Content uploaded by Dinesh Attota
Author content
All content in this area was uploaded by Dinesh Attota on Feb 15, 2023
Content may be subject to copyright.
Towards Application of Speech Analysis in
Predicting Learners’ Performance
Dinesh Chowdary Attota
Department of Computer Science
Kennesaw State University
Marietta, GA, USA
dattota@students.kennesaw.edu
Nasrin Dehbozorgi
Department of Software Engineering
Kennesaw State University
Marietta, GA, USA
dnasrin@kennesaw.edu
Abstract—In this work in progress, we propose a model for
analysis of students’ verbal conversation during teamwork to
predict their academic performance based on expressed emotions.
Our previous studies support the link between an individual’s
attitude and emotional states during the cognitive process with
their performance in the given context [1], [2]. Traditionally
the learners’ affective states were assessed by having them fill
out standard surveys. More recently the researchers have been
using advanced methods to extract students’ emotions from
their writings by using Natural Language Processing (NLP)
models. These models are applied to data collected from different
sources such as discussion forums, team chats, students’ reflective
surveys, and journals. In this research, we take one step further
by recording students audio in class as they converse about the
course topic in low-stake teams and extract emotions from their
conversations by NLP methods. The main contributions of the
proposed model are 1) the audio transcription component 2)
the multi-class emotion analysis unit and 3) the performance
prediction model based on input data. SpeechBrain pre-trained
models with transformer language models were applied for
automated transcription of audio data and converting them to
embedding vectors. NLP methods were applied for sentiment
analysis. Next, we formed the feature set by combining the
extracted emotions with students’ formative assessment grades
during the semester to implement a prediction model. We further
analyzed which features in the feature set have a higher impact
on the students’ academic performance. The early result of this
research is promising as we found high accuracy in the predicted
scores of the students.
Index Terms—Automatic Speech Recognition (ASR), emotion
analysis, predictive model, academic performance, NLP
I. INT ROD UC TI ON
The application of advanced technologies for a more re-
alistic interpretation of human speech by computers is being
more popular in both academia and industrial domains. Au-
tomated Speech Recognition (ASR) is a critical component
of conversational technology by which computers detect and
convert spoken language into text and bring together linguis-
tics, computer science, and other disciplines of study. Due to
effective training and decoding techniques, end-to-end ASR
has garnered interest as a means of directly combining acoustic
and linguistic models (AMs and LMs) [3], [4]. Numerous
models for ASR have been developed, including attention-
based encoder-decoder architectures [5], [6], Recurrent Neural
Network (RNN)-powered transducers [7], and Connectionist
Temporal Classification (CTC). Transformers have taken the
role of RNNs in recent years, exceeding bi-directional RNNs
in terms of performance [8], [9]. Transfer learning is one
of the most remarkable models and algorithms in the wide
family of machine learning methods and algorithms. Transfer
learning is a broad term that encompasses all strategies that
use supplemental resources to enhance model learning for the
target problem domain. With a high level of variation and
dynamics, it is practically impossible for researchers in speech
and language to train the model from a single data source [10].
We can depend on more clever algorithms that allow learning
from a wide range of languages, data sets, and topics, and to
constantly adapt the model.
A. Emotion Recognition:
The study of people’s feelings or emotions toward a sub-
ject is called emotion recognition. It is critical to evaluate
students’ emotions in the educational context and notably
during the collaborative learning process, in order to adapt and
enhance content delivery methods [11]. Holistic evaluation of
students’ academic performance is critical for both students
and educators, enabling them to identify the ones at risk and
intervene early in order to lessen the likelihood of failure
[12], [13]. When attempting to gauge a student’s academic
progress, formative and summative tests and assignments are
often used [14]. Such evaluations provide useful information,
such as trends and patterns connected to the educational
process, which may be utilized to better understand the overall
learning state of the students. Grade-based assessment and
evaluation have some pitfalls too, especially in collaborative
environments in which it’s a challenge to assess an individual’s
contribution to the teamwork. Research suggests there is a
correlation between students’ emotions and their academic
achievement [15]. Sentimental data such as anger, fear, joy,
surprise, and sadness can be used as complementary data
points to students’ grades to make the evaluation process more
efficient [16]. In the following sections, we discuss the related
work and present the proposed model followed by a case study
and data analysis.
