=Paper=
{{Paper
|id=Vol-2491/paper60
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2491/paper60.pdf
|volume=Vol-2491
|dblpUrl=https://dblp.org/rec/conf/bnaic/RopkeREN19
}}
==None==
https://ceur-ws.org/Vol-2491/paper60.pdf
Training a Speech-to-Text Model for Dutch on
the Corpus Gesproken Nederlands ?,??
Willem Röpke, Roxana Rădulescu, Kyriakos Efthymiadis, and Ann Nowé
Artificial Intelligence Research Group, Vrije Universiteit Brussel, Belgium
{Willem.Ropke,Roxana.Radulescu,Kyriakos.Efthymiadis,Ann.Nowe}@vub.be
Abstract. Speech-to-text, also known as Speech Recognition, is a tech-
nology that is able to recognize and transcribe spoken language into
text. In subsequent steps, this transcription can be used to complete a
multitude of tasks, such as providing automatic subtitles or parsing voice
commands. In recent years, Speech-to-Text models have dramatically im-
proved thanks partially to advances in Deep Learning methods. Starting
from the open-source project DeepSpeech, we train speech-to-text mod-
els for Dutch, using the Corpus Gesproken Nederlands (CGN). First, we
contribute a pre-processing pipeline for this dataset, to make it suitable
for the task at hand, obtaining a ready-to-use speech-to-text dataset
for Dutch. Second, we investigate the performance of Dutch and Flemish
models trained from scratch, establishing a baseline for the CGN dataset
for this task. Finally, we investigate the issue of transferring speech-to-
text models between related languages. In this case, we analyse how a
pre-trained English model can be transferred and fine-tuned for Dutch.
Keywords: Speech Recognition · Corpus Gesproken Nederlands · Deep-
Speech · Speech-to-Text
1 Introduction
The field of Artificial Intelligence has managed to achieve unprecedented re-
sults in the last couple of decades, across many fields, ranging from computer
vision, e.g., beating human performance at the image recognition task [31], to
reinforcement learning, e.g., achieving superhuman performance at the game of
Go [33]. Many of these developments are already being applied to everyday life,
in the form of product recommendation systems [32], vehicles with partially
autonomous capabilities [9] or financial fraud detecting systems [36].
This work will take a closer look at another one of these fields, namely Speech-
to-Text (STT), otherwise referred to as Speech Recognition (SR). Simply put,
speech-to-text is the ability of a machine to “hear” spoken language and tran-
scribe it. In subsequent steps, this transcription can be used to perform certain
?
Copyright c 2019 for this paper by its authors. Use permitted under Creative Com-
mons License Attribution 4.0 International (CC BY 4.0).
??
Part of this work was carried out by the first author for his bachelor thesis [30],
under the supervision of the other authors.
2 W. Röpke et al.
actions or tasks. In recent years we have seen the practical application of this
process in consumer directed systems such as Google Assistant1 , Siri2 , or Alexa3
[21]. STT as a field has many other practical applications, ranging from health-
care [7] and helping people with disabilities [23], to controlling military aircrafts
[5, 34]. This potential societal impact makes STT an attractive research sub-
ject. However, STT has faced a multitude of difficulties along its history. For
starters, early models were simply not expressive enough to grasp the complex-
ity of human speech and later models like artificial neural networks required
massive datasets and substantial computing power, which up until recently have
not been available.
A vital component for obtaining high quality STT systems is thus having
a large and well curated dataset. There are currently numerous STT datasets
available for English, such as LibriSpeech [26], TED-LIUM [16] and VoxCeleb
[2, 22]. However, projects such as Common Voice4 aim to facilitate the creation
of community curated datasets for a larger variety of languages. It currently
contains data across 29 languages, including 23 hours of recordings for Dutch.
This relatively small size has led us to look for a different dataset. For this
work we have selected the Corpus Gesproken Nederlands [24, 35]5 . This corpus
consists of 900 hours of spoken Dutch in both the Flemish and Dutch dialects.
A detailed description together with how we pre-process CGN for this work is
presented in Section 4.
For training our speech-to-text Dutch models we rely on Mozilla’s open-
source project DeepSpeech6 , which offers an implementation of the neural archi-
tecture presented by Hannun et al. [14]. This allows us to train speech-to-text
models for Dutch, using our desired data. We offer an overview of the theoretical
concepts involved in creating this deep learning architecture in Section 2, while
Section 3 presents other lines of work related to this topic.
