mpoli/initial-phonetic-learning

Modeling the initial state of early phonetic learning

CPC checkpoints for:

Poli, M., Schatz, T., Dupoux, E., & Lavechin, M. (2025). Modeling the initial state of early phonetic learning in infants. Language Development Research, 5(1), 1–34. https://doi.org/10.34842/y89t-6q31

Code: https://github.com/mxmpl/initial-phonetic-learning

Each model may first be pretrained (on ambient sounds, multilingual speech, or another language). It is then trained in a development phase on 1 to 500 hours of American English or Japanese.

Layout

checkpoints/<pretraining>_<development>/<pretraining split>_<development split>/
    checkpoint_<epoch>.pt
    checkpoint_args.json
    checkpoint_logs.json
    checkpoint_args_orig.json    (only for models that started from a pretrained model)

Examples:

Path	Model
checkpoints/scratch_null/null_null	Untrained model (random weights)
checkpoints/ambient-500h_null/s000_null	Pretrained on 500 h of ambient sounds, no development
checkpoints/scratch_jpn-004h/null_s034	No pretraining, trained on 4 h of Japanese (split 34)
checkpoints/ambient-500h_jpn-004h/s000_s034	Ambient-sound pretraining, then the same 4 h of Japanese
checkpoints/eng-500h_jpn-004h/s000_s124	500 h of English, then 4 h of Japanese (split 124)
checkpoints/eng-001h_jpn-020h/s097_s000	1 h of English (split 97), then 20 h of Japanese (split 0)

Naming

Part	Values	Meaning
pretraining	scratch	No pretraining (random initialization)
	ambient-500h	500 h of ambient sounds (Animal Sound Archive and AudioSet, no speech or music)
	multilingual-500h	500 h of VoxPopuli in 18 non-Germanic languages (about 28 h each)
	eng-<hours>, jpn-<hours>	A model trained on English or Japanese, used as the starting point
development	eng-<hours>, jpn-<hours>	The language and amount of speech the model was then trained on
	null	No development phase
hours	001h 002h 004h 020h 100h 500h	Zero-padded
split	s<k>	Split k of the training set at that duration; a full 500 h set is s000
	null	No data at that stage

• eng is American English audiobooks from LibriVox, selected to match the Japanese corpus in number of speakers and amount of speech per speaker.
• jpn is the Corpus of Spontaneous Japanese (CSJ).
• Each corpus has about 500 h, cut into mutually exclusive splits of 1, 2, 4, 20 and 100 h. Each condition has 15 models per language at 1, 2, 4 and 20 h, 5 at 100 h, and 1 at 500 h.

Conditions

Folders (in checkpoints/)	Models	Description
scratch_null, ambient-500h_null, multilingual-500h_null	3	Initial states, before any speech experience
scratch_{eng,jpn}-*	132	No pretraining (baseline)
ambient-500h_{eng,jpn}-*	132	Pretraining on ambient sounds (Experiment 1)
multilingual-500h_{eng,jpn}-*	102	Pretraining on multilingual speech (Experiment 2)
eng-500h_jpn-*, jpn-500h_eng-*	132	Cross-lingual pretraining from the 500 h model of the other language (Experiment 3, appendix; with 2 h models not used in the paper)
{eng,jpn}-{001h,…,100h}_{jpn,eng}-020h	130	Cross-lingual pretraining from each smaller model of the other language, then the same 20 h split (not used)

Files

• checkpoint_<epoch>.pt: model weights at the best epoch by validation accuracy. 5 % of each training set was held out for validation.
• checkpoint_args.json: training arguments.
• checkpoint_logs.json: training and validation loss and accuracy for each epoch.
• checkpoint_args_orig.json: the training arguments of the pretraining run (only for models that started from a pretrained model).

Notes

• Epoch numbers continue from pretraining. A model that started from a pretrained one keeps the original epoch counter and logs. The first 100 entries of its checkpoint_logs.json come from pretraining, and training on the new data starts after them.
• Paths inside the JSON files (for example pathDB, pathCheckpoint, load) are absolute paths from the cluster the models were trained on, kept as they were. They are not needed to load a model.