II. RE LATE D WOR K
Using deep neural networks to combine different types
of information, such as audio and language, makes it eas-
2022 IEEE Frontiers in Education Conference (FIE) | 978-1-6654-6244-0/22/$31.00 ©2022 IEEE | DOI: 10.1109/FIE56618.2022.9962701
Authorized licensed use limited to: Kennesaw State University. Downloaded on January 23,2023 at 16:19:51 UTC from IEEE Xplore. Restrictions apply.
ier for computers to recognize emotions from speech [17].
Researchers in [18] incorporate an end-to-end framework for
sentiment analysis using speech recognition. In this work,
they utilized a single layer multiplicative LSTM (mLSTM)
[19] model with 4,096 nodes to encode the entire text input.
The ASR model computes character probabilities for each
frame and then extracts the final transcription (greedy de-
coding). They have not used any language model to rectify
any spelling errors or out-of-vocabulary words. This ASR
model is trained on five datasets: LibriSpeech, TED-LIUM v3,
Mozilla Common Voice, and VoxForge. Each dataset contains
around 1000 hours of English-read speech. The only pre-
processing technique used is the conversion of recording to a
WAV file using a single-channel 16-bit signed integer format
with a sampling rate of 16,000. They have achieved 69.9% of
accuracy.
Empathy-based interactive dialogue management is another
approach presented by the authors in [20]. Emotions were
extracted from raw speech input using the CNN deep learning
model. In this approach, Deep Learning models like DNN-
HMMs have been used to train raw audio for the acoustic
model of the Kaldi speech recognition system. The sentiment
of identified speech was examined using a CNN-based classi-
fier and Word2Vec in experiments done with the TED-LIUM
dataset. An accuracy of 65.7 was achieved by avoiding feature
engineering approaches in voice emotion identification. When
trained on domain data, CNN sentiment analysis yielded an
82.5 F-score.
On the other hand, research in [18] developed a strategy
for integrating an ASR system with a character-level recurrent
neural network for sentiment recognition. Earlier neural emo-
tion detection models in the context of human-robot interaction
[21] have been extended in their work. Experiments have been
carried out to resolve the disparities in the performance of
spoken sentiment identification when there is no transcript.
They employed mLSTM (multiplicative Long Short Term
Memory) to represent the entire textual input. This study
combined five different freely available datasets to train the
model. On the Stanford Sentimental Treebank, the model
outperformed more sophisticated architectures in identifying
emotion based on the next character prediction in the given
context.
Use of Transformers based language models has been
increased in recent years [22], [23], [24], [25]. Originally
suggested for machine translation, the transformer model is
an encoder-decoder based on self-attention. When applied to
ASR, the transformer model has shown promising results and
enhanced Word Error Rates(WERs) over RNN-based systems
[26]. According to LibriSpeech test data, the suggested stream-
ing transformer architecture in [18] produces a WER of 2.8%
and 7.3%, which is the best documented streaming end-to-
end ASR result for this job to the researchers’ knowledge.
Likewise, research work in [27] presented an approach for
recognizing emotions in speech using transfer learning from
automatic speech recognition. They attained a 71.7 percent ac-
curacy for anger, excitement, sadness, and neutrality emotion
content using speech data.
Automatic sentiment recognition in natural audio streams
was the subject of a study published in [28] by a team of
researchers. Text sentiment detection models were developed
using POS tagging and Maximum Entropy Modelling (MEM).
An algorithm called ME is used to estimate ratings based
on text collected from reviews. It makes use of Stanford’s
Log-linear POS tagger for POS tagging. Switchboard and
Fisher corpora have been used to train the speech recognition
system, while Mel Frequency Cepstral Coefficients (MFCC)
characteristics have been employed to train the acoustic model
in this study. The baseline model without tuning had an
accuracy of 92.1%, whereas the baseline mode, with certain
parts of speech such as Nouns, had an accuracy of 94.8%.
