Atiwong Suchato
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Atiwong Suchato has a total of 18 co-authors. Here are the top co-authors by number of shared publications:
- Proadpran Punyabukkana
Chulalongkorn University
10 shared publications - Ekapol Chuangsuwanich
Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University, Bangkok, Thailand
2 shared publications - Sirinart Tangruamsub
Spoken Language Systems Research Group, Department of Computer Engineering, Faculty of Engineering, Chulalongkorn University, Bangkok, Thailand
2 shared publications Chai Wutiwiwatchai
HLT, National Electronics and Computer Technology Center (NECTEC), Bangkok 10400, Thailand
1 shared publication- Chai Wutiwiwatchai
National Electronics and Computer Technology Center (NECTEC)
1 shared publication - Jirut Polohakul
Department of Computer Engineering, Chulalongkorn University, Bangkok, Thailand
1 shared publication
Latest Publications
- Open Access
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 90-97
Pioneering Inclusion: A Low-Cost, Low-Tech Approach to Educating Thailand's First Blind Engineering Student
- 14 July 2025
- Published by Association for Computing Machinery (ACM)
- p. 90-97
Open AccessPioneering Inclusion: A Low-Cost, Low-Tech Approach to Educating Thailand's First Blind Engineering Student
- Open Access
- 14 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 86938-86953
A Real Estate Valuation Model Using Boosted Feature Selection
To estimate real estate values, a complex valuation model based on artificial neural network (ANN) has been established as a successful means in modern machine learning research, specifically when high-dimensional data are available. Unfortunately, the real estate data in many locations, such as Thailand, are quite limited in terms of features. Hence, it becomes mandatory to reduce the complexity using feature selection techniques. These techniques aim to improve performance by identifying significant factors and help decrease the computational overload and model construction. However, due to the lack of explicability and interpretability in ANNs, the analysis of input factors cannot be explained directly by model composition. In this research, we apply a combination of a boosting strategy and input sensitivity analysis in an improved Garson's algorithm to perform feature selection that can adjust its selection criteria through each iteration on an ANN model. This proposed technique is then compared with other traditional feature selection techniques using synthetic data and real-world house valuation data. The results show that our model can maintain the sensitivity coefficient for every informative feature. The technique of this study provides a set of features that influences the house price and implies the character of each specific area. It also contributes to statistical improvement in most valuation cases, with a lower error than that obtained without a selected feature dataset, and it placed in the top 24% of the public leaderboard in the Zillow Prize competition.
- 14 June 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 86938-86953
Open AccessA Real Estate Valuation Model Using Boosted Feature Selection
To estimate real estate values, a complex valuation model based on artificial neural network (ANN) has been established as a successful means in modern machine learning research, specifically when high-dimensional data are available. Unfortunately, the real estate data in many locations, such as Thailand, are quite limited in terms of features. Hence, it becomes mandatory to reduce the complexity using feature selection techniques. These techniques aim to improve performance by identifying significant factors and help decrease the computational overload and model construction. However, due to the lack of explicability and interpretability in ANNs, the analysis of input factors cannot be explained directly by model composition. In this research, we apply a combination of a boosting strategy and input sensitivity analysis in an improved Garson's algorithm to perform feature selection that can adjust its selection criteria through each iteration on an ANN model. This proposed technique is then compared with other traditional feature selection techniques using synthetic data and real-world house valuation data. The results show that our model can maintain the sensitivity coefficient for every informative feature. The technique of this study provides a set of features that influences the house price and implies the character of each specific area. It also contributes to statistical improvement in most valuation cases, with a lower error than that obtained without a selected feature dataset, and it placed in the top 24% of the public leaderboard in the Zillow Prize competition.
