SBE-RCUK: CompCog: Modeling the Development of Phonetic Representations

听众对语音的处理是根据他们的母语来调整的。例如，对英语[l]和[r]进行分类的日本听众不依赖于与英语母语听众所依赖的语音信号的相同方面。本项目使用计算模型来研究儿童如何发展语言特定的感知策略。更好地理解这种感知学习过程可以更好地诊断和治疗具有感知基础的发展性语言障碍，并可以深入了解成年后学习第二语言时听众面临的困难。建立儿童如何从周围的语音中学习母语的计算模型也可以改善低资源语言（世界上没有多少人使用的语言或缺乏大规模注释数据库等数字资源的语言）的语音技术，最终导致系统使用很少或没有转录的音频更有效地学习。这样的系统可以成为记录和分析濒危语言和少数民族语言的重要工具，并有助于使语音技术更普遍地可用。一系列模拟测试的假设，即儿童的语音处理可以成为专门为他们的母语，通过一个过程的维度学习，不依赖于声音类别的知识。提出了两种使用维度学习的模型，借鉴了在低资源自动语音识别中表现良好的表示学习方法，其中没有大量的标记训练数据。第一个模型依赖于时间信息作为声音类别知识的代理，而第二个模型依赖于自上而下的信息，从类似的话，婴儿已被证明使用。每个模型都是在特定语言的语音记录上训练的，并评估其预测具有该语言背景的成人和婴儿如何区分声音的能力。这项研究将产生新的方法来训练和测试自然主义语音录音的语言认知模型，并有可能显着影响儿童如何以及何时学习母语声音的理论。

Listeners' processing of speech is tuned to their native language. For example, Japanese listeners categorizing English [l] and [r] do not rely on the same aspects of the speech signal that native English listeners do. This project uses computational models to investigate how children develop language-specific perceptual strategies. A better understanding of this perceptual learning process could lead to better diagnosis and treatment of developmental language impairments that have a perceptual basis and can provide insight into the difficulties that listeners face when learning a second language in adulthood. Building computational models of how children learn their native language from the speech around them can also lead to improved speech technology for low-resource languages (languages that are not spoken by many people in the world or that lack digital resources such as large-scale, annotated databases), ultimately leading to systems that learn more effectively using little or no transcribed audio. Such systems could become important tools for documenting and analyzing endangered and minority languages and could help make speech technology more universally available.A series of simulations tests the hypothesis that children's processing of speech can become specialized for their native language through a process of dimension learning that does not rely on knowledge of sound categories. Two models that use dimension learning are proposed, drawing on representation learning methods that have performed well in low-resource automatic speech recognition, where extensive labeled training data are not available. The first model relies on temporal information as a proxy for sound category knowledge, while the second model relies on top-down information from similar words, which infants have been shown to use. Each model is trained on speech recordings from a particular language and is evaluated on its ability to predict how adults and infants with that language background discriminate sounds. The research will yield new methods for training and testing cognitive models of language with naturalistic speech recordings and has the potential to significantly impact theories of how and when children learn about the sounds of their native language.