You can find a searchable list of my publications below. My Google Scholar profile contains an up-to-date overview of my citations. I also have a ResearchGate profile with most of my full-texts.
I am a strong proponent of Open Access, especially after having spent more than four years as a researcher at an institution with a very limited number of journal subscriptions. For each entry below where I am legally allowed to share a full-text, you can find it as the first link of the entry.
2023
Viggen, Erlend Magnus; Singstad, Bjørn-Jostein; Time, Eirik; Mishra, Siddharth; Berg, Eirik
Assisted cement log interpretation using machine learning Journal Article
In: SPE Drilling & Completion, vol. 38, iss. 02, pp. 220–234, 2023.
Abstract | BibTeX | Tags: machine learning, well logging | Links:
@article{viggen_assisted_2022,
title = {Assisted cement log interpretation using machine learning},
author = {Erlend Magnus Viggen and Bjørn-Jostein Singstad and Eirik Time and Siddharth Mishra and Eirik Berg},
url = {https://hdl.handle.net/11250/3062647, Post-print at NTNU Open},
doi = {10.2118/209529-PA},
year = {2023},
date = {2023-06-14},
urldate = {2023-06-14},
journal = {SPE Drilling & Completion},
volume = {38},
issue = {02},
pages = {220–234},
abstract = {The Assisted Cement Log Interpretation Project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the code will be made available as open source.},
keywords = {machine learning, well logging},
pubstate = {published},
tppubtype = {article}
}
The Assisted Cement Log Interpretation Project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the code will be made available as open source.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals. By using high-quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with cased-hole logging interpretation software. By containerizing the code using an “API First” design principle (API: application programming interface), the applicability of this add-on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the code will be made available as open source.
2022
Time, Eirik; Viggen, Erlend Magnus; Mishra, Siddharth; Berg, Eirik
Assisted cement log interpretation Proceedings Article
In: SPE Norway Subsurface Conference 2022, pp. 15, Society of Petroleum Engineers, 2022.
Abstract | BibTeX | Tags: machine learning, well logging | Links:
@inproceedings{time_assisted_2022,
title = {Assisted cement log interpretation},
author = {Eirik Time and Erlend Magnus Viggen and Siddharth Mishra and Eirik Berg},
doi = {10.2118/209529-MS},
year = {2022},
date = {2022-04-27},
urldate = {2022-04-27},
booktitle = {SPE Norway Subsurface Conference 2022},
pages = {15},
publisher = {Society of Petroleum Engineers},
abstract = {The Assisted Cement Log Interpretation project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals in the cased hole logging unit. By using high quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with the cased hole logging interpretation software. By containerizing the code using an "API First" design principle, the applicability of this add- on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the results are made available as open source.},
keywords = {machine learning, well logging},
pubstate = {published},
tppubtype = {inproceedings}
}
The Assisted Cement Log Interpretation project has used machine learning (ML) to create a tool that interprets cement logs by predicting a predefined set of annular condition codes used in the cement log interpretation process.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals in the cased hole logging unit. By using high quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with the cased hole logging interpretation software. By containerizing the code using an "API First" design principle, the applicability of this add- on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the results are made available as open source.
The development of a cement log interpretation tool speeds up the log interpretation process and enables expert knowledge to be efficiently shared when training new professionals in the cased hole logging unit. By using high quality and consistent training data sets, the project has trained a model that will support unbiased and consistent interpretations over time.
The tool consists of a training and a prediction tool integrated with the cased hole logging interpretation software. By containerizing the code using an "API First" design principle, the applicability of this add- on tool is broad. The ML model is trained using selected and engineered features from cement logs, and the tool predicts an annular condition code according to the cement classification system for each depth segment in the log. The interpreters can easily fetch a complete cement log interpretation prediction for the log and use that as a template for their final interpretation. The ML model can easily be retrained with new data sets to improve accuracy even further.
To improve cement log interpretation consistency in the industry, the results are made available as open source.
2021
Viggen, Erlend Magnus; Løvstakken, Lasse; Måsøy, Svein-Erik; Merciu, Ioan Alexandru
Better automatic interpretation of cement evaluation logs through feature engineering Journal Article
In: SPE Journal, vol. 26, no. 05, pp. 2894–2913, 2021, ISSN: 1930-0220.
