Citation Context

...s in the performance on native speech and non-native speech. We use one filter based algorithm, namely Minimum Redundancy Maximum Relevance (MRMR) [13], and one wrapper based, Genetic Algorithms (GA) =-=[14]-=- to achieve this. These features are then used to train SVM models with Gaussian kernels, one of the better performing machine learning algorithms for binary class problems. The classifier system, thu...

Citation Context

...errors, by finding significant performance differences in the performance on native speech and non-native speech. We use one filter based algorithm, namely Minimum Redundancy Maximum Relevance (MRMR) =-=[13]-=-, and one wrapper based, Genetic Algorithms (GA) [14] to achieve this. These features are then used to train SVM models with Gaussian kernels, one of the better performing machine learning algorithms ...

Citation Context

...ASR is not able to recognize the pronounced word correctly, unless it is pronounced with a native accent. Explicit feedback can be in the form of a score, for example, Goodness of Pronunciation (GOP) =-=[2, 3]-=- which gives the students a reasonable feedback on their performance. While many such CAPT systems provide scores that are highly correlated to scores provided by humans [4], this sort of feedback doe...

Citation Context

...uage Learning (CALL) systems. While many CAPT systems have focussed on providing intelligent training methods to help improve pronunciation, some other systems rely on providing feedback to the users =-=[1]-=-. A few methods use Automatic Speech Recognition (ASR) to give implicit or explicit feedback regarding the quality of the pronunciation. Implicit feedback is provided when the ASR is not able to recog...

Citation Context

...ated to scores provided by humans [4], this sort of feedback does not tell the student what the specific mistake is or what he or she should change in the pronunciation in order to get a higher score =-=[5]-=-. Other systems like the ISLE project [6] try to provide explicit feedback based on the specific first language (L1) and second langauge (L2) pair, by highlighting the locus of the error in the word. ...

Citation Context

... of Pronunciation (GOP) [2, 3] which gives the students a reasonable feedback on their performance. While many such CAPT systems provide scores that are highly correlated to scores provided by humans =-=[4]-=-, this sort of feedback does not tell the student what the specific mistake is or what he or she should change in the pronunciation in order to get a higher score [5]. Other systems like the ISLE proj...

Citation Context

...is sort of feedback does not tell the student what the specific mistake is or what he or she should change in the pronunciation in order to get a higher score [5]. Other systems like the ISLE project =-=[6]-=- try to provide explicit feedback based on the specific first language (L1) and second langauge (L2) pair, by highlighting the locus of the error in the word. Using such explicit knowledge, specific a...

Citation Context

...plicit knowledge, specific and typical pronunciation errors may be accurately detected, but idiosyncratic errors are often ignored. Several machine learning approcahes like Hidden Markov Models (HMM) =-=[7]-=-, Linear Discriminant Analysis (LDA) [8] and Support Vector Machines [9] have been used to detect errors as well as to provide feedback. While reported performances of such methods in CAPT systems hav...

Citation Context

... described in [11]. Given the acoustic signal and the text of what the subject is supposed to have uttered, the acoustic signal is segmented into the sequence of phonemes using an HMM based alignment =-=[12]-=-. We use the native speech for training our models and test them on non-native speech uttered by the L2 language learner. The input to the classification framework are the acoustic segments of individ...

Citation Context

...ately detected, but idiosyncratic errors are often ignored. Several machine learning approcahes like Hidden Markov Models (HMM) [7], Linear Discriminant Analysis (LDA) [8] and Support Vector Machines =-=[9]-=- have been used to detect errors as well as to provide feedback. While reported performances of such methods in CAPT systems have been improving rapidly, the low accuracy of such feedback systems may ...

Citation Context

... of such feedback systems may discourage students from utilizing them [4]. Another approach commonly used in CAPT systems is to provide feedback on the fluency of the speaker on a global scale (e.g., =-=[10]-=-). The score can be given for a sentence, or group of utterances, based on segmental, or prosodic features. Some of the studies use known texts, while others make use of the students L1 to make sound ...

Citation Context

...ASR is not able to recognize the pronounced word correctly, unless it is pronounced with a native accent. Explicit feedback can be in the form of a score, for example, Goodness of Pronunciation (GOP) =-=[2, 3]-=- which gives the students a reasonable feedback on their performance. While many such CAPT systems provide scores that are highly correlated to scores provided by humans [4], this sort of feedback doe...

Citation Context

...ronunciation errors may be accurately detected, but idiosyncratic errors are often ignored. Several machine learning approcahes like Hidden Markov Models (HMM) [7], Linear Discriminant Analysis (LDA) =-=[8]-=- and Support Vector Machines [9] have been used to detect errors as well as to provide feedback. While reported performances of such methods in CAPT systems have been improving rapidly, the low accura...

Citation Context

...terances while trying to mimic the words or sentences uttered by the virtual tutor. More details on the test subjects and their performances before and after the training using VILLE are described in =-=[15]-=-. The data was cleaned up to remove instances of hesitations or completely incorrect utterances in the data. In this experiment, the native and non-native speakers recorded the same set of utterances,...

Citation Context

...d confusing for the particular student. 2. Ensemble of Classifiers The block diagram of the ensemble classification framework we used in this study is illustrated in Figure 1, previously described in =-=[11]-=-. Given the acoustic signal and the text of what the subject is supposed to have uttered, the acoustic signal is segmented into the sequence of phonemes using an HMM based alignment [12]. We use the n...