III. PROP OS ED ME TH OD
In this section, we discuss our proposed model for automatic
speech transcription and extracting different types of emotions
from transcribed speech. Fig 1depicts the high-level architec-
ture of the model, which is composed of a speech transcription
unit, an emotion recognition unit, and a regression model to
predict the students’ performance based on the conversations
captured from their audio input and score in-class activities.
A. Speech Transcription Unit
We used SpeechBrain [29], an open-source conversational
AI framework that runs on PyTorch, to create voice transcrip-
tion. SpeechBrain is a transformer-based end-to-end ASR that
includes a large number of language translation models that
create WAV to vector embeddings that can be used to predict
text. SpeechBrain enables a user-friendly and flexible imple-
mentation of cutting-edge speech technologies such as speech
recognition, speaker recognition, voice augmentation, speech
Fig. 1: High Level Architecture of the Proposed Approach
Authorized licensed use limited to: Kennesaw State University. Downloaded on January 23,2023 at 16:19:51 UTC from IEEE Xplore. Restrictions apply.
separation, language identification, and multi-microphone sig-
nal processing. This ASR has been pre-trained on the corpus
of LibriSpeech [30] and is publicly available on HuggingFace.
This ASR consists of three components of Unigram Tokenizer,
Neural Language Model (Transformer LM), and Acoustic
model with CTC.
The Unigram Tokenizer converts the words into subwords and
is trained using LibriSpeech train transcriptions [31]. The tok-
enization starts with a large vocabulary and gradually reduces
the size of the vocabulary until it reaches the target vocabulary
size. A Unigram model examines each token independent of
the previous one such that the probability of token X given
the preceding context is just X. So a Unigram language model
would always anticipate the most popular token. A token’s
likelihood is its frequency in the original corpus divided by
the total of all tokens’ frequencies in the lexicon (to make sure
the probabilities sum up to 1).
The Neural Language Model consists of a deep learning
model, which is trained on a dataset that consists of 10M
words. The neural language model predicts the most likely
sequence of words among numerous text strings. The output
of the previous tokenization component will be passed to
this language model to predict the probabilities of words at
different timestamps [32].
Finally, the acoustic model analyzes the waveform of speech
and predicts the most likely phonemes in the speech. The
language model generates a matrix containing the character
probabilities for each timestamp. The matrix that is generated
by the neural language is decoded with the Connectionist
Temporal Classification (CTC) algorithm [33]
B. Emotion Recognition
We applied Google’s Text-To-Text-Transfer Transformer
(T5) basic fine-tuning model to extract different emotion
classes of joy, anger, fear, sadness, surprise, and love. T5 is
based on the research reported in [34], which conducts a large-
scale empirical survey to ascertain the most effective transfer
learning approaches. Colossal Clean Crawled Corpus (C4)
dataset [35], which is two orders of magnitude bigger than
Wikipedia, is used to train the T5 model. The trained model
is flexible enough to be used for a wide range of important
downstream tasks. T5 generates text-to-text formats, where
the inputs and outputs are both text strings, unlike BERT,
which only outputs a class label or a span of the input. Using
graph representations of the text, the pre-trained model of T5
maintains semantic relations between words in a phrase and
extracts patterns. CNN uses the extracted patterns to infer the
underlying emotion of the phrase they are fed into to get their
predictions. The self-attention feature analyzes a sentence and
places it into one of many emotion categories based on the
keywords that are found in it [36]. These keywords correspond
to the level of interest that a student has in a certain topic
or assignment. Thus, the pre-trained T5 model enabled us
to extract six distinct types of emotions from the student’s
conversation. In this study, the emotion of ”love” is intertwined
with passion and interest.
C. Linear Regression Prediction Model
We used multiple linear regression (MLR) to predict the
performance of the students based on the extracted emotions
combined with tests and assignment grades. Multiple linear
regression figures out the relationship between two or more
independent variables and a dependent variable by fitting a
linear equation to the data. Each independent variable x has a
corresponding value in the dependent variable y. The equation
1represents the regression line for n observations.
yi=β0+β1xi1+β2xi2+β3xi3+... +βnxin (1)
The least-squares model finds the line that best fits the
observed data by minimizing the sum of the squares of
the vertical deviations between each data point and the line
(vertical deviation is 0 when the point lies on the fitted line).