Abstract | BibTeX | Tags: machine learning, well logging | Links:
@article{viggen_better_2021b,
title = {Better automatic interpretation of cement evaluation logs through feature engineering},
author = {Erlend Magnus Viggen and Lasse Løvstakken and Svein-Erik Måsøy and Ioan Alexandru Merciu},
doi = {10.2118/204057-PA},
issn = {1930-0220},
year = {2021},
date = {2021-10-13},
urldate = {2021-10-13},
journal = {SPE Journal},
volume = {26},
number = {05},
pages = {2894–2913},
abstract = {We investigate systems to automatically interpret cement evaluation logs using supervised machine learning (ML). Such systems can provide instant rough interpretations that may then be used as a basis for human interpretation. Here, we compare the performance of two approaches, one previously published and one new. The previous approach is based on deep convolutional neural networks (CNNs) that autonomously learn to extract features from well log data, whereas the new approach uses feature engineering, in which we use our own domain knowledge to extract features.
We base this work on a data set of approximately 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (BQ; six ordinal classes) and hydraulic isolation (HI; two classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m in length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For BQ, the CNNs' interpretation exactly matches the reference 51.6% of the time. It does not miss by more than one class 88.5% of the time. For HI, the CNNs match the reference 86.7% of the time. The best-performingfeature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4, 89.5, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this data set may be particularly liable to overfitting. This may favor approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalizable. It may also favor simpler classifiers with less overfitting capacity.
This paper shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are easier to apply and deploy while also performing better than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.},
keywords = {machine learning, well logging},
pubstate = {published},
tppubtype = {article}
}
We investigate systems to automatically interpret cement evaluation logs using supervised machine learning (ML). Such systems can provide instant rough interpretations that may then be used as a basis for human interpretation. Here, we compare the performance of two approaches, one previously published and one new. The previous approach is based on deep convolutional neural networks (CNNs) that autonomously learn to extract features from well log data, whereas the new approach uses feature engineering, in which we use our own domain knowledge to extract features.
We base this work on a data set of approximately 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (BQ; six ordinal classes) and hydraulic isolation (HI; two classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m in length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For BQ, the CNNs' interpretation exactly matches the reference 51.6% of the time. It does not miss by more than one class 88.5% of the time. For HI, the CNNs match the reference 86.7% of the time. The best-performingfeature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4, 89.5, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this data set may be particularly liable to overfitting. This may favor approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalizable. It may also favor simpler classifiers with less overfitting capacity.
This paper shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are easier to apply and deploy while also performing better than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.
We base this work on a data set of approximately 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (BQ; six ordinal classes) and hydraulic isolation (HI; two classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m in length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For BQ, the CNNs' interpretation exactly matches the reference 51.6% of the time. It does not miss by more than one class 88.5% of the time. For HI, the CNNs match the reference 86.7% of the time. The best-performingfeature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4, 89.5, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this data set may be particularly liable to overfitting. This may favor approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalizable. It may also favor simpler classifiers with less overfitting capacity.
This paper shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are easier to apply and deploy while also performing better than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.
Viggen, Erlend Magnus; Løvstakken, Lasse; Merciu, Ioan Alexandru; Måsøy, Svein-Erik
Better automatic interpretation of cement evaluation logs through feature engineering Proceedings Article
In: SPE/IADC International Drilling Conference and Exhibition, pp. 28, 2021.
Abstract | BibTeX | Tags: machine learning, well logging | Links:
@inproceedings{viggen_better_2021a,
title = {Better automatic interpretation of cement evaluation logs through feature engineering},
author = {Erlend Magnus Viggen and Lasse Løvstakken and Ioan Alexandru Merciu and Svein-Erik Måsøy},
doi = {10.2118/204057-MS},
year = {2021},
date = {2021-03-09},
urldate = {2021-03-09},
booktitle = {SPE/IADC International Drilling Conference and Exhibition},
pages = {28},
abstract = {We build systems to automatically interpret cement evaluation logs using supervised machine learning (ML). Such systems can provide instant rough interpretations that may then be used as a basis for human interpretation. Here, we compare the performance of two approaches: A previously published approach based on deep convolutional neural networks (CNNs) that autonomously learn to extract features from well log data, and a feature-engineering approach where we use our own domain knowledge to extract features.
We base this work on a dataset of around 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (6 ordinal classes) and hydraulic isolation (2 classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For bond quality, the CNNs’ interpretation exactly matches the reference 51.6% of the time. 88.5% of the time, it does not miss by more than one class. For hydraulic isolation, the CNNs match the reference 86.7% of the time. The best-performing feature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4%, 89.5%, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this dataset may be particularly liable to overfitting. This may favour approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalisable. It may also favour simpler classifiers with less overfitting capacity.