- Open Access
- 4 May 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 68139-68150
Real Estate Recommendation Approach for Solving the Item Cold-Start Problem
The item cold-start problem occurs when a recommendation system cannot recommend new items owing to record deficiencies and new listing omissions. When searching for real estate, users can register a concurrent interest in recent and prior projects. Thus, an approach to recommend cold-start and warm-start items simultaneously must be determined. Furthermore, unrequired membership and stop-by behavior cause real estate recommendations to have many cold-start and new users. This characteristic encourages the use of a content-based approach and a session-based recommendation system. Herein, we propose a real estate recommendation approach for solving the item cold-start problem with acceptable warm-start item recommendations in the many-cold-start-users scenario. We modify a session-based recommendation system and employ existing mechanisms to efficiently deal with sequential and context information for the next-interacted item’s encoded attribute prediction. Subsequently, we use the nearest-neighbors approach using weighted cosine similarity to determine conforming candidates. We use Recall@K and MRR@K with the top-n recommendation to evaluate warm-start and cold-start item recommendations among different applied mechanisms and against the baselines. The results demonstrate the effectiveness of efficiently integrating the information and the difficulty in performing well in warm-start and cold-start item recommendations simultaneously. Our proposed approach illustrates the capability of solving the item cold-start problem while yielding promising results in both recommendations although neither result is the best. We believe that our approach provides a suitable compromise between both recommendations and that it will benefit recommendation tasks focusing on both recommendations.
- 4 May 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 68139-68150
Open AccessReal Estate Recommendation Approach for Solving the Item Cold-Start Problem
The item cold-start problem occurs when a recommendation system cannot recommend new items owing to record deficiencies and new listing omissions. When searching for real estate, users can register a concurrent interest in recent and prior projects. Thus, an approach to recommend cold-start and warm-start items simultaneously must be determined. Furthermore, unrequired membership and stop-by behavior cause real estate recommendations to have many cold-start and new users. This characteristic encourages the use of a content-based approach and a session-based recommendation system. Herein, we propose a real estate recommendation approach for solving the item cold-start problem with acceptable warm-start item recommendations in the many-cold-start-users scenario. We modify a session-based recommendation system and employ existing mechanisms to efficiently deal with sequential and context information for the next-interacted item’s encoded attribute prediction. Subsequently, we use the nearest-neighbors approach using weighted cosine similarity to determine conforming candidates. We use Recall@K and MRR@K with the top-n recommendation to evaluate warm-start and cold-start item recommendations among different applied mechanisms and against the baselines. The results demonstrate the effectiveness of efficiently integrating the information and the difficulty in performing well in warm-start and cold-start item recommendations simultaneously. Our proposed approach illustrates the capability of solving the item cold-start problem while yielding promising results in both recommendations although neither result is the best. We believe that our approach provides a suitable compromise between both recommendations and that it will benefit recommendation tasks focusing on both recommendations.
- 1 July 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Northern Thai Dialect Text to Speech
Each of the dialects of Thai Language has a distinct identity associated with its accents. The conversation between different native speakers of these dialects despite their standard language origination cannot be avoided when visiting each region. Communication with people who understand only the Northern Thai Dialect (NTD) brought us to the idea of inventing the Northern Thai Dialect Text to Speech (NTD-TTS). This idea derives from the same concept as a translating program; after getting text input in the Center Thai Dialect (CTD), the TTS system will translate and synthesize speech output in NTD. TTS used a software structure and modified two components: Grapheme to Phoneme (G2P) and Speech models. The NTD-G2P conversion was created by using rule-based and dictionary-based approaches. It was evaluated by 100 randomly selected sentences from ORCHID. The NTD-G2P reports a conversion accuracy of 83.19% on the syllable level and it is used for implementing the NTD-corpus. The sentence selections were presented to train the NTD speech model. The selection chosen covers 95.32% in the first percentile of phoneme distribution in the NTD-corpus. After connecting the speech models to the TTS system, the whole system was evaluated with Mean Opinion Score (MOS) and the comprehension on the syllable level by the native speakers. The NTD-MOS evaluations indicated that the accent, naturalness, and intelligibility of synthetic speech ranged from “acceptable” to “good”. The test set of the NTD-TTS system earned a good MOS and high comprehension percentage from the NTD native listeners. The results are 3.73 in the accent, 3.68 in the naturalness, 3.63 in the intelligibility, and the comprehension percentage is 97.16%.