The variance can be calculated using Mean Squared Error
(MSE) shown in equation 2.
v2=Pe2
i
n−p−1(2)
D. Case study
To test the developed model we conducted a case study
by collecting speech data from a CS1 active learning class
[37], [38] where students worked in the same low-stakes teams
throughout the semester [39]. The class met twice a week
for 75 minutes. An average of 40 minutes was dedicated
to teamwork in each class that we used to record students’
speeches. For this purpose, a recorder with dual microphones
was connected to the members of the team and the total
number of 28 students were recorded during 5 weeks of
the semester. Every class began with a mini-quiz on the
prep material [40]. After a brief poll quiz on prep-work
students were given a mini-lecture if they didn’t comprehend
the subject. Then, after completing a graded class activity,
students were requested to complete an exit form (minute
paper) separately [41]. Students were tested in 4 tests, 4 major
assignments, and 4 lab tests. The students’ final scores were
based on their performance in the final exam as well as class
and lab tests, assignments, class activities, polls, and prep
work quizzes. This study uses students’ final grades as the
performance metric to determine if emotions and other low-
stake grade data points can predict their performance.
For extracting emotions from students’ speech we used
audio files, each containing the recordings of conversations
between two students in a team. To accommodate the em-
bedding vectors of the transcription model in GPU memory,
we segmented the original audio file into multiple parts with
a duration of 30 seconds for each part. The audio segments
were given as input to the SpeechBrain thereby extracting
embedding vectors and passed to decoders to extract the
relevant text. The extracted transcribed text of an audio chunk
is appended with the transcription of previous audio chunks in
each dataset. The transcriptions were passed to the Google T5
fine-tuned pre-trained model for emotion extraction. A sample
Authorized licensed use limited to: Kennesaw State University. Downloaded on January 23,2023 at 16:19:51 UTC from IEEE Xplore. Restrictions apply.
Fig. 2: Emotion Trends of a Sample Team
visualization of the trends of the emotion of a single team over
the semester is presented in figure 2. In this plot, each emotion
is color coded and the size of the circles reflects the intensity
of the emotion in the team conversation at the specified time.
IV. DATA ANALYSI S
We evaluated if the proposed model can predict stu-
dents’ final grades in the course based on different data
points collected throughout the semester. These data points
included varying emotion classes, personality traits of the
students and their peers in a team (measured by Big
Five Personality tool [1], lab tests, lecture tests, assign-
ment tests, prep-work quiz grades, and class activity scores.
For the prediction model, we used the scikit-learn [42]
library to implement multiple linear regression on differ-
ent feature sets (combinations of data points). Our analysis
showed that students’ personality has the least weight in
predicting the performance and the feature set of [anger,
fear, joy, surprise, sadness, love, total number of words,
class activity grade, preparation grade, assignment grade,
lecture test grade, lab test grade] predicted the final score
more accurately. Furthermore, we implemented the variable
clustering technique [43] to find out which variables in the
feature set have more predictive power. Variables were divided
into multiple sets of disjoint clusters. Each cluster has a linear
combination of its corresponding variables. The clustering
method uses (1−R2) measure to find the most significant
variable. The (1 −R2)ratio is defined below in equation 3.
1−R2=(1 −R2)(own)
(1 −R2)(nearest)(3)
For a variable where 1−R2goes nearest to zero is the best
representation of a cluster to use. When we applied the vari-
able clustering technique to the feature set we observed that
the variables of [joy, lecture test grade, preparation grade,
love] have more predictive power in predicting the students’
academic performance.
The preliminary result shows that the students’ positive
emotions (joy and love) and their preparedness before attend-
ing the class as well as their score in the main lecture tests
can determine if they will earn high final grades or not. The
early result of applying our model to students’ data using the
selected feature set, shows impressively good performance and
high accuracy. The difference between the predicted scores and
the actual scores of the students ranges anywhere between -1
to 1. One reason for this high accuracy could be the limited
data we had on a small sample size. More analysis should be
done to come up with generic conclusions.
