder.

Pepare the data for Thomas regression model (Experiment 2A)

Prepare utterances

You will need to install the Thomas corpus https://gin.g-node.org/LAAC-LSCP/thomas.

To install the thomas corpus using datalad, run the following commands:

datalad install -r git@gin.g-node.org:/LAAC-LSCP/thomas.git
cd thomas
datalad get annotations/cha/*
datalad get recordings/raw/*

Once installed, you can run this command to extract utterances, their cleaned version and the timemarks:

python src/create_thomas_corpus.py -c [PATH_TO_THOMAS_CORPUS] -o data/Thomas

Where [PATH_TO_THOMAS_CORPUS] is the path to the installed Thomas corpus.

In the created folder, orthographic contains the raw annotations without cleaning. The cleaned folder contains the cleaned version of the annotations. And timemarks contains the onsets and offsets of each utterance in the audios. All of these are aligned, meaning that the i^th line of each file corresponds to the i^th line of the other files.

The filename.txt files contain the raw filenames and months.txt files contain the ages of the child in months.

Prepare inputs for the regression model

> python src/prepare_childes_corpus.py -i data/Thomas/
> python src/prepare_input_files.py -c data/Thomas/ -a [AUDIO_FOLDER] -m checkpoints/librispeech_360.arpa

Where [AUDIO_FOLDER] is the path to the audio folder of the data installed from the GIN repository. The audio folder is recordings/raw/.

Pepare the data for Librispeech regression model (Experiment 2B)

Create the inputs for the regression model:

> python src/prepare_librispeech_corpus.py -i [LIBRISPEECH_TRAIN-CLEAN-100_FOLDER] -o data/Librispeech/model_inputs
> python src/prepare_input_files.py -c data/Librispeech/ -a [LIBRISPEECH_TRAIN-CLEAN-100_FOLDER] -m checkpoints/librispeech_360.arpa

Where [LIBRISPEECH_TRAIN-CLEAN-100_FOLDER] is the path to the folder containing the librispeech train-clean-100.

Prepare the Providence test data

As for the Thomas corpus, you will also need to install the providence corpus https://gin.g-node.org/LAAC-LSCP/providence.

Prepare utterances

Extract the utterances of the providence corpus:

python src/create_providence_corpus_new.py -i [PREPARED_CSV] -c [PATH_TO_PROVIDENCE_CORPUS] -o data/Providence/

Where [PREPARED_CSV] is the CSV aleady prepared with cleaned utterances, the timemarks, etc.

Prepare inputs for the regression model

Create the inputs for the model:

> python src/prepare_childes_corpus.py -i data/Providence/
> python src/prepare_input_files.py -c data/Providence/ -a [AUDIO_FOLDER] -m checkpoints/librispeech_360.arpa

Where [AUDIO_FOLDER] is the path to the audio folder of the data installed from the GIN repository. The audio folder is recordings/raw/.

Run the trainings

Experiment 1A: n-gram language model training on Librispeech train-clean-360

The model is already trained during the data prepration and is saved on checkpoints/librispeech_360.arpa. So we will not retrain it again.

Experiment 2A: Entropy predictions training on Thomas

Run the regression model training on Thomas:

python src/train.py -c configs/thomas.yaml

The trained model will be stored in the folder checkpoints as Thomas_30h_Librispeech360_en.pt.

Experiment 2B: Entropy predictions training on Librispeech train-clean-100

Run the regression model training on librispeech train-clean-100:

python src/train.py -c configs/librispeech.yaml

The trained model will be stored in the folder checkpoints as Librispeech_100h_Librispeech360_en.pt.

Run the testing

Experiment 1A: Text entropies

python src/compute_entropies_ngram_lm.py

This will create a csv file named Librispeech_360h.csv in the folder results.

Experiment 2A: Entropy predictions using Thomas test data

python src/compute_entropies_whisper.py -c configs/test.yaml -m checkpoints/Thomas_30h_Librispeech360_en.pt

This will create a csv file named Thomas_30h_Librispeech360_en.csv in the folder results.

Experiment 2B: Entropy predictions using Librispeech test data

python src/compute_entropies_whisper.py -c configs/test.yaml -m checkpoints/Librispeech_100h_Librispeech360_en.pt

This will create a csv file named Librispeech_100h_Librispeech360_en.csv in the folder results.

Analysis

Prepare the CSVs for analysis

Experiment 1A

python src/prepare_for_analysis.py -i results/Librispeech_360h.csv

This will create a csv file named Librispeech_360h_analysis.csv in the folder results.

Experiment 1B

python src/prepare_for_analysis_hubert.py -i results/HuBERT-nat_entropy_ngram-2-merge-False_mmap.csv -c [CHILDES_PATH_PROVIDENCE]

This will create a csv file named HuBERT-nat_entropy_ngram-2-merge-False_mmap_analysis.csv in the folder results.

Experiment 1C

python src/prepare_for_analysis_hubert.py -i results/HuBERT-tts_entropy_ngram-2-merge-False_mmap.csv -c [CHILDES_PATH_PROVIDENCE]

This will create a csv file named HuBERT-tts_entropy_ngram-2-merge-False_mmap_analysis.csv in the folder results.

Experiment 2A

python src/prepare_for_analysis.py -i results/Thomas_30h_Librispeech_en.csv

This will create a csv file named Thomas_30h_Librispeech_en_analysis.csv in the folder results.

Experiment 2B

python src/prepare_for_analysis.py -i results/Librispeech_100h_Librispeech360_en.csv

This will create a csv file named Librispeech_100h_Librispeech360_en_analysis.csv in the folder results.

Plottings

You can reproduce the figures of the paper with this notebook: analysis/plots.Rmd

Mixed Linear Models

You can reproduce the mixed linear models of the paper with this notebook: analysis/models.Rmd

Correlations

We give the MLU, IPSyn and VOCD already computed on the Providence corpus. The CSV file is in extra/chi.kideval.csv.

But before computing the correlations with the entropy metric, you will need to merge chi.kideval.csv with the CSVs produced in the previous experiments.

For the experiments 1A, 2A and 2B, you can prepare the CSVs for computing the correlations using this command:

python src/merge_metrics.py -i [CSV_FOR_ANALYSIS]

Where [CSV_RESULTS_FOR_ANALYSIS]is the path to CSV results already prepared for analysis.

For the experiments 1B and 1C, you can prepare the CSVs for computing the correlations using this command:

python src/merge_metrics_hubert.py -i [CSV_FOR_ANALYSIS]

Once done, you can use the notebook analysis/correlations.Rmd to compute correlations.