We investigate two types of models trained using DeepSpeech on the CGN.
First we consider training a model from scratch and establish a baseline for
how well CGN can perform when training an end-to-end speech-to-text model
on a neural architecture. Afterwards, we also investigate if we can boost the
obtained performance by taking advantage of a high performing model trained
on an extensive English dataset and fine-tune it using CGN. An overview of our
results is presented in Section 5.
Contributions We can summarize the contribution of our work as follows:
(i) we provide a full analysis and contribute a pre-processing pipeline for the
CGN dataset aimed for use in an end-to-end neural architecture for speech-to-
text; (ii) we train and evaluate STT models for Dutch using an end-to-end deep
1
https://assistant.google.com
2
https://www.apple.com/siri/
3
https://developer.amazon.com/alexa
4
https://voice.mozilla.org/en/datasets
5
https://ivdnt.org/downloads/tstc-corpus-gesproken-nederlands
6
https://github.com/mozilla/DeepSpeech
Training a Speech-To-Text Model for Dutch 3
0.3
0.2
0.1
amplitude
0.0
0.1
0.2
0.3
0.0 0.5 1.0 1.5 2.0 2.5
time [s]
(a) Waveform of the utterance “eindelijk hebben de olifanten de boodschap begrepen”.
(b) Spectrogram of the utterance “eindelijk hebben de olifanten de boodschap be-
grepen”.
Fig. 1: Two different representations of the same utterance.
learning approach in the form of the open-source project DeepSpeech, providing
a benchmark for Dutch models on the CGN dataset; (iii) we investigate and
evaluate the possibility of transfer learning between English and Dutch STT
models.
2 Background
2.1 Speech Spectrograms
The task of speech-to-text is to find a mapping between natural language ut-
terances and their corresponding written form representation, i.e., a textual
transcriptions. The input for the DeepSpeech end-to-end deep learning systems
consists of the representation of the audio file in the frequency domain, i.e., a
spectrogram. Speech spectrogram allows one to visualise the spectrum of fre-
quencies of an audio signal and identify phonetic information of the given input.
Figure 1b presents an example for the spectrogram of the utterance “eindelijk
hebben de olifanten de boodschap begrepen” (a fragment from a Flemish audio
file in component h, lessons recorded in a classroom, Table 1), while Figure 1a
4 W. Röpke et al.
presents the same utterance as a waveform plot (representations created using
the Parselmouth library [19]).
2.2 Recurrent Neural Networks
Speech data has a sequential character, meaning that order plays an impor-
tant part in understanding or processing the given information. This has proven
hard to tackle using regular feed-forward neural networks. Recurrent neural net-
works (RNNs) are artificial neural networks specialized in processing sequential
data [10] (e.g., time series, DNA sequences or speech).
Given an input x representing a time-series of length T , we are interested in
predicting a sequence of character probabilities that will form our transcription.
After obtaining this prediction, DeepSpeech uses the CTC loss [11] to measure
the prediction error for training the network. Every component of x consists of a
vector of audio features, in the form of frequency information as detailed above.
The characters we consider in order to train our Dutch models belong to the
set: {a, b, c, . . ., z, ’, &, -, space}. We have included the special characters ’,
& and - because they were already present in multiple utterances in the CGN.
It would be feasible to exclude these characters from the set with some extra
pre-processing steps.
Fig. 2: Recurrent neural architecture for the DeepSpeech project, figure adapted
from [14].
Figure 2 presents the architecture used in the DeepSpeech project. The net-
work has a simple architecture, consisting in 5 hidden layers, out of which only
(f )
the fourth one is a bi-directional recurrent layer with forward (ht ) and back-
(b)
ward (ht ) units. This allows the network to not only capture past information
Training a Speech-To-Text Model for Dutch 5
at a given time-step t, but to actually output predictions that depend on the
entire input sequence, as we have two recurrent layers that move both forward
(from the start to the end of the sequence) and backwards (from the end to the
start of the sequence) through time. When passing through the first layer, for
every time-step t, we consider both the spectrogram frame xt , together with the
surrounding C frames on each side.
This architecture has led to very good results when trained on English con-
versational speech data, reaching a word error rate of 16% on the Switchboard
dataset, when trained on both the Switchboard and Fisher datasets [14]. More
recent models using this architecture have scored a word error rate as low as
8.22% on the LibriSpeech clean test corpus.