This article shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are relatively easy to apply and deploy, with better results than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.},
keywords = {machine learning, well logging},
pubstate = {published},
tppubtype = {inproceedings}
}
We build systems to automatically interpret cement evaluation logs using supervised machine learning (ML). Such systems can provide instant rough interpretations that may then be used as a basis for human interpretation. Here, we compare the performance of two approaches: A previously published approach based on deep convolutional neural networks (CNNs) that autonomously learn to extract features from well log data, and a feature-engineering approach where we use our own domain knowledge to extract features.
We base this work on a dataset of around 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (6 ordinal classes) and hydraulic isolation (2 classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For bond quality, the CNNs’ interpretation exactly matches the reference 51.6% of the time. 88.5% of the time, it does not miss by more than one class. For hydraulic isolation, the CNNs match the reference 86.7% of the time. The best-performing feature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4%, 89.5%, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this dataset may be particularly liable to overfitting. This may favour approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalisable. It may also favour simpler classifiers with less overfitting capacity.
This article shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are relatively easy to apply and deploy, with better results than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.
We base this work on a dataset of around 60 km of well log data. Specialist interpreters have classified these logs according to the bond quality (6 ordinal classes) and hydraulic isolation (2 classes) of solids outside the casing. We train the ML systems to reproduce these reference interpretations in segments of 1 m length. The CNNs directly receive log data as a collection of 2D images and 1D curves. In the feature-engineering approach, we combine the extracted features with various classifiers.
For bond quality, the CNNs’ interpretation exactly matches the reference 51.6% of the time. 88.5% of the time, it does not miss by more than one class. For hydraulic isolation, the CNNs match the reference 86.7% of the time. The best-performing feature-based classifier, which is an ensemble of individual classifiers, provides better results of 57.4%, 89.5%, and 88.9%, respectively.
Our results indicate two main reasons why feature-based classifiers may perform particularly well on this task. First, there is some subjectivity inherent in the well log interpretations that are used to train and test ML systems. Second, well logs comprise many different and complex pieces of data. For these reasons, this dataset may be particularly liable to overfitting. This may favour approaches based on feature engineering, where we apply our domain knowledge to extract a few pieces of essential information from the data instead of leaving the job of understanding the data to an ML system that may misinterpret spurious patterns as generalisable. It may also favour simpler classifiers with less overfitting capacity.
This article shows how petroleum researchers and engineers can implement automatic interpretation systems for cement evaluation logs using ML methods that are relatively easy to apply and deploy, with better results than an approach based on autonomous feature extraction. This approach could also be adapted for automatic interpretation of other types of well log data.
2020
Viggen, Erlend Magnus; Merciu, Ioan Alexandru; Løvstakken, Lasse; Måsøy, Svein-Erik
Automatic interpretation of cement evaluation logs from cased boreholes using supervised deep neural networks Journal Article
In: Journal of Petroleum Science and Engineering, vol. 195, pp. 17, 2020, ISSN: 0920-4105.
Abstract | BibTeX | Tags: machine learning, well logging | Links:
@article{viggen_automatic_2020,
title = {Automatic interpretation of cement evaluation logs from cased boreholes using supervised deep neural networks},
author = {Erlend Magnus Viggen and Ioan Alexandru Merciu and Lasse Løvstakken and Svein-Erik Måsøy},
url = {https://www.sciencedirect.com/science/article/pii/S0920410520306100},
doi = {10.1016/j.petrol.2020.107539},
issn = {0920-4105},
year = {2020},
date = {2020-12-01},
urldate = {2020-12-01},
journal = {Journal of Petroleum Science and Engineering},
volume = {195},
pages = {17},
abstract = {The integrity of cement in cased boreholes is typically evaluated using well logging. However, well logging results are complex and can be ambiguous, and decisions associated with significant risks may be taken based on their interpretation. Cement evaluation logs must therefore be interpreted by trained professionals. To aid these interpreters, we propose a system for automatically interpreting cement evaluation logs, which they can use as a basis for their own interpretation. This system is based on deep convolutional neural networks, which we train in a supervised manner using a dataset of around 60 km of interpreted well log data. Thus, the networks learn the connections between data and interpretations during training. More specifically, the task of the networks is to classify the bond quality (among 6 ordinal classes) and the hydraulic isolation (2 classes) in each 1 m depth segment of each well based on the surrounding 13 m of well log data. We quantify the networks' performance by comparing over all segments how well the networks' interpretations of unseen data match the reference interpretations. For bond quality, the networks’ interpretation exactly matches the reference 51.6% of the time and is off by no more than one class 88.5% of the time. For hydraulic isolation, the interpretations match the reference 86.7% of the time. For comparison, a random-guess baseline gives matches of 16.7%, 44.4%, and 50%, respectively. We also compare with how well human reinterpretations of the log data match the reference interpretations, finding that the networks match the reference somewhat better. This may be linked to the networks learning and sharing the biases of the team behind the reference interpretations. An analysis of the results indicates that the subjectivity inherent in the interpretation process (and thereby in the reference interpretations we used for training and testing) is the main reason why we were not able to achieve an even better match between the networks and the reference.},
