Greenberg Steven, Christiansen Thomas U
Silicon Speech, Hidden Valley Lake, CA, 95467, USA.
Oticon, Kongebakken 9, DK-2765, Smørum, Denmark.
Atten Percept Psychophys. 2019 May;81(4):884-896. doi: 10.3758/s13414-019-01666-y.
Over a long and distinguished career, Randy Diehl has elucidated the brain mechanisms underlying spoken language processing. The present study touches on two of Randy's central interests, phonetic features and Bayesian statistics. How does the brain go from sound to meaning? Traditional approaches to the study of speech intelligibility and word recognition are unlikely to provide a definitive answer. A finer-grained, Bayesian-inspired approach may help. In this study, listeners identified 11 Danish consonants spoken in a Consonant + Vowel + [l] environment. Each syllable was filtered so that only a portion of the original audio spectrum was presented. Three-quarter-octave bands of speech, centered at 750, 1,500, and 3,000 Hz, were presented individually and in combination. The conditional, posterior probabilities associated with decoding the phonetic-features Voicing, Manner, and Place of Articulation were computed from confusion matrices to delineate the perceptual flow of phonetic information processing. Analysis of the conditional probabilities associated with both correct and incorrect feature decoding suggest that Manner of articulation is linked to the decoding of Voicing (but not vice-versa), and that decoding of Place of articulation is associated with decoding of Manner of articulation (but not the converse). Such feature-decoding asymmetries may reflect processing strategies in which the decoding of lower-level features, such as Voicing and Manner, is leveraged to enhance the recognition of more complex linguistic elements (e.g., phonetic segments, syllables, and words), especially in adverse listening conditions. Such asymmetric feature decoding patterns are consistent with a hierarchical, perceptual flow model of phonetic processing.
在漫长而卓越的职业生涯中，兰迪·迪尔阐明了口语处理背后的大脑机制。本研究涉及兰迪的两个核心兴趣点：语音特征和贝叶斯统计。大脑是如何从声音转换为意义的？传统的语音清晰度和单词识别研究方法不太可能给出一个明确的答案。一种更精细的、受贝叶斯启发的方法可能会有所帮助。在这项研究中，听众识别了在辅音+元音+[l]环境中说出的11个丹麦辅音。每个音节都经过滤波，以便只呈现原始音频频谱的一部分。以750、1500和3000赫兹为中心的四分之三倍频程语音频段分别单独呈现和组合呈现。根据混淆矩阵计算与解码语音特征浊音、发音方式和发音部位相关的条件后验概率，以描绘语音信息处理的感知流程。对与正确和错误特征解码相关的条件概率的分析表明，发音方式与浊音解码相关（反之则不然），发音部位的解码与发音方式的解码相关（反之则不然）。这种特征解码不对称可能反映了一种处理策略，即利用较低层次特征（如浊音和发音方式）的解码来增强对更复杂语言元素（如语音片段、音节和单词）的识别，尤其是在不利的听力条件下。这种不对称特征解码模式与语音处理的分层感知流模型一致。
