Noraini Seman
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Noraini Seman has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Nursuriati Jamil
Universiti Teknologi MARA
14 shared publications Zainab Abu Bakar
7 shared publicationsNordin Abu Bakar
Universiti Teknologi MARA (UiTM)
5 shared publicationsNorizah Ardi
Academy of Languages Studies, Universiti Teknologi MARA, 40450 Shah Alam, Selangor, Malaysia
5 shared publicationsIzzad Ramli
Faculty of Computer and Mathematical Sciences, Universiti Teknologi MARA, Malaysia
4 shared publicationsRaseeda Hamzah
Computer Science Department, Faculty of Computer & Mathematical Sciences, Universiti Teknologi MARA (UiTM), 40450 Shah Alam, Selangor Darul Ehsan, Malaysia
4 shared publications
Latest Publications
- Open Access
- 25 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 388-393
Enhancing Pretrained Multilingual Machine Translation Model with Code-Switching: A Study on Chinese, English and Malay Language
In the field of multilingual machine translation, many pretrained language models have achieved the inspiring results. However, the results based on pretrained models are not yet very satisfactory for low-resource languages. This paper investigates how to leverage code-switching data to fine-tune pretrained multilingual machine translation model, in order to boost the performance of few-shot low resource machine translation. By utilizing a multilingual mixed corpus, the code-switching method can enhance the cross-linguistic generalization ability of the model and improve its overall understanding of the language. In this paper, on the smaller size model of FLORES-101 benchmark, we use the code-switching data augmentation method to achieve the results of benchmark's larger model on six direction pairs of three languages, Chinese, English and Malay. This paper studied various corpus mixture mechanisms to construct the data in code-switching, and the experimental findings show that the results using the code-switching fine-tuning model improve the spBLEU score by an average of 2 to 3 points over the results without code-switching.
- 25 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 388-393
Open AccessEnhancing Pretrained Multilingual Machine Translation Model with Code-Switching: A Study on Chinese, English and Malay Language
In the field of multilingual machine translation, many pretrained language models have achieved the inspiring results. However, the results based on pretrained models are not yet very satisfactory for low-resource languages. This paper investigates how to leverage code-switching data to fine-tune pretrained multilingual machine translation model, in order to boost the performance of few-shot low resource machine translation. By utilizing a multilingual mixed corpus, the code-switching method can enhance the cross-linguistic generalization ability of the model and improve its overall understanding of the language. In this paper, on the smaller size model of FLORES-101 benchmark, we use the code-switching data augmentation method to achieve the results of benchmark's larger model on six direction pairs of three languages, Chinese, English and Malay. This paper studied various corpus mixture mechanisms to construct the data in code-switching, and the experimental findings show that the results using the code-switching fine-tuning model improve the spBLEU score by an average of 2 to 3 points over the results without code-switching.
- Open Access
- 1 October 2021
- Published by UUM Press, Universiti Utara Malaysia in Journal of Information and Communication Technology
- Vol. 20 (4) , 489-510
An Iterated Two-Step Sinusoidal Pitch Contour Formulation for Expressive Speech Synthesis
Intonation generation in expressive speech such as storytelling is essential to produce high quality Malay language expressive speech synthesizer. Intonation generation, for instance explicit control, has shown good performance in terms of intelligibility with reasonably natural speech; thus, it was selected in this research. This approach modifies the prosodic features, such as pitch contour, intensity, and duration, to generate the intonation. However, modification of pitch contour remains a problem because the desired pitch contour is not achieved. This paper formulated an improved pitch contour algorithm to develop a modified pitch contour resembling the natural pitch contour. In this work, the syllable pitch contours of nine storytellers were extracted from their storytelling speeches to create an expressive speech syllable dataset called STORY_DATA. All the shapes of pitch contours from STORY_DATA were analyzed and clustered into the standard six main pitch contour clusters for storytelling. The clustering was performed using one minus the Pearson product moment correlation. Then, an improved iterative two-step sinusoidal pitch contour formulation was introduced to modify the pitch contours of a neutral speech into an expressive pitch contour of natural speeches. Overall, the improved pitch contour formulation was able to achieve 93 percent high correlated matches, indicating the high resemblance as compared to the previous pitch contour formulation at 15 percent. Therefore, the improved formula can be used in a text-to-speech (TTS) synthesizer to produce a more natural expressive speech. The paper also discovered unique expressive pitch contours in the Malay language that need further investigations in the future.
