skip to main content

Title: Neural network vs. HMM speech recognition systems as models of human cross-linguistic phonetic perception
The way listeners perceive speech sounds is largely determined by the language(s) they were exposed to as a child. For example, native speakers of Japanese have a hard time discriminating between American English /ɹ/ and /l/, a phonetic contrast that has no equivalent in Japanese. Such effects are typically attributed to knowledge of sounds in the native language, but quantitative models of how these effects arise from linguistic knowledge are lacking. One possible source for such models is Automatic Speech Recognition (ASR) technology. We implement models based on two types of systems from the ASR literature—hidden Markov models (HMMs) and the more recent, and more accurate, neural network systems—and ask whether, in addition to showing better performance, the neural network systems also provide better models of human perception. We find that while both types of systems can account for Japanese natives’ difficulty with American English /ɹ/ and /l/, only the neural network system successfully accounts for Japanese natives’ facility with Japanese vowel length contrasts. Our work provides a new example, in the domain of speech perception, of an often observed correlation between task performance and similarity to human behavior.
Award ID(s):
Publication Date:
Journal Name:
Proceedings of the Conference on Cognitive Computational Neuroscience
Sponsoring Org:
National Science Foundation
More Like this
  1. Before they even speak, infants become attuned to the sounds of the language(s) they hear, processing native phonetic contrasts more easily than nonnative ones. For example, between 6 to 8 mo and 10 to 12 mo, infants learning American English get better at distinguishing English and [l], as in “rock” vs. “lock,” relative to infants learning Japanese. Influential accounts of this early phonetic learning phenomenon initially proposed that infants group sounds into native vowel- and consonant-like phonetic categories—like and [l] in English—through a statistical clustering mechanism dubbed “distributional learning.” The feasibility of this mechanism for learning phonetic categories has been challenged, however. Here, we demonstrate that a distributional learning algorithm operating on naturalistic speech can predict early phonetic learning, as observed in Japanese and American English infants, suggesting that infants might learn through distributional learning after all. We further show, however, that, contrary to the original distributional learning proposal, our model learns units too brief and too fine-grained acoustically to correspond to phonetic categories. This challenges the influential idea that what infants learn are phonetic categories. More broadly, our work introduces a mechanism-driven approach to the study of early phonetic learning, together with a quantitative modeling framework that can handlemore »realistic input. This allows accounts of early phonetic learning to be linked to concrete, systematic predictions regarding infants’ attunement.« less
  2. We evaluate several publicly available off-the-shelf (commercial and research) automatic speech recognition (ASR) systems on dialogue agent-directed English speech from speakers with General American vs. non-American accents. Our results show that the performance of the ASR systems for non-American accents is considerably worse than for General American accents. Depending on the recognizer, the absolute difference in performance between General American accents and all non-American accents combined can vary approximately from 2% to 12%, with relative differences varying approximately between 16% and 49%. This drop in performance becomes even larger when we consider specific categories of non-American accents indicating a need for more diligent collection of and training on non-native English speaker data in order to narrow this performance gap. There are performance differences across ASR systems, and while the same general pattern holds, with more errors for non-American accents, there are some accents for which the best recognizer is different than in the overall case. We expect these results to be useful for dialogue system designers in developing more robust inclusive dialogue systems, and for ASR providers in taking into account performance requirements for different accents.
  3. Purpose The “bubble noise” technique has recently been introduced as a method to identify the regions in time–frequency maps (i.e., spectrograms) of speech that are especially important for listeners in speech recognition. This technique identifies regions of “importance” that are specific to the speech stimulus and the listener, thus permitting these regions to be compared across different listener groups. For example, in cross-linguistic and second-language (L2) speech perception, this method identifies differences in regions of importance in accomplishing decisions of phoneme category membership. This research note describes the application of bubble noise to the study of language learning for 3 different language pairs: Hindi English bilinguals' perception of the /v/–/w/ contrast in American English, native English speakers' perception of the tense/lax contrast for Korean fricatives and affricates, and native English speakers' perception of Mandarin lexical tone. Conclusion We demonstrate that this technique provides insight on what information in the speech signal is important for native/first-language listeners compared to nonnative/L2 listeners. Furthermore, the method can be used to examine whether L2 speech perception training is effective in bringing the listener's attention to the important cues.
  4. Training acoustic models with sequentially incoming data – while both leveraging new data and avoiding the forgetting effect – is an essential obstacle to achieving human intelligence level in speech recognition. An obvious approach to leverage data from a new domain (e.g., new accented speech) is to first generate a comprehensive dataset of all domains, by combining all available data, and then use this dataset to retrain the acoustic models. However, as the amount of training data grows, storing and retraining on such a large-scale dataset becomes practically impossible. To deal with this problem, in this study, we study several domain expansion techniques which exploit only the data of the new domain to build a stronger model for all domains. These techniques are aimed at learning the new domain with a minimal forgetting effect (i.e., they maintain original model performance). These techniques modify the adaptation procedure by imposing new constraints including (1) weight constraint adaptation (WCA): keeping the model parameters close to the original model parameters; (2) elastic weight consolidation (EWC): slowing down training for parameters that are important for previously established domains; (3) soft KL-divergence (SKLD): restricting the KL-divergence between the original and the adapted model output distributions; andmore »(4) hybrid SKLD-EWC: incorporating both SKLD and EWC constraints. We evaluate these techniques in an accent adaptation task in which we adapt a deep neural network (DNN) acoustic model trained with native English to three different English accents: Australian, Hispanic, and Indian. The experimental results show that SKLD significantly outperforms EWC, and EWC works better than WCA. The hybrid SKLD-EWC technique results in the best overall performance.« less
  5. Lierler, Yuliya ; Morales, Jose F ; Dodaro, Carmine ; Dahl, Veroniica ; Gebser, Martin ; Tekle, Tuncay (Ed.)
    Knowledge representation and reasoning (KRR) systems represent knowledge as collections of facts and rules. Like databases, KRR systems contain information about domains of human activities like industrial enterprises, science, and business. KRRs can represent complex concepts and relations, and they can query and manipulate information in sophisticated ways. Unfortunately, the KRR technology has been hindered by the fact that specifying the requisite knowledge requires skills that most domain experts do not have, and professional knowledge engineers are hard to find. One solution could be to extract knowledge from English text, and a number of works have attempted to do so (OpenSesame, Google's Sling, etc.). Unfortunately, at present, extraction of logical facts from unrestricted natural language is still too inaccurate to be used for reasoning, while restricting the grammar of the language (so-called controlled natural language, or CNL) is hard for the users to learn and use. Nevertheless, some recent CNL-based approaches, such as the Knowledge Authoring Logic Machine (KALM), have shown to have very high accuracy compared to others, and a natural question is to what extent the CNL restrictions can be lifted. In this paper, we address this issue by transplanting the KALM framework to a neural natural languagemore »parser, mStanza. Here we limit our attention to authoring facts and queries and therefore our focus is what we call factual English statements. Authoring other types of knowledge, such as rules, will be considered in our followup work. As it turns out, neural network based parsers have problems of their own and the mistakes they make range from part-of-speech tagging to lemmatization to dependency errors. We present a number of techniques for combating these problems and test the new system, KALMFL (i.e., KALM for factual language), on a number of benchmarks, which show KALMFL achieves correctness in excess of 95%.« less