V. CONCLUSION
In this paper, we proposed a model to predict students’
performance based on the emotions they express in their con-
versations as they worked in teams as well as their formative
assessment scores during the semester. We used SpeechBrain
to transcribe the recorded speech and by using a transformer-
based emotion recognizer (T5) extracted the emotion classes.
Data shows the performance of the model is promising as
the predicted values are very close to students’ actual grades.
However, this work is considered as proof of concept and
it’s early to provide solid conclusions. The focus of this
paper was to present the core functionality of the model
with a limited sample size. In future work, we will work
on a larger sample size of students in different classes and
will be extending the functionality of the model by creating
a dynamic dashboard that allows educators to visualize the
various emotional patterns of the students as they work in
class. This system allows instructors to adjust the content
delivery pace and methods according to both educational
and emotional feedback presented to them. This research has
the potential to help instructors get better insights into the
students’ progress earlier in the semester and apply required
interventions accordingly.
REFERENCES
[1] N. Dehbozorgi, M. Lou Maher, and M. Dorodchi, “Sentiment analysis
on conversations in collaborative active learning as an early predictor of
performance,” in 2020 IEEE Frontiers in Education Conference (FIE),
2020, pp. 1–9.
[2] N. Dehbozorgi, “Sentiment analysis on verbal data from team discus-
sions as an indicator of individual performance,” Ph.D. dissertation, The
University of North Carolina at Charlotte, 2020.
[3] S. Karita, N. Yalta, S. Watanabe, M. Delcroix, A. Ogawa, and
T. Nakatani, “Improving transformer-based end-to-end speech recog-
nition with connectionist temporal classification and language model
integration,” 09 2019, pp. 1408–1412.
Authorized licensed use limited to: Kennesaw State University. Downloaded on January 23,2023 at 16:19:51 UTC from IEEE Xplore. Restrictions apply.
[4] Y. Wang, A. Mohamed, D. Le, C. Liu, A. Xiao, J. Mahadeokar,
H. Huang, A. Tjandra, X. Zhang, F. Zhang, C. Fuegen, G. Zweig, and
M. L. Seltzer, “Transformer-based acoustic modeling for hybrid speech
recognition,” in ICASSP 2020 - 2020 IEEE International Conference
on Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 6874–
6878.
[5] S. Zhang, E. Loweimi, P. Bell, and S. Renals, “On the usefulness of self-
attention for automatic speech recognition with transformers,” in 2021
IEEE Spoken Language Technology Workshop (SLT). IEEE, 2021, pp.
89–96.
[6] H. Le, J. Pino, C. Wang, J. Gu, D. Schwab, and L. Besacier, “Dual-
decoder transformer for joint automatic speech recognition and multi-
lingual speech translation,” arXiv preprint arXiv:2011.00747, 2020.
[7] J. Li, R. Zhao, H. Hu, and Y. Gong, “Improving rnn transducer modeling
for end-to-end speech recognition,” in 2019 IEEE Automatic Speech
Recognition and Understanding Workshop (ASRU), 2019, pp. 114–121.
[8] E. Tsunoo, Y. Kashiwagi, T. Kumakura, and S. Watanabe, “Towards
online end-to-end transformer automatic speech recognition,” arXiv
preprint arXiv:1910.11871, 2019.
[9] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in
neural information processing systems, vol. 30, 2017.
[10] J. Cho, M. K. Baskar, R. Li, M. Wiesner, S. H. Mallidi, N. Yalta,
M. Karafi´
at, S. Watanabe, and T. Hori, “Multilingual sequence-to-
sequence speech recognition: Architecture, transfer learning, and lan-
guage modeling,” in 2018 IEEE Spoken Language Technology Workshop
(SLT), 2018, pp. 521–527.
[11] N. Dehbozorgi, M. L. Maher, and M. Dorodchi, “Emotion mining from
speech in collaborative learning.”
[12] S. M. Jayaprakash, E. W. Moody, E. J. Laur´
ıa, J. R. Regan, and J. D.
Baron, “Early alert of academically at-risk students: An open source
analytics initiative,” Journal of Learning Analytics, vol. 1, no. 1, pp.