keywords = {machine learning, well logging},
pubstate = {published},
tppubtype = {article}
}
The integrity of cement in cased boreholes is typically evaluated using well logging. However, well logging results are complex and can be ambiguous, and decisions associated with significant risks may be taken based on their interpretation. Cement evaluation logs must therefore be interpreted by trained professionals. To aid these interpreters, we propose a system for automatically interpreting cement evaluation logs, which they can use as a basis for their own interpretation. This system is based on deep convolutional neural networks, which we train in a supervised manner using a dataset of around 60 km of interpreted well log data. Thus, the networks learn the connections between data and interpretations during training. More specifically, the task of the networks is to classify the bond quality (among 6 ordinal classes) and the hydraulic isolation (2 classes) in each 1 m depth segment of each well based on the surrounding 13 m of well log data. We quantify the networks' performance by comparing over all segments how well the networks' interpretations of unseen data match the reference interpretations. For bond quality, the networks’ interpretation exactly matches the reference 51.6% of the time and is off by no more than one class 88.5% of the time. For hydraulic isolation, the interpretations match the reference 86.7% of the time. For comparison, a random-guess baseline gives matches of 16.7%, 44.4%, and 50%, respectively. We also compare with how well human reinterpretations of the log data match the reference interpretations, finding that the networks match the reference somewhat better. This may be linked to the networks learning and sharing the biases of the team behind the reference interpretations. An analysis of the results indicates that the subjectivity inherent in the interpretation process (and thereby in the reference interpretations we used for training and testing) is the main reason why we were not able to achieve an even better match between the networks and the reference.
2018
Gelderblom, Femke B.; Tronstad, Tron V.; Viggen, Erlend Magnus
Subjective evaluation of a noise-reduced training target for deep neural network-based speech enhancement Journal Article
In: IEEE/ACM Transactions on Audio, Speech, and Language Processing, vol. 27, no. 3, pp. 583–594, 2018, ISSN: 2329-9290.
Abstract | BibTeX | Tags: machine learning, speech enhancement | Links:
@article{gelderblom_subjective_2018,
title = {Subjective evaluation of a noise-reduced training target for deep neural network-based speech enhancement},
author = {Femke B. Gelderblom and Tron V. Tronstad and Erlend Magnus Viggen},
url = {https://erlend-viggen.no/wp-content/uploads/2018/11/gelderblom_subjective_2018_post-print.pdf, Full-text},
doi = {10.1109/TASLP.2018.2882738},
issn = {2329-9290},
year = {2018},
date = {2018-11-21},
journal = {IEEE/ACM Transactions on Audio, Speech, and Language Processing},
volume = {27},
number = {3},
pages = {583–594},
abstract = {Speech enhancement systems aim to improve the quality and intelligibility of noisy speech. In this study, we compare two speech enhancement systems based on deep neural networks. The speech intelligibility and quality of both systems was evaluated subjectively, by a Speech Recognition Test based on Hagerman sentences and a translation of the ITU-T P.835 recommendation, respectively. Results were compared with the objective measures STOI and POLQA. Neither STOI nor POLQA reliably predicted subjective results. While STOI anticipated improvement, subjective results for both models showed degradation of speech intelligibility. POLQA results were overall hardly affected, while the subjective results showed significant changes in overall quality, both positive and negative, in many of the tests. One of the systems was trained to remove all noise; a strategy that is common in speech enhancement systems found in the literature. The other system was trained to only reduce the noise such that the signal-to-noise ratio increased with 10 dB. The latter system subjectively outperformed the system that attempted to remove noise completely. From this, we conclude that objective evaluation cannot replace subjective evaluation until a measure that reliably predicts intelligibility and quality for deep neural network based systems has been identified. Results further indicate that it may be beneficial to move away from more aggressive noise removal strategies towards noise reduction strategies that cause less speech distortion.},
keywords = {machine learning, speech enhancement},
pubstate = {published},
tppubtype = {article}
}
Speech enhancement systems aim to improve the quality and intelligibility of noisy speech. In this study, we compare two speech enhancement systems based on deep neural networks. The speech intelligibility and quality of both systems was evaluated subjectively, by a Speech Recognition Test based on Hagerman sentences and a translation of the ITU-T P.835 recommendation, respectively. Results were compared with the objective measures STOI and POLQA. Neither STOI nor POLQA reliably predicted subjective results. While STOI anticipated improvement, subjective results for both models showed degradation of speech intelligibility. POLQA results were overall hardly affected, while the subjective results showed significant changes in overall quality, both positive and negative, in many of the tests. One of the systems was trained to remove all noise; a strategy that is common in speech enhancement systems found in the literature. The other system was trained to only reduce the noise such that the signal-to-noise ratio increased with 10 dB. The latter system subjectively outperformed the system that attempted to remove noise completely. From this, we conclude that objective evaluation cannot replace subjective evaluation until a measure that reliably predicts intelligibility and quality for deep neural network based systems has been identified. Results further indicate that it may be beneficial to move away from more aggressive noise removal strategies towards noise reduction strategies that cause less speech distortion.
