---
configs:
- config_name: "English"
  default: True
  data_files:
  - split: train
    path: Eng-NA/train.csv
  - split: valid
    path: Eng-NA/valid.csv
- config_name: "French"
  data_files:
  - split: train
    path: French/train.csv
  - split: valid
    path: French/valid.csv
- config_name: "German"
  data_files:
  - split: train
    path: German/train.csv
  - split: valid
    path: German/valid.csv
- config_name: "Spanish"
  data_files:
  - split: train
    path: Spanish/train.csv
  - split: valid
    path: Spanish/valid.csv
language:
- en
- de
- fr
- es
tags:
- language modeling
- cognitive modeling
pretty_name: Phonemized Child Directed Speech
size_categories:
- 100K<n<1M
---

# Phonemized Child Directed Speech Dataset

This dataset contains utterance downloaded from CHILDES which have been pre-processed and converted to phonemic transcriptions by this [processing script](https://github.com/codebyzeb/Corpus-Phonemicizers). Many of the columns from CHILDES have been preserved in case they may be useful for experiments (e.g. number of morphemes, part-of-speech tags, etc.). The key columns added by the processing script are as follows:

| Column | Description |
|:----|:-----|
| `is_child`| Whether the utterance was spoken by a child or not. Note that this is set to `False` for all utterances in this dataset, but the processing script has the ability to preserve child utterances.|
| `processed_gloss`| The pre-processed orthographic utterance. This includes lowercasing, fixing English spelling and adding punctuation marks. This is based on the [AOChildes](https://github.com/UIUCLearningLanguageLab/AOCHILDES) preprocessing.|
| `phonemized_utterance`| A phonemic transcription of the utterance, space-separated with word boundaries marked with the `WORD_BOUNDARY` token.|
| `language_code`| Language code used for producing the phonemic transcriptions. May not match the `language` column provided by CHILDES (e.g. Eng-NA and Eng-UK tend to be transcribed with eng-us and eng-gb). |
| `character_split_utterance`| A space separated transcription of the utterance, produced simply by splitting the processed gloss by character. This is intended to have a very similar format to `phonemized_utterance` for studies comparing phonetic to orthographic transcriptions. |

The last two columns are designed for training character-based (phoneme-based) language models using a simple tokenizer that splits around whitespace. The `processed_gloss` column is suitable for word-based (or subword-based) language models with standard tokenizers. 

Note that the data has been sorted by the `target_child_age` column, which stores child age in months. This can be used to limit the training data according to a maximum child age, if you wish. 

Each subset of the data is split into a training split containing most of the utterances and an in-distribution validation split containing 10,000 utterances. The following languages are included (ordered by number of phonemes):

| Language | Description | Speakers | Utterances | Words | Phonemes
|:----|:-----|:-----|:----|:-----|:-----|
| English | Taken from 44 corpora in Eng-NA collection of CHILDES and phonemized using language code `en-us`. | 2,692 | 1,646,954 | 7,090,066 | 21,932,139
| German | Taken from 10 corpora in German collection of CHILDES and phonemized using language code `ge`. | 627 | 850,888 | 3,893,168 | 14,032,948
| French | Taken from 11 corpora in French collection of CHILDES and phonemized using language code `fr-fr`. | 722 | 432,133 | 1,995,063 | 5,510,523
| Spanish | Taken from 18 corpora in Spanish collection of CHILDES and phonemized using language code `es`. | 562 | 286,462 | 1,266,366 | 4,511,261