Machine Learning For Non-Native Human Spoken-Word Recognition

This study establishes that machine learning can be cultivated in a supervised or unsupervised approach. When likened to supervised learning, unsupervised techniques perform in the beginning, when they are anomalous, but as they tune and adjust themselves, performance upturns. However, in the use of various machine learning techniques, there are tendencies for speech recognition problems. Speech recognition is a functioning research field in natural language handling and has many meaningful real-world applications. One of the main challenges in speech recognition is to comprehend speech by non-native speakers.

This is as a follow of the diverse accent that non-native speakers possess. However, this study also pinpoints several framework that could result in this discrepancy in speech recognition. One of which “accent” was highlighted. Accent is a soft biometric trait that can be inferred from pronunciation and articulation patterns characterizing the speaking style of an individual (Christos et al. , 2015). Furthermore, reasons were presented on why the deep learning approaches have been selected and preferred. Machine learning and deep learning algorithms can then be engaged to score spoken English after learning from the manual evaluation of the training samples. Secondary data were employed for this study. Extracted from (Kukanov, 2015). Analytical Comparison was made between CNN and DBN-DNN, and CNN was proven to be a more suitable framework. Other relevant discussions were made, and the conclusion was reasonably arrived at.

INTRODUCTION: Speech recognition is the capacity of a machine, database or program to recognize phrases, expressions or words in the spoken language and translate them to a machine-readable layout. The detection of non-native speakers’ dialog is, by itself, a very stimulating task. The discourse of speech recognition has been around for years, but the question in mind should be why it is now salient. The motivation is that deep learning has finally made speech recognition accurate enough to be useful outside of a carefully controlled environment. Machine learning is the awareness that there are universal algorithms that can tell something fascinating and attention-grabbing about a set of data shorn of you having to inscribe any custom code particular to the problem.

As an alternative of writing codes, data is inputted to the generic algorithm, and it constructs its logic grounded on the data. It suffices to mention that with actual world disposition, nonetheless, the miscellany of users, stresses serious considerations. Though the prerogative that all users should be presented identical access to speech recognition is not solid. This is synonymous to the analogy that people with pitiable reading skills do not have the equal access to newspapers as the vastly literate (Laura, 2001)In simple terms, Machine learning is a canopy term that shelters different kinds of generic algorithms. Moreover, non-native recognition plays a key need for boundary control security systems (Bennie, 2007). Such recognition systems help security officials to identify immigrants with a counterfeit and forged permit or ID by spotting actual country of articulated foreign accent. Also, it does appear that speech recognition applications are on a flight to become a default interface for information spreading systems. Housing users whose language use are somehow compromised are not just a research problem but also a momentous useful concern. Research Question: The study focused on answering the question: What is the best possible machine learning algorithm interface for speech recognition for non-native speakers considering the acoustic characteristics of the accent induced?

METHODOLOGY: There are few considerations, however that seem principally expedient for encoding non-native speech. Modes, choice, lexical, syntactic soundness, accent, and fluency are facets of spoken English that can both label disparity in native speech and be used to differentiate it from native speakers. “Accent commonly come from articulation habit of the speaker in his or her language” (Pascale & Wai Kat, 1999). As it is widely known that beginners of a language are mostly exposed to preliminary grammar in the premature stage of their study, yet, imperfect mastery of syntax is one of the sorts that can make an even vastly proficient speech as non-native. Lately, perception between native and non-native speech has been battered by means of binary classification structures. These frameworks essentially depend on prosodic, cepstral, speech recognition based or N-gram language sorts, and engage support vector machines (SVMs) for classification.

AUTOMATIC SPEECH RECOGNITION (ASR): Efforts to construct Automatic Speech Recognition (ASR) systems were first made in the 1950s. These initial speech recognition systems tried to relate a set of grammatical and syntactical rules to pinpoint speech. The system could only identify the word if the spoken words stick to a certain rule set. An Automatic Speech Recognition (ASR) module forms the root of virtually all the spoken Language evaluation systems (Yu et al. , 2015). An ASR front-end component for most state-of-the-art evaluation systems gives word speculations about the responses given by the person available for the assessment. Consequently, it can be predicted that a huge amount of data, more precisely a pool of non-native speech, and careful transcriptions of each piece of that speech, would be required to train this type of ASR module. Moreover, there are no doubts that this would implicate human effort in transcribing the whole speech collection. Despite the advancement of Automatic Speech Recognition (ASR) systems that have led to supporters is still an issue in developing robust ASR systems that deliver high performance across diverse user groups (Russell & Najafian, 2015). The problem of the present ASR systems that these are working mostly with the native speech only, and the correctness affectedly lessens when words are articulated with an unusual pronunciation (foreign accent). However, human language has copious concessions to its guidelines.