2017
Gelderblom, Femke B.; Tronstad, Tron V.; Viggen, Erlend Magnus
Subjective Intelligibility of Deep Neural Network-Based Speech Enhancement Presentation
22.08.2017.
BibTeX | Tags: machine learning, speech enhancement | Links:
@misc{gelderblom_subjective_2017b,
title = {Subjective Intelligibility of Deep Neural Network-Based Speech Enhancement},
author = {Femke B. Gelderblom and Tron V. Tronstad and Erlend Magnus Viggen},
url = {https://www.researchgate.net/publication/319243091_Poster_Subjective_intelligibility_of_deep_neural_network-based_speech_enhancement, Poster on ResearchGate},
year = {2017},
date = {2017-08-22},
keywords = {machine learning, speech enhancement},
pubstate = {published},
tppubtype = {presentation}
}
Gelderblom, Femke B; Tronstad, Tron V; Viggen, Erlend Magnus
Subjective Intelligibility of Deep Neural Network-Based Speech Enhancement Proceedings Article
In: INTERSPEECH 2017, pp. 1968–1972, ISCA, 2017.
Abstract | BibTeX | Tags: machine learning, speech enhancement | Links:
@inproceedings{gelderblom_subjective_2017,
title = {Subjective Intelligibility of Deep Neural Network-Based Speech Enhancement},
author = {Femke B Gelderblom and Tron V Tronstad and Erlend Magnus Viggen},
url = {https://www.researchgate.net/publication/319184981_Subjective_Intelligibility_of_Deep_Neural_Network-Based_Speech_Enhancement, Full-text on ResearchGate},
doi = {10.21437/Interspeech.2017-1041},
year = {2017},
date = {2017-01-01},
urldate = {2017-01-01},
booktitle = {INTERSPEECH 2017},
pages = {1968--1972},
publisher = {ISCA},
abstract = {Recent literature indicates increasing interest in deep neural networks for use in speech enhancement systems. Currently, these systems are mostly evaluated through objective measures of speech quality and/or intelligibility. Subjective intelligibility evaluations of these systems have so far not been reported. In this paper we report the results of a speech recognition test with 15 participants, where the participants were asked to pick out words in background noise before and after enhancement using a common deep neural network approach. We found that, although the objective measure STOI predicts that intelligibility should improve or at the very least stay the same, the speech recognition threshold, which is a measure of intelligibility, deteriorated by 4 dB. These results indicate that STOI is not a good predictor for the subjective intelligibility of deep neural network-based speech enhancement systems. We also found that the postprocessing technique of global variance normalisation does not significantly affect subjective intelligibility.},
keywords = {machine learning, speech enhancement},
pubstate = {published},
tppubtype = {inproceedings}
}
Recent literature indicates increasing interest in deep neural networks for use in speech enhancement systems. Currently, these systems are mostly evaluated through objective measures of speech quality and/or intelligibility. Subjective intelligibility evaluations of these systems have so far not been reported. In this paper we report the results of a speech recognition test with 15 participants, where the participants were asked to pick out words in background noise before and after enhancement using a common deep neural network approach. We found that, although the objective measure STOI predicts that intelligibility should improve or at the very least stay the same, the speech recognition threshold, which is a measure of intelligibility, deteriorated by 4 dB. These results indicate that STOI is not a good predictor for the subjective intelligibility of deep neural network-based speech enhancement systems. We also found that the postprocessing technique of global variance normalisation does not significantly affect subjective intelligibility.
This page is powered by the excellent TeachPress plugin for WordPress.