Auxiliary Language Model Because an RNN will output a character by
character transcription of the input, the network will often produce phonetically
correct interpretations, but incorrect spellings. To remedy this, an N-gram lan-
guage model is attached to the output. What this language model does is to try
and find the most likely sequence of words instead of characters. Afterwards the
sequence that maximizes a scoring function that takes the output of the RNN
and the language model into account is chosen as the final output.
3 Related Work
Advances in the field of Deep Learning, together with newly developed archi-
tectures have led to spectacular results for the speech-to-text domain [1, 12–14].
However, state-of-the-art models are usually trained and benchmarked on En-
glish datasets.
Research into STT using neural architectures for Dutch has been very lim-
ited, with a few exceptions [8]. However, approaches based on Hidden Markov
Models (HMMs) and statistical language models has been extensively explored
in the literature [3, 6, 17, 25]. A notable initiative was Project SPRAAK7 [4, 27],
an open source speech-to-text package for Dutch, also based on an HMM ar-
chitecture. Kessens and Van Leeuwen [20] have also proposed the N-Best 2008
benchmark dataset (i.e., the news broadcast and telephone dialogue components
from the CGN), in order to evaluate the SPRAAK system.
For our work, we rely on the architecture developed by the open-source
project DeepSpeech. Other similar frameworks include PyTorch-Kaldi [29] and
wav2letter++ [28]. Finally, a similar initiative to this current work has also been
applied to other languages, as for example Russian [18].
4 Pipeline for Processing the Corpus Gesproken
Nederlands
The Spoken Dutch Corpus is a database of Dutch as spoken by adults from the
Netherlands and Flanders. Its contents entail almost 9 million words, adding to a
7
https://www.spraak.org
6 W. Röpke et al.
total of 900 hours of spoken Dutch. The goal of this project was to create a corpus
that would form a plausible sample of contemporary Dutch as spoken in the
Netherlands and Flanders. In creating the corpus, the developers have attempted
to assemble it to optimally suit the needs of diverse research disciplines and
applications. For this reason, all data is sorted by type of speech into multiple
components as shown in Table 1. Each component is divided into Flemish audio
files and Dutch audio files. This represents a split of 76.23% and 23.77% of Dutch
and Flemish audio files respectively.
Table 1: Components available in the Corpus Gesproken Nederlands dataset with
their corresponding descriptions.
Component Description
comp-a spontaneous conversations (face-to-face)
comp-b interviews with teachers of Dutch
comp-c spontaneous telephone dialogues (recorded via a switchboard)
comp-d spontaneous telephone dialogues (recorded on MD with local interface)
comp-e simulated business negotiations
comp-f interviews/discussions/debates (broadcast)
comp-g (political) discussions/debates/meetings (non-broadcast)
comp-h lessons recorded in a classroom
comp-i live (eg sport) commentaries (broadcast)
comp-j newsreports/reportages (broadcast)
comp-k news (broadcast)
comp-l commentaries/columns/reviews (broadcast)
comp-m ceremonious speeches/sermons
comp-n lectures/seminars
comp-o read speech
Out of the box, the Corpus Gesproken Nederlands does not come ready to
efficiently train a speech-to-text model. Neural-based STT architectures, includ-
ing DeepSpeech, will not perform well when trained on long audio files. The
usual approach is to split each file in smaller chunks and then use those for the
training process. As a consequence, we have processed the CGN to better fit
our needs. We do this in multiple parts, starting with first splitting the audio
files into shorter segments, and afterwards presenting two approaches targeted
at reducing the noise in the dataset. All the code used to build this pipeline is
available at: https://github.com/wilrop/Import-CGN.
4.1 Segmenting the Audio Files
The average length of an audio file in the original dataset is approximately 234
seconds, or 3 minutes and 54 seconds with a standard deviation of 283. Having a
dataset for STT with audio files this long and a variance this high would make it
Training a Speech-To-Text Model for Dutch 7
very difficult to efficiently train a model. For this reason we have opted to split
audio files into smaller segments.
In the XML transcriptions of the corpus, timestamps for the beginning and
ending of each sentence in an audio file were provided. We used these timestamps
to group several sentences together in a new audio file until the length of this
file would reach a certain threshold, in our case 3 seconds. Using this strategy
we had a good balance between the amount of files and the length of these files.