- 1 July 2017
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-6
Northern Thai Dialect Text to Speech
Each of the dialects of Thai Language has a distinct identity associated with its accents. The conversation between different native speakers of these dialects despite their standard language origination cannot be avoided when visiting each region. Communication with people who understand only the Northern Thai Dialect (NTD) brought us to the idea of inventing the Northern Thai Dialect Text to Speech (NTD-TTS). This idea derives from the same concept as a translating program; after getting text input in the Center Thai Dialect (CTD), the TTS system will translate and synthesize speech output in NTD. TTS used a software structure and modified two components: Grapheme to Phoneme (G2P) and Speech models. The NTD-G2P conversion was created by using rule-based and dictionary-based approaches. It was evaluated by 100 randomly selected sentences from ORCHID. The NTD-G2P reports a conversion accuracy of 83.19% on the syllable level and it is used for implementing the NTD-corpus. The sentence selections were presented to train the NTD speech model. The selection chosen covers 95.32% in the first percentile of phoneme distribution in the NTD-corpus. After connecting the speech models to the TTS system, the whole system was evaluated with Mean Opinion Score (MOS) and the comprehension on the syllable level by the native speakers. The NTD-MOS evaluations indicated that the accent, naturalness, and intelligibility of synthetic speech ranged from “acceptable” to “good”. The test set of the NTD-TTS system earned a good MOS and high comprehension percentage from the NTD native listeners. The results are 3.73 in the accent, 3.68 in the naturalness, 3.63 in the intelligibility, and the comprehension percentage is 97.16%.
- 1 May 2016
- Published by EJournal Publishing in Lecture Notes on Software Engineering
- Vol. 4 (2) , 95-98
Accessible Video-Call Application on Android for the Blind
- 1 May 2016
- Published by EJournal Publishing in Lecture Notes on Software Engineering
- Vol. 4 (2) , 95-98
Accessible Video-Call Application on Android for the Blind
- Open Access
- 1 January 2014
- Published by Wiley in Mathematical Problems in Engineering
- Vol. 2014 (1)
Language Recognition Using Latent Dynamic Conditional Random Field Model with Phonological Features
Spoken language recognition (SLR) has been of increasing interest in multilingual speech recognition for identifying the languages of speech utterances. Most existing SLR approaches apply statistical modeling techniques with acoustic and phonotactic features. Among the popular approaches, the acoustic approach has become of greater interest than others because it does not require any prior language‐specific knowledge. Previous research on the acoustic approach has shown less interest in applying linguistic knowledge; it was only used as supplementary features, while the current state‐of‐the‐art system assumes independency among features. This paper proposes an SLR system based on the latent‐dynamic conditional random field (LDCRF) model using phonological features (PFs). We use PFs to represent acoustic characteristics and linguistic knowledge. The LDCRF model was employed to capture the dynamics of the PFs sequences for language classification. Baseline systems were conducted to evaluate the features and methods including Gaussian mixture model (GMM) based systems using PFs, GMM using cepstral features, and the CRF model using PFs. Evaluated on the NIST LRE 2007 corpus, the proposed method showed an improvement over the baseline systems. Additionally, it showed comparable result with the acoustic system based on i‐vector. This research demonstrates that utilizing PFs can enhance the performance.
- 1 January 2014
- Published by Wiley in Mathematical Problems in Engineering
- Vol. 2014 (1)
Open AccessLanguage Recognition Using Latent Dynamic Conditional Random Field Model with Phonological Features
Spoken language recognition (SLR) has been of increasing interest in multilingual speech recognition for identifying the languages of speech utterances. Most existing SLR approaches apply statistical modeling techniques with acoustic and phonotactic features. Among the popular approaches, the acoustic approach has become of greater interest than others because it does not require any prior language‐specific knowledge. Previous research on the acoustic approach has shown less interest in applying linguistic knowledge; it was only used as supplementary features, while the current state‐of‐the‐art system assumes independency among features. This paper proposes an SLR system based on the latent‐dynamic conditional random field (LDCRF) model using phonological features (PFs). We use PFs to represent acoustic characteristics and linguistic knowledge. The LDCRF model was employed to capture the dynamics of the PFs sequences for language classification. Baseline systems were conducted to evaluate the features and methods including Gaussian mixture model (GMM) based systems using PFs, GMM using cepstral features, and the CRF model using PFs. Evaluated on the NIST LRE 2007 corpus, the proposed method showed an improvement over the baseline systems. Additionally, it showed comparable result with the acoustic system based on i‐vector. This research demonstrates that utilizing PFs can enhance the performance.
- 1 October 2008
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 268-273
Improving Segment-based Speech Recognition by Recovering Missing Segments in Segment Graphs ¿ A Thai Case Study
In segment-based speech recognition systems, the quality of the segmentation step is a major factor highly affecting their accuracies. This paper proposes methods to reduce missing segments caused by boundary insertion errors in segment graphs, which, in the case of Thai, could be generated from a probabilistic segmentation with limited speech resources. Acoustic discontinuities and manners of articulation are used to verify boundaries of the segment graph. Segments are added to the graph in the case of possible falsely detected boundaries. With the proposed insertion error eliminations, the best phonetic recognition accuracy achieved shows a 13.66% error reduction.