6–47, 2014.
[13] D. Wiliam*, C. Lee, C. Harrison, and P. Black, “Teachers developing
assessment for learning: Impact on student achievement,” Assessment in
education: principles, policy & practice, vol. 11, no. 1, pp. 49–65, 2004.
[14] S. M. Brookhart, “Successful students’ formative and summative uses
of assessment information,” Assessment in Education: Principles, Policy
& Practice, vol. 8, no. 2, pp. 153–169, 2001. [Online]. Available:
https://doi.org/10.1080/09695940123775
[15] R. Pekrun, T. Goetz, W. Titz, and R. P. Perry, “Academic emotions
in students’ self-regulated learning and achievement: A program of
qualitative and quantitative research,” Educational psychologist, vol. 37,
no. 2, pp. 91–105, 2002.
[16] N. Dehbozorgi and D. P. Mohandoss, “Aspect-based emotion analysis
on speech for predicting performance in collaborative learning,” in 2021
IEEE Frontiers in Education Conference (FIE), 2021, pp. 1–7.
[17] V. Rozgi´
c, S. Ananthakrishnan, S. Saleem, R. Kumar, and R. Prasad,
“Ensemble of svm trees for multimodal emotion recognition,” in Pro-
ceedings of The 2012 Asia Pacific Signal and Information Processing
Association Annual Summit and Conference, 2012, pp. 1–4.
[18] E. Lakomkin, M. A. Zamani, C. Weber, S. Magg, and S. Wermter,
“Incorporating end-to-end speech recognition models for sentiment
analysis,” in 2019 International Conference on Robotics and Automation
(ICRA), 2019, pp. 7976–7982.
[19] B. Krause, L. Lu, I. Murray, and S. Renals, “Multiplicative lstm for
sequence modelling,” arXiv preprint arXiv:1609.07959, 2016.
[20] D. Bertero, F. B. Siddique, C.-S. Wu, Y. Wan, R. H. Y. Chan, and
P. Fung, “Real-time speech emotion and sentiment recognition for
interactive dialogue systems,” in Proceedings of the 2016 conference
on empirical methods in natural language processing, 2016, pp. 1042–
1047.
[21] E. Lakomkin, M. A. Zamani, C. Weber, S. Magg, and S. Wermter,
“Emorl: continuous acoustic emotion classification using deep reinforce-
ment learning,” in 2018 IEEE International Conference on Robotics and
Automation (ICRA). IEEE, 2018, pp. 4445–4450.
[22] B. Xue, J. Yu, J. Xu, S. Liu, S. Hu, Z. Ye, M. Geng, X. Liu, and H. Meng,
“Bayesian transformer language models for speech recognition,” in
ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech
and Signal Processing (ICASSP). IEEE, 2021, pp. 7378–7382.
[23] Z. Dai, Z. Yang, Y. Yang, J. Carbonell, Q. V. Le, and R. Salakhutdi-
nov, “Transformer-xl: Attentive language models beyond a fixed-length
context,” arXiv preprint arXiv:1901.02860, 2019.
[24] K. Irie, A. Zeyer, R. Schl¨
uter, and H. Ney, “Language modeling with
deep transformers,” arXiv preprint arXiv:1905.04226, 2019.
[25] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert: Pre-training
of deep bidirectional transformers for language understanding,” arXiv
preprint arXiv:1810.04805, 2018.
[26] S. Karita, N. Chen, T. Hayashi, T. Hori, H. Inaguma, Z. Jiang,
M. Someki, N. E. Y. Soplin, R. Yamamoto, X. Wang et al., “A
comparative study on transformer vs rnn in speech applications,” in
2019 IEEE Automatic Speech Recognition and Understanding Workshop
(ASRU). IEEE, 2019, pp. 449–456.
[27] S. Zhou and H. Beigi, “A transfer learning method for speech emo-
tion recognition from automatic speech recognition,” arXiv preprint
arXiv:2008.02863, 2020.
