Hirokazu Masataki’s research while affiliated with NTT Communication Science Laboratories and other places

... context information) as well as fixed-length output. It is now possible to treat a discrete symbolic series of variable length as a continuous vector of fixed length [5,6,7,8,9,10]. ...

Reference:

A Proposal for a Method of Determining Contextual Semantic Frames by Understanding the Mutual Objectives and Situations Between Speech Recognition and Interlocutors

... The authors in [2] discuss the possibility of differentiation between read and spontaneous speech by just looking at the intonation or prosody. Read and spontaneous speech classification based on variance of GMM supervectors has been studied in [1]. From a speaker role characterization perspective, in [6] the authors use acoustic and linguistic features derived from an automatic speech recognition system to characterize and detect spontaneous speech. ...

Reference:

A Novel Scheme to classify Read and Spontaneous Speech

... In contrast, latent words LMs (LWLMs) (Deschacht et al., 2012) are clearly effective for outof domain tasks. We employed the LWLM to speech recognition and the resulting performance was significantly superior in out-of domain tasks while the performance was comparable in domainmatched task to conventional LMs (Masumura et al., 2013a;Masumura et al., 2013b). LWLMs are generative models that employ a latent word space. ...

Reference:

Hierarchical Latent Words Language Models for Robust Modeling to Out-Of Domain Tasks

... Using linguistic features, such as part-of-speech tags and semantic role labels, they proposed a method to estimate these intentions through logistic regression assuming that users' utterances are given in the form of explicit sentences. Conventionally, supplemental prosodic features such as F0 have been used for intention recognition [7,8,9,10]. Fujie et al. proposed a method for estimating whether the user's attitude to the system is positive or negative via Bayesian discrimination using para-linguistic information such as fundamental frequency (F0) [4]. ...

Reference:

Recognition of Intentions of Users’ Short Responses for Conversational News Delivery System

... Wei et al. (2014aWei et al. ( ,b, 2013 use submodular function-based subset selection on generated transcripts to find a minimal set of ASR training data and Wu et al. (2007) use an entropy measure for the same. Asami et al. (2015) employ a joint Kullback-Leibler divergencebased subset selection on out-of-domain samples for ASR adaptation across acoustic characteristics such as speaker, noise and recording devices. Similarly, Liu et al. (2015) study subset selection to obtain low-vocabulary speech corpora for ASR, while Kirchhoff and Bilmes (2014) use a submodular approach for data selection in machine translation. ...

Reference:

DITTO: Data-efficient and Fair Targeted Subset Selection for ASR Accent Adaptation

... In addition, the h-LWLMs are related to other extended modeling of LWLMs. One related modeling is latent words RNN LMs (LWRNNLMs) [23], [24] that use the RNN modeling for latent variable modeling instead of n-gram modeling. The h-LWLMs differ from these extended models by taking into account the hierarchical structure of latent words. ...

Reference:

Hierarchical Latent Words Language Models for Automatic Speech Recognition

... They achieved 0.29% and 0.51% reductions in WERs. A Viterbi approximation of latent word language models (LWLMs) for ASR was proposed by the authors of [32]; they concluded that the combination of an n-gram approximation method and the Viterbi approximation method improved ASR performance. Confusing words is another factor that affects the understanding of speech. ...

Reference:

Automatic Speech Recognition: Systematic Literature Review

... As a special case of SLU, spoken utterance classification (SUC) aims at classifying the observed utterance into one of the predefined semantic classes L = {l 1 , ..., l k } (Masumura et al., 2018). Thus, a semantic classifier is trained to maximize the class-posterior probability for a given observation, W = {w 1 , w 2 , ..., w j }, representing a sequence of tokens. ...

Reference:

Multimodal Audio-textual Architecture for Robust Spoken Language Understanding

... In the past, statistical machine translation (Cucu et al. 2013;D'Haro and Banchs 2016) was used for this purpose. With the development of neural network-based language models, autoregressive sequence-to-sequence models are used for error correction (Tanaka et al. 2018;Liao et al. 2023), like neural machine translation. Moreover, with the advancement of attention mechanisms (Chan et al. 2016;Inaguma and Kawahara 2023), research utilizing the Transformer architecture (Vaswani et al. 2017) for error correction has demonstrated strong performance (Mani et al. 2020;Leng et al. 2021bLeng et al. , 2023. ...

Reference:

MathSpeech: Leveraging Small LMs for Accurate Conversion in Mathematical Speech-to-Formula

... The original SR system was used in a preliminary study with an endoscopic simulator (10). In the present study, we upgraded the VoiceRex system by introducing the latest technology (11,12). We also upgraded the text-processing method and vocabulary database of the system. ...

Reference:

Speech Recognition System Generates Highly Accurate Endoscopic Reports in Clinical Practice