- 1 October 2021
- Published by UUM Press, Universiti Utara Malaysia in Journal of Information and Communication Technology
- Vol. 20 (4) , 489-510
Open AccessAn Iterated Two-Step Sinusoidal Pitch Contour Formulation for Expressive Speech Synthesis
Intonation generation in expressive speech such as storytelling is essential to produce high quality Malay language expressive speech synthesizer. Intonation generation, for instance explicit control, has shown good performance in terms of intelligibility with reasonably natural speech; thus, it was selected in this research. This approach modifies the prosodic features, such as pitch contour, intensity, and duration, to generate the intonation. However, modification of pitch contour remains a problem because the desired pitch contour is not achieved. This paper formulated an improved pitch contour algorithm to develop a modified pitch contour resembling the natural pitch contour. In this work, the syllable pitch contours of nine storytellers were extracted from their storytelling speeches to create an expressive speech syllable dataset called STORY_DATA. All the shapes of pitch contours from STORY_DATA were analyzed and clustered into the standard six main pitch contour clusters for storytelling. The clustering was performed using one minus the Pearson product moment correlation. Then, an improved iterative two-step sinusoidal pitch contour formulation was introduced to modify the pitch contours of a neutral speech into an expressive pitch contour of natural speeches. Overall, the improved pitch contour formulation was able to achieve 93 percent high correlated matches, indicating the high resemblance as compared to the previous pitch contour formulation at 15 percent. Therefore, the improved formula can be used in a text-to-speech (TTS) synthesizer to produce a more natural expressive speech. The paper also discovered unique expressive pitch contours in the Malay language that need further investigations in the future.
- Open Access
- 1 February 2021
- Published by Institute of Advanced Engineering and Science in Bulletin of Electrical Engineering and Informatics
- Vol. 10 (1) , 365-373
Visualizing stemming techniques on online news articles text analytics
Stemming is the process to convert words into their root words by the stemming algorithm. It is one of the main processes in text analytics where the text data needs to go through stemming process before proceeding to further analysis. Text analytics is a very common practice nowadays that is practiced toanalyze contents of text data from various sources such as the mass media and media social. In this study, two different stemming techniques; Porter and Lancaster are evaluated. The differences in the outputs that are resulted from the different stemming techniques are discussed based on the stemming error and the resulted visualization. The finding from this study shows that Porter stemming performs better than Lancaster stemming, by 43%, based on the stemming error produced. Visualization can still be accommodated by the stemmed text data but some understanding of the background on the text data is needed by the tool users to ensure that correct interpretation can be made on the visualization outputs.
- 1 February 2021
- Published by Institute of Advanced Engineering and Science in Bulletin of Electrical Engineering and Informatics
- Vol. 10 (1) , 365-373
Open AccessVisualizing stemming techniques on online news articles text analytics
Stemming is the process to convert words into their root words by the stemming algorithm. It is one of the main processes in text analytics where the text data needs to go through stemming process before proceeding to further analysis. Text analytics is a very common practice nowadays that is practiced toanalyze contents of text data from various sources such as the mass media and media social. In this study, two different stemming techniques; Porter and Lancaster are evaluated. The differences in the outputs that are resulted from the different stemming techniques are discussed based on the stemming error and the resulted visualization. The finding from this study shows that Porter stemming performs better than Lancaster stemming, by 43%, based on the stemming error produced. Visualization can still be accommodated by the stemmed text data but some understanding of the background on the text data is needed by the tool users to ensure that correct interpretation can be made on the visualization outputs.
- Open Access
- 1 October 2019
- Published by Institute of Advanced Engineering and Science in Indonesian Journal of Electrical Engineering and Computer Science
- Vol. 16 (1) , 283-291
Hybrid methods of brandt’s generalised likelihood ratio and short-term energy for malay word speech segmentation
Speech segmentation is an important part for speech recognition, synthesizing and coding. Statistical based approach detects segmentation points via computing spectral distortion of the signal without prior knowledge of the acoustic information proved to be able to give good match, less omission but lot of insertion. In this study the segmentation is done both manually and automatically using Malay words in traditional Malay poetry. This study proposed a hybrid method of Brandt’s generalized likelihood ratio (GLR) and short-term energy algorithm. The Brandt’s algorithm tries to estimate the abrupt change in energy to determine the segmentation points. A total of five Pantun are used in read mode and spoken by one male student in a noise free room. Experiments are conducted to see the the accuracy, insertion, and omission of the segmentation points. Experimental results show on average 80% accuracy with 0.2 second time tolerance for automatic segmentation with the algorithm having no knowledge of the acoustic characteristics.