The Flemish set now has a mean of a bit more than 6 seconds with a standard
deviation of 1.6 and the Dutch set has a mean of 6 seconds with a standard
deviation of 1.5.
4.2 Handling Noise in the Dataset
After analysing the output of our first iteration of models it quickly became
obvious that we were dealing with an excessively noisy dataset, thereby making
it more difficult to accurately train and score our models. To tackle this problem,
we have developed two methods.
Our first solution to this problem was to generate a smaller dataset, that
would exclude certain components completely. We hypothesized that a lot of the
noise in the data comes from face-to-face speech or spontaneous dialog, a noisy
settings even for humans. These components would have a greater likelihood of
people interrupting each other or talking over each other, resulting in incorrect
transcriptions. We have identified components a, c, d, f and g to contain the
highest amount of this type of speech (see Table 1 for a description of these
components). By excluding these components we hoped to still have an adequate
amount of audio files in the dataset, but with better quality. This dataset will
also be referred to as the small dataset from here on after.
The second method tries to tackle the noise in the dataset in a different way.
When inspecting transcriptions in the dataset, we noticed that several times,
completely different transcriptions would start at overlapping timestamps. Be-
cause there is no way to know which transcription, if any, is correct in this case
without a laborious manual process, we decided to leave out all instances where
this issues occurred. Using this method we have generated the non overlapping
dataset. Finally, we have also created a third dataset, by combining the two noise
handling techniques presented above: the small non overlapping dataset. These
partitions were generated for Dutch, Flemish and also for a combination of the
two, having obtained in the end 9 datasets. Tables 2, 3, 4 present the amount of
hours present per split in each of the created datasets.
5 Training a Speech-To-Text Dutch Model
We present here the results we obtained for each subset we generated from the
CGN dataset, for both models trained from scratch and also by fine-tuning a
state-of-the-art English model.
8 W. Röpke et al.
Table 2: Amount of hours of audio in the splits for each Flemish dataset.
Train Validation Test
Small non overlapping 24.2 6.0 7.6
Non overlapping 34.1 8.5 10.6
Small 60.3 15.0 18.8
Table 3: Amount of hours of audio in the splits for each Dutch dataset.
Train Validation Test
Small non overlapping 39.2 9.8 12.2
Non overlapping 61.3 15.3 19.2
Small 81.8 20.5 25.6
Auxiliary Language Model for Dutch Similar to the original approach
present by Hannun et al. [14], to aid our prediction process we have created an
5-gram language model for Dutch. We have done this only once, because the
Dutch and Flemish dialects follow the same grammar and spelling. The 5-gram
language model used in our experiments was trained on a collection of over one
million phrases, adding up to a vocabulary of more than 160 thousand words,
gathered from the CGN. We used KenLM8 , which is shown to provide faster
and smaller language queries than other systems [15] to efficiently generate a
language model.
5.1 Experimental Setting
Splitting the Dataset We have opted to split each dataset into 80%–20% for
training and testing. We also create a validation dataset by making again an
80%–20% split on the training data.
In order to obtain high-performing models, we do a parameter search for
the number of hidden nodes per layer, learning rate and dropout, as shown in
Table 5. We fix the rest of the parameters and present the exact values used
during the training process in Table 6.
The metrics used to evaluate our results are the Word Error Rate (WER) and
Character Error Rate (CER). The CER is computed as the Levenshtein distance,
8
https://github.com/kpu/kenlm
Table 4: Amount of hours of audio in the splits for each combined dataset.
Train Validation Test
Small non overlapping 63.4 15.9 19.8
Non overlapping 95.4 23.8 29.8
Small 142.1 35.5 44.5
Training a Speech-To-Text Model for Dutch 9
Table 5: All variables and their values experimented with as parameters during
training.
Variable Values
Amount of hidden nodes per layer 512, 1024 and 2048
Learning rate 0.001, 0.00055 and 0.0001
Dropout 0.2, 0.35 and 0.5
Table 6: Flags set in training of our models that were kept constant.
Flag Value
–train batch size 64
–dev batch size 32
–test batch size 32
–epochs 30
–use warpctc True
–early stop True
–es steps 10
–es mean th 0.1
–es std th 0.1
which represents the minimum number of operations required to transform the
ground truth into the output, divided by the total amount of characters in this
ground truth. The WER is defined as the Levenshtein distance on word level
divided by the amount of words in the original text.
