Simulating Early Phonetic and Word Learning Without Linguistic Categories

Area of Science:

• Cognitive Science
• Developmental Psychology
• Computational Linguistics

Background:

• Infants develop sensitivity to native language sounds and word forms before speaking.
• Traditional views attribute these changes to acquiring linguistic categories (phonemes, words).
• Skepticism exists due to limited evidence of early category knowledge in infants.

Purpose of the Study:

• To investigate if linguistic categories are necessary for early word learning.
• To propose and test a computational model that explains infant speech perception development.
• To challenge the necessity of pre-existing linguistic categories in early language acquisition.

Main Methods:

• Introduction of STELA, a predictive coding algorithm.
• STELA extracts statistical patterns from continuous raw speech data.
• Analysis of learned representations to understand emergent properties.

Main Results:

• STELA successfully reproduces developmental patterns in phonetic and word form learning.
• The algorithm functions without relying on explicit linguistic categories (phonemes, words).
• No requirement for explicit word segmentation was needed for STELA's learning.

Conclusions:

• Linguistic categories may not be prerequisites for early language acquisition.
• Statistical pattern extraction from continuous speech is a viable mechanism for early learning.
• Linguistic categories might emerge as a result of learning, not as an initial requirement.