- 1 October 2019
- Published by Institute of Advanced Engineering and Science in Indonesian Journal of Electrical Engineering and Computer Science
- Vol. 16 (1) , 283-291
Open AccessHybrid methods of brandt’s generalised likelihood ratio and short-term energy for malay word speech segmentation
Speech segmentation is an important part for speech recognition, synthesizing and coding. Statistical based approach detects segmentation points via computing spectral distortion of the signal without prior knowledge of the acoustic information proved to be able to give good match, less omission but lot of insertion. In this study the segmentation is done both manually and automatically using Malay words in traditional Malay poetry. This study proposed a hybrid method of Brandt’s generalized likelihood ratio (GLR) and short-term energy algorithm. The Brandt’s algorithm tries to estimate the abrupt change in energy to determine the segmentation points. A total of five Pantun are used in read mode and spoken by one male student in a noise free room. Experiments are conducted to see the the accuracy, insertion, and omission of the segmentation points. Experimental results show on average 80% accuracy with 0.2 second time tolerance for automatic segmentation with the algorithm having no knowledge of the acoustic characteristics.
- 1 May 2017
- Published by American Scientific Publishers in Advanced Science Letters
- Vol. 23 (5) , 4202-4205
Mouth Segmentation of Viseme Using Biased Normalized Cuts and Mathematical Morphology
Traditionally, work on Automatic Speech Recognition (ASR) focused in recognizing speech using spoken audio data only which is susceptible to noise. The new approach of Audio-Visual Speech Recognition (AVSR) is made to mimic the human lip reading and it can recognize the spoken words based on the lip images, known asviseme. The mouth segmentation is a vital step in AVSR which is used further for recognition of the phonemes from the viseme images. Detecting mouth shape accurately is a difficult task since mouth is highly misshapen and different in shape, size and color. Thus, this paper proposed using biased normalized cuts and mathematical morphology algorithms to segment the mouth region of viseme images. Experiments are conducted on 80 images comprising 40 mouth images of a female and 40 mouth images of a male. This proposed method of biased normalized cuts and mathematical morphology are identified as a promising method in segmenting the mouth region of viseme images with an accuracy rate of 80.66% and 79.8% for female and male, respectively.
- 1 May 2017
- Published by American Scientific Publishers in Advanced Science Letters
- Vol. 23 (5) , 4202-4205
Mouth Segmentation of Viseme Using Biased Normalized Cuts and Mathematical Morphology
Traditionally, work on Automatic Speech Recognition (ASR) focused in recognizing speech using spoken audio data only which is susceptible to noise. The new approach of Audio-Visual Speech Recognition (AVSR) is made to mimic the human lip reading and it can recognize the spoken words based on the lip images, known asviseme. The mouth segmentation is a vital step in AVSR which is used further for recognition of the phonemes from the viseme images. Detecting mouth shape accurately is a difficult task since mouth is highly misshapen and different in shape, size and color. Thus, this paper proposed using biased normalized cuts and mathematical morphology algorithms to segment the mouth region of viseme images. Experiments are conducted on 80 images comprising 40 mouth images of a female and 40 mouth images of a male. This proposed method of biased normalized cuts and mathematical morphology are identified as a promising method in segmenting the mouth region of viseme images with an accuracy rate of 80.66% and 79.8% for female and male, respectively.
- Open Access
- 3 October 2016
- Published by EDP Sciences in MATEC Web of Conferences
- Vol. 77, 04003
Rule-Based Storytelling Text-to-Speech (TTS) Synthesis
MATEC Web of Conferences, open access proceedings in Materials science, Engineering and Chemistry
- 3 October 2016
- Published by EDP Sciences in MATEC Web of Conferences
- Vol. 77, 04003
Open AccessRule-Based Storytelling Text-to-Speech (TTS) Synthesis
MATEC Web of Conferences, open access proceedings in Materials science, Engineering and Chemistry
- 13 August 2016
- Published by Springer Nature
- p. 563-570
Prosody Analysis of Malay Language Storytelling Corpus
- 13 August 2016
- Published by Springer Nature
- p. 563-570
Prosody Analysis of Malay Language Storytelling Corpus
- 17 November 2015
- Published by Springer Nature
- p. 127-139
Bimodality Streams Integration for Audio-Visual Speech Recognition Systems
- 17 November 2015
- Published by Springer Nature
- p. 127-139
Bimodality Streams Integration for Audio-Visual Speech Recognition Systems
- Open Access
- 1 January 2015
- Published by Elsevier in Procedia Computer Science
- Vol. 76, 417-424
An Improved Syllabification for a Better Malay Language Text-to-Speech Synthesis (TTS)
- 1 January 2015
- Published by Elsevier in Procedia Computer Science
- Vol. 76, 417-424
Open AccessAn Improved Syllabification for a Better Malay Language Text-to-Speech Synthesis (TTS)
- Open Access
- 1 January 2015
- Published by Elsevier in Procedia Computer Science
- Vol. 51, 620-629
Blending Sentence Optimization Weights of Unsupervised Approaches for Extractive Speech Summarization
- 1 January 2015
- Published by Elsevier in Procedia Computer Science
- Vol. 51, 620-629
Open AccessBlending Sentence Optimization Weights of Unsupervised Approaches for Extractive Speech Summarization