5.2 Training a Dutch Model from Scratch
We have trained models for each of the sub-datasets we created for Dutch,
Flemish and the combined data: small, non-overlapping and the small non-
overlapping datasets. This has also allowed us to evaluate the noise handling
techniques developed for CGN.
Table 7: Best performing models on the Flemish datasets.
Nodes Learning Rate Dropout WER CER
Small Non Overlapping 2048 0.00055 0.35 0.386307 0.231521
Non Overlapping 2048 0.00055 0.35 0.469610 0.295521
Small 2048 0.00010 0.35 0.519958 0.366938
Table 7 presents our results on the Flemish data, together with the parame-
ters that lead to these results. We note again that the Flemish data represents
less than a quarter of the CGN dataset. We hypothesize that the high WER
10 W. Röpke et al.
and CER values here are also caused by the limited amount of data that was
available for the training process. Notice that removing the files with overlap-
ping timestamps proved to be a better noise handling strategy compared to the
removal of face-to-face speech and the spontaneous dialogues (i.e., Non Overlap-
ping versus Small). However, the best results of 38.6% WER and 23.1% CER
were obtained when combining the two noise handling strategies.
Table 8 presents the same results on the Dutch data. We note that due
to a higher amount of data available for training, the performance of our best
model has increased to 34.2% WER and 20.5% CER. Notice also that for this
dataset, removing face-to-face speech and spontaneous dialogue proves to be
more efficient than removing the files having overlapping transcriptions. After a
short analysis, we observed that this is due to a smaller number of overlapping
timestamps in the Dutch dataset.
Table 8: Best performing models on the Dutch datasets.
Nodes Learning Rate Dropout WER CER
Small Non Overlapping 2048 0.00055 0.35 0.342087 0.205417
Non Overlapping 2048 0.00010 0.50 0.466391 0.277180
Small 2048 0.00055 0.20 0.419772 0.273178
Finally, looking at Table 9, we remark that having more data proves re-
ally beneficial for our case, obtaining in the end a performance of about 30%
WER and 17.2% CER on the combined dataset, when using both noise han-
dling techniques. We also note here that the best performing models for the non
overlapping and small datasets had 1024 hidden nodes per layer. The experi-
ments on these datasets were solely ran on 512 and 1024 hidden nodes per layer,
due to only having a marginal increase going from 1024 to 2048 for all other
experiments.
Table 9: Best performing models on the combined datasets.
Nodes Learning Rate Dropout WER CER
Small Non Overlapping 2048 0.0001 0.20 0.299879 0.172096
Non Overlapping 1024 0.0001 0.35 0.408456 0.261991
Small 1024 0.0001 0.20 0.460147 0.326860
5.3 Transfer Learning from English to Dutch
In order to create the best possible models for Dutch, we have attempted to
apply transfer learning, by fine-tuning a well performing English model on our
Training a Speech-To-Text Model for Dutch 11
datasets. The model that we have selected was provided in Project DeepSpeech
and has a WER of 8.22% on the LibriSpeech clean test corpus. All experiments
for transfer learning were conducted using the small non overlapping dataset,
which gave the best performing models for all previous experiments.
In these experiments we have trained each neural network in cycles of five
epochs, evaluating the resulting network after every cycle. In the following cy-
cle, the network resumed training from the best checkpoint of the previous cycle.
This was done for 5, 10, 15, 20, 25 and 30 epochs. We have also tried to push
the best performing networks for 40 epochs, but noticed a decrease in perfor-
mance which can most likely be attributed to overfitting. To obtain a baseline
for the performance of the English neural network without any fine-tuning we
have evaluated this neural network (combined with the Dutch M-gram language
model) on the test sets of Flemish, Dutch and the combined languages. We can
clearly see in Table 10 that we obtain a poor initial performance.
Table 10: Performance of the English model (combined with the Dutch n-gram
language model) on our datsets before fine-tuning.
WER CER
Flemish 0.978428 0.611835
Dutch 0.965037 0.631461
Combined 0.972322 0.624038
Table 11 presents the results of our fine-tuning process for Flemish, Dutch
and the combined data. We notice a large boost in performance, especially for
the Flemish and combined datasets, which reach a WER of 22.9% and 23.6%
respectively. This result emphasises the importance of having well-curated and
large enough datasets, when training neural STT models. Furthermore, we can
conclude that the close relationship between English and Dutch allows us to
leverage the features learned while training on the English dataset for our Dutch
models.