[28] L. Kaushik, A. Sangwan, and J. H. L. Hansen, “Sentiment extraction
from natural audio streams,” in 2013 IEEE International Conference on
Acoustics, Speech and Signal Processing, 2013, pp. 8485–8489.
[29] M. Ravanelli, T. Parcollet, P. Plantinga, A. Rouhe, S. Cornell, L. Lu-
gosch, C. Subakan, N. Dawalatabad, A. Heba, J. Zhong, J.-C. Chou,
S.-L. Yeh, S.-W. Fu, C.-F. Liao, E. Rastorgueva, F. Grondin, W. Aris,
H. Na, Y. Gao, R. D. Mori, and Y. Bengio, “SpeechBrain: A general-
purpose speech toolkit,” 2021, arXiv:2106.04624.
[30] V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech:
An asr corpus based on public domain audio books,” in 2015 IEEE
International Conference on Acoustics, Speech and Signal Processing
(ICASSP), 2015, pp. 5206–5210.
[31] T. Kudo, “Subword regularization: Improving neural network trans-
lation models with multiple subword candidates,” arXiv preprint
arXiv:1804.10959, 2018.
[32] K. Li, Z. Liu, T. He, H. Huang, F. Peng, D. Povey, and S. Khudan-
pur, “An empirical study of transformer-based neural language model
adaptation,” in ICASSP 2020 - 2020 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP), 2020, pp. 7934–
7938.
[33] H. Scheidl, S. Fiel, and R. Sablatnig, “Word beam search: A connec-
tionist temporal classification decoding algorithm,” in 2018 16th Inter-
national Conference on Frontiers in Handwriting Recognition (ICFHR),
2018, pp. 253–258.
[34] C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena,
Y. Zhou, W. Li, and P. J. Liu, “Exploring the limits of trans-
fer learning with a unified text-to-text transformer,” arXiv preprint
arXiv:1910.10683, 2019.
[35] J. Dodge, M. Sap, A. Marasovic, W. Agnew, G. Ilharco, D. Groeneveld,
and M. Gardner, “Documenting the english colossal clean crawled
corpus,” arXiv e-prints, pp. arXiv–2104, 2021.
[36] E. Saravia, H.-C. T. Liu, Y.-H. Huang, J. Wu, and Y.-S. Chen, “Carer:
Contextualized affect representations for emotion recognition,” in Pro-
ceedings of the 2018 conference on empirical methods in natural
language processing, 2018, pp. 3687–3697.
[37] N. Dehbozorgi, S. MacNeil, M. L. Maher, and M. Dorodchi, “A
comparison of lecture-based and active learning design patterns in cs
education,” in 2018 IEEE Frontiers in Education Conference (FIE),
2018, pp. 1–8.
[38] N. Dehbozorgi, “Active learning design patterns for cs education,” in
Proceedings of the 2017 ACM Conference on International Computing
Education Research, 2017, pp. 291–292.
[39] N. Dehbozorgi, M. L. Maher, and M. Dorodchi, “Does self-efficacy cor-
relate with positive emotion and academic performance in collaborative
learning?” in 2021 IEEE Frontiers in Education Conference (FIE), 2021,
pp. 1–8.
[40] M. Maher, N. Dehbozorgi, M. Dorodchi, and S. Macneil, “Design
patterns for active learning,” Faculty Experiences in Active Learning:
A Collection of Strategies for Implementing Active Learning Across
Disciplines, pp. 130–158, 2020.
[41] N. Dehbozorgi and S. MacNeil, “Semi-automated analysis of reflections
as a continuous course,” in 2019 IEEE Frontiers in Education Confer-
ence (FIE), 2019, pp. 1–5.
[42] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion,
O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vander-
plas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch-
esnay, “Scikit-learn: Machine learning in Python,” Journal of Machine
Learning Research, vol. 12, pp. 2825–2830, 2011.
[43] R. Sanche and K. Lonergan, “Variable reduction for predictive modeling
with clustering,” in Casualty Actuarial Society Forum. Citeseer, 2006,
pp. 89–100.
Authorized licensed use limited to: Kennesaw State University. Downloaded on January 23,2023 at 16:19:51 UTC from IEEE Xplore. Restrictions apply.