The way words and phrases are articulated can be enormously altered by dialects, assents, and mannerisms. First, there is a disparity in what is said by the speaker. For open vocabulary systems, there is no way to gather training data for every conceivable utterance or even every possible word. Second, there is dissimilarity due to differences between speakers. Different people have different voices and accents and ways of speaking. Third, there is variation in noise conditions. Anything in the acoustic data that is not the signal is noise, and thus noise can include background sounds, microphone specific artifacts, and other effects. (George, 2015) Hence, to accomplish Automatic Speech Recognition, we make use of Deep Learning Algorithm. Therefore, for this study, Deep Learning Algorithm will be considered as our methodology. It may also interest one to know that Deep learning canvassers who know almost nothing about language translation are knitting together comparatively simple machine learning resolutions that are thrashing the best expert-built language translation systems in our world today.

EXPERIMENT AND RESULT: In machine learning, a neural network (Deep learning) is a construction especially used for grouping or regression tasks when the extraordinary dimensionality and non-linearity of the data make these tasks unlikely to accomplish. In the case of a visual data, the standard is to engage Convolutional Neural Networks (CNN). CNNs are directly inspired by the hierarchy of the cells in visual neuroscience (Hubel & Wiesel, 1962). It is important to note that the Neural network itself is not an algorithm, but rather a charter for many other machine learning algorithms to work collectively and process multifarious and complicated data inputs. Siniscalchi, et al. , (2013) already established that manner and place of articulation attributes can efficiently characterize any spoken language along the same lines as in the automatic speech attribute transcription (ASAT) model for Automatic Speech Recognition (Lee & Siniscalchi, 2013)Consider the table below (extracted from (Kukanov, 2015)):(For analysis, the data-set used by Kukanov, (2015) consist of Data from each language were pooled together to obtain around 6 hours of training, 2. 5 hours of test and 0. 5 hours of cross-validation datasets. The sample rate is 8kHz, encompassing six different languages: English, Spanish, Japanese, Mandarin, Hindi, and German. In Tables 7. 1 & 7. 2, we recap the outcome DNN models of the articulatory attribute detectors, which display the lowest value of the AvgEER. Relate the performance of DNNs versus CNNs using DET plots and AvgEER, CNNs steadily outperform DNNs through a different range of tuned parameters; resulting in 18. 8% relative error reduction for manner and 10. 3% for place compare with the best DBN-DNN models. What this tells us is that CNNs have better generalization ability than DNNs particularly when the amount of training data is restricted (Jui-Ting Huang & Gong, 2015).

CONCLUSION AND DISCUSSION: Non-native human speech is multifarious; thus, it constrains quite a few studies in their research to a selected language speaking group or nation. Programmed assessment of some sides of spoken language adeptness, including grammar, content appropriateness, vocabulary, and dialog rationality, profoundly on how correctly the input speech can be accepted. While state-of-the-art acoustic models constructed on deep neural networks have meaningfully upgraded recognition performance of native speech, correct recognition results are still puzzling to achieve when the input is unstructured non-native speech. This is due, in large part, to the point that non-native spoken reactions tend to encompass substantively complex amounts of pronunciation errors and flawed phrases.

In other to determine the DNN model parameter for maximum accurateness, a good number of experiments have been carried out before extracting our data from its source. In order to measure the performance of the system of any data set and parallel the models generally, one could use the DET (Detection Error Trade-offs) curves approach or the AvgEER. Also, when the CNNs was compared with the DBN-DNNs, the CNN provides a reduction producing 18. 8 percent for the manner and 10. 3 percent for the place relative AvgEER improvement (refer to the tables above). It is not flawless that speech recognition technology has gotten to the point at which it can make judgments as to the exactness of pronunciations that corresponds to human judgments at an acceptable level (Philippe et al. , 1998). However, the truth is that our present world, with the constantly evolving technological knowledge base, we can now confidently say we are exposed to highly effective speech recognition systems like Google, Amazon, and so on, with minimal speech detection errors (even for non-native spoken-words) and somewhat close to being perfect.