- 1 October 2008
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 268-273
Improving Segment-based Speech Recognition by Recovering Missing Segments in Segment Graphs ¿ A Thai Case Study
In segment-based speech recognition systems, the quality of the segmentation step is a major factor highly affecting their accuracies. This paper proposes methods to reduce missing segments caused by boundary insertion errors in segment graphs, which, in the case of Thai, could be generated from a probabilistic segmentation with limited speech resources. Acoustic discontinuities and manners of articulation are used to verify boundaries of the segment graph. Segments are added to the graph in the case of possible falsely detected boundaries. With the proposed insertion error eliminations, the best phonetic recognition accuracy achieved shows a 13.66% error reduction.
- 1 October 2008
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 262-267
Concept Extraction from Conversational Speech using Segment Graphs
In this paper, a concept extraction method, an important element in spoken dialog systems when the nature of users' spoken queries are unstructured, is proposed and tested in a Thai automated international telephone call inquiry service. The method unifies concept constraints with other constraints, ranging from acoustic to language constraints, in a single segment-based probabilistic framework. With concept knowledge available, un-recalled keywords are reduced by 37%. Regarding concept extraction and goal identification, the proposed method is shown to be superior to the baseline system. Concepts are extracted at more than 80% of both recall and precision rates.
- 1 October 2008
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 262-267
Concept Extraction from Conversational Speech using Segment Graphs
In this paper, a concept extraction method, an important element in spoken dialog systems when the nature of users' spoken queries are unstructured, is proposed and tested in a Thai automated international telephone call inquiry service. The method unifies concept constraints with other constraints, ranging from acoustic to language constraints, in a single segment-based probabilistic framework. With concept knowledge available, un-recalled keywords are reduced by 37%. Regarding concept extraction and goal identification, the proposed method is shown to be superior to the baseline system. Concepts are extracted at more than 80% of both recall and precision rates.
- 1 March 2007
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 253-260
Thai Speech Keyword Spotting using Heterogeneous Acoustic Modeling
This paper illustrates the use of acoustic modeling of three different structures, including syllables, fillers and keywords, for keyword spotting. Filler models and syllable models are applied to capture out-of-vocabulary words, while keyword models extract significant words from speech utterances. Grammatical details are utilized with syllable models to add extra domain constraints. This improves the system's ability to detect non-keyword vocabularies. Filler models associating with syllable models reduce false alarm of keyword detection. Three kinds of filler models are described. Different types of filler models perform differently in keyword spotting of utterances with only one keyword and ones with multiple keywords. Experiments are conducted on a telephone call transferring via Thai spoken language domain. The proposed method is compared with a limited vocabulary speech recognition and keyword spotting using a reward function. For single- keyword utterances, the best accuracy obtained using the proposed method is approximately 70%, which is better than the ones from LVSR and spotting via reward functions. For multiple-keyword utterances, the best precision and recall rates are 72% and 65%, respectively. These are marginally better than ones obtained from limited vocabulary speech recognition, while typical reward function approach yields the rates of less than 50%.
- 1 March 2007
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 253-260
Thai Speech Keyword Spotting using Heterogeneous Acoustic Modeling
This paper illustrates the use of acoustic modeling of three different structures, including syllables, fillers and keywords, for keyword spotting. Filler models and syllable models are applied to capture out-of-vocabulary words, while keyword models extract significant words from speech utterances. Grammatical details are utilized with syllable models to add extra domain constraints. This improves the system's ability to detect non-keyword vocabularies. Filler models associating with syllable models reduce false alarm of keyword detection. Three kinds of filler models are described. Different types of filler models perform differently in keyword spotting of utterances with only one keyword and ones with multiple keywords. Experiments are conducted on a telephone call transferring via Thai spoken language domain. The proposed method is compared with a limited vocabulary speech recognition and keyword spotting using a reward function. For single- keyword utterances, the best accuracy obtained using the proposed method is approximately 70%, which is better than the ones from LVSR and spotting via reward functions. For multiple-keyword utterances, the best precision and recall rates are 72% and 65%, respectively. These are marginally better than ones obtained from limited vocabulary speech recognition, while typical reward function approach yields the rates of less than 50%.