Modeling generalization in perceptual learning of speech

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Science & Research — Mathematics & Computational Sciences, Social Sciences & Behavioral Studies, Research Methodology & Innovation · Depth: Expert, quick

Summary

A study on speech perception generalization proposes that listeners treat generalization as an inference problem, rather than solely relying on acoustic similarity. Researchers exposed listeners to shifted vowel pronunciations and tested them on minimal pairs varying in lexical frequency. They found learning effects were strongest when the exposure direction aligned with a high-frequency alternative in mixed-frequency pairs, but absent for low-frequency pairs. To explain this, three Bayesian belief-updating models were formalized: a talker-specific model, a mixture-of-expectations model, and a hierarchical Bayesian model. The talker-specific model captured most generalization patterns due to its sensitivity to token-level acoustic properties but overpredicted learning for low-frequency pairs. The hierarchical model best recovered the exposure-control contrast, indicating lexical frequency may constrain how learned representations are applied. This work offers a computationally explicit framework for understanding contextual factors in speech perception generalization.

Key takeaway

For research scientists modeling speech perception, this work suggests you should incorporate lexical frequency as a critical contextual factor. Your models of generalization should move beyond purely acoustic similarity, considering how listeners infer the applicability of learned phonetic mappings. This implies that robust speech perception models need to account for uncertainty in applying updated representations, particularly across varying lexical contexts.

Key insights

Generalization in speech perception is an inference problem, constrained by lexical frequency, not just acoustic similarity.

Principles

• Lexical frequency constrains learned representation application.
• Generalization involves inferring mapping applicability.
• Acoustic similarity alone is insufficient for generalization.

Method

A Bayesian belief-updating framework was used to model generalization, comparing talker-specific, mixture-of-expectations, and hierarchical Bayesian models against experimental data.

Topics

• Speech Perception
• Generalization Learning
• Bayesian Modeling
• Lexical Frequency
• Phonetic Mapping
• Acoustic Similarity

Best for: AI Scientist, Research Scientist

Related on AIssential

• Perceptual compensation for tonal context in self-supervised speech models — Self-supervised speech models like wav2vec2.0 may require supervised fine-tuning to achieve perceptual compensation for complex phonological features like Mandarin tones.
• A Theoretical Framework for Acoustic Neighbor Embeddings — A theoretical framework provides a probabilistic interpretation of acoustic neighbor embedding distances, simplifying phonetic similarity measurement.
• N-Grams and Markov Assumptions: The First Predictive Models of Language — N-grams were the first predictive language models, using statistical counts over short word sequences to anticipate the next word.
• SpeechDx: A Multi-Task Benchmark for Clinical Speech AI — SpeechDx establishes a multi-task benchmark to evaluate and advance general-purpose clinical speech AI representations.
• Lexical Tone is Hard to Quantize: Probing Discrete Speech Units in Mandarin and Yor\`ub\'a — Current DSU quantization methods struggle to reliably encode lexical tone, prioritizing phonetic structure instead.
• Prior over Evidence: Stereotype-Driven Diagnosis in LLM-Based L2 Pronunciation Feedback — LLMs often prioritize pretraining stereotypes over actual acoustic evidence in L2 pronunciation diagnosis.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.