Table 11: Best performing models with transfer learning.
Epochs Learning Rate Dropout WER CER
Flemish 20 0.0001 0.15 0.229560 0.133830
Dutch 30 0.0001 0.20 0.255273 0.145501
Combined 30 0.0001 0.20 0.235773 0.134833
12 W. Röpke et al.
6 Conclusion
In this work we have explored how well speech-to-text neural models for Dutch
can perform on the Corpus Gesproken Nederlands. To this end we used the archi-
tecture proposed by the DeepSpeech project, which we tuned for our purposes.
Because the CGN was not directly usable for this task, we provided a full pro-
cessing pipeline that also includes two noise handling techniques. We evaluated
these approaches, by training from scratch models for Dutch, Flemish and a
combination of the two. We conclude that, in most cases, removing the files with
overlapping transcriptions was more efficient than removing face-to-face speech
and spontaneous dialogue, however the best results were obtained when com-
bining both techniques. For models trained from scratch, our best performing
models reached a WER of 30.0% and a CER of 17.2%.
A second step in our quest to obtain good quality STT models for Dutch
was the attempt to fine-tune a high-performing English model for our dataset.
Using this approach we obtained a boost in performance, reaching a WER of
23.0% and CER of 13.4%. This is a strong indication that in the absence of a
large volume of high quality data for a specific language, transfer learning could
be used to enable successful training on smaller datasets.
For the future, we are interested in further improving and curating Dutch
datasets for speech-to-text. This includes developing further techniques for han-
dling noisy data, but also including other smaller datasets such at the one avail-
able through the Common Voice project. We are also interested in developing
different neural architectures and also using other frameworks such as PyTorch-
Kaldi for this task.
References
1. Amodei, D., Ananthanarayanan, S., Anubhai, R., Bai, J., Battenberg, E., Case,
C., Casper, J., Catanzaro, B., Cheng, Q., Chen, G., et al.: Deep speech 2: End-
to-end speech recognition in english and mandarin. In: International conference on
machine learning. pp. 173–182 (2016)
2. Chung, J.S., Nagrani, A., Zisserman, A.: Voxceleb2: Deep speaker recognition. In:
INTERSPEECH (2018)
3. Demuynck, K., Puurula, A., Van Compernolle, D., Wambacq, P.: The esat 2008
system for n-best dutch speech recognition benchmark. pp. 339 – 344 (01 2010).
https://doi.org/10.1109/ASRU.2009.5373311
4. Demuynck, K., Roelens, J., Compernolle, D.V., Wambacq, P.: Spraak: An open
source” speech recognition and automatic annotation kit”. In: Ninth Annual Con-
ference of the International Speech Communication Association (2008)
5. Draper, M., Calhoun, G., Ruff, H., Williamson, D., Barry, T.: Manual versus speech
input for unmanned aerial vehicle control station operations. In: Proceedings of
the Human Factors and Ergonomics Society Annual Meeting. vol. 47, pp. 109–113.
SAGE Publications Sage CA: Los Angeles, CA (2003)
6. Duchateau, J., Uytsel, D.H.V., Hamme, H.V., Wambacq, P.: Statistical language
models for large vocabulary spontaneous speech recognition in dutch. In: Ninth
European Conference on Speech Communication and Technology (2005)
Training a Speech-To-Text Model for Dutch 13
7. Durling, S., Lumsden, J.: Speech recognition use in healthcare applications. In:
Proceedings of the 6th international conference on advances in mobile computing
and multimedia. pp. 473–478. ACM (2008)
8. Evers, J.: Using lstms to improve dutch language models (2017)
9. Fagnant, D.J., Kockelman, K.: Preparing a nation for autonomous vehicles: oppor-
tunities, barriers and policy recommendations. Transportation Research Part A:
Policy and Practice 77, 167–181 (2015)
10. Goodfellow, I., Bengio, Y., Courville, A.: Deep learning. MIT press (2016)
11. Graves, A., Fernández, S., Gomez, F., Schmidhuber, J.: Connectionist temporal
classification: labelling unsegmented sequence data with recurrent neural networks.
In: Proceedings of the 23rd international conference on Machine learning. pp. 369–
376. ACM (2006)
12. Graves, A., Jaitly, N.: Towards end-to-end speech recognition with recurrent neural
networks. In: International Conference on Machine Learning. pp. 1764–1772 (2014)
13. Graves, A., Mohamed, A.r., Hinton, G.: Speech recognition with deep recurrent
neural networks. In: Acoustics, speech and signal processing (icassp), 2013 ieee
international conference on. pp. 6645–6649. IEEE (2013)
14. Hannun, A., Case, C., Casper, J., Catanzaro, B., Diamos, G., Elsen, E., Prenger,
R., Satheesh, S., Sengupta, S., Coates, A., et al.: Deep speech: Scaling up end-to-
end speech recognition. arXiv preprint arXiv:1412.5567 (2014)
15. Heafield, K.: KenLM: faster and smaller language model queries. In: Proceedings
of the EMNLP 2011 Sixth Workshop on Statistical Machine Translation. pp. 187–
197. Edinburgh, Scotland, United Kingdom (July 2011), https://kheafield.com/
papers/avenue/kenlm.pdf
16. Hernandez, F., Nguyen, V., Ghannay, S., Tomashenko, N., Estève, Y.: TED-LIUM
3: twice as much data and corpus repartition for experiments on speaker adapta-
tion. In: International Conference on Speech and Computer. pp. 198–208. Springer
(2018)
17. Huijbregts, M., Ordelman, R., de Jong, F.M.: A spoken document retrieval appli-
cation in the oral history domain. In: Proceedings of 10th international conference
Speech and Computer, Patras, Greece (SPECOM 2005). vol. 2, pp. 699–702. Uni-
versity of Patras/WCL Moscow State Linguistics Uni. (2005)
18. Iakushkin, O., Fedoseev, G., Shaleva, A., Degtyarev, A., Sedova, O.: Russian-
language speech recognition system based on deepspeech (2018)
19. Jadoul, Y., Thompson, B., de Boer, B.: Introducing Parselmouth:
A Python interface to Praat. Journal of Phonetics 71, 1–15 (2018).
https://doi.org/https://doi.org/10.1016/j.wocn.2018.07.001
20. Kessens, J., Van Leeuwen, D.: N-best: The northern- and southern-dutch bench-
mark evaluation of speech recognition technology. vol. 2, pp. 1354–1357 (01 2007)
21. López, G., Quesada, L., Guerrero, L.A.: Alexa vs. siri vs. cortana vs. google as-
sistant: a comparison of speech-based natural user interfaces. In: International
Conference on Applied Human Factors and Ergonomics. pp. 241–250. Springer
(2017)
22. Nagrani, A., Chung, J.S., Zisserman, A.: Voxceleb: a large-scale speaker identifica-
tion dataset. In: INTERSPEECH (2017)
23. Noyes, J., Frankish, C.: Speech recognition technology for individuals with disabil-
ities. Augmentative and Alternative Communication 8(4), 297–303 (1992)
24. Oostdijk, N.: The Spoken Dutch Corpus. Overview and First Evaluation. In: LREC
(2000)
25. Ordelman, R., Ordelman, R.: Dutch Speech Recognition in Multimedia Information
Retrieval. Ph.D. thesis, Netherlands (10 2003)
14 W. Röpke et al.
26. Panayotov, V., Chen, G., Povey, D., Khudanpur, S.: Librispeech: an asr corpus
based on public domain audio books. In: 2015 IEEE International Conference on
Acoustics, Speech and Signal Processing (ICASSP). pp. 5206–5210. IEEE (2015)
27. Pelemans, J., Demuynck, K., Wambacq, P.: Dutch automatic speech recognition
on the web: Towards a general purpose system. In: Thirteenth Annual Conference
of the International Speech Communication Association (2012)
28. Pratap, V., Hannun, A., Qiantong Xu, J.C., Kahn, J., Synnaeve, G., Liptchin-
sky, V., Collobert, R.: wav2letter++: The fastest open-source speech recognition
system. arXiv preprint arXiv:1812.07625 (2018)
29. Ravanelli, M., Parcollet, T., Bengio, Y.: The pytorch-kaldi speech recognition
toolkit. In: In Proc. of ICASSP (2019)
30. Röpke, W.: Building a Speech-to-Text Engine for Dutch. Bachelor thesis, Vrije Uni-
versiteit Brussel (2019), https://ai.vub.ac.be/files/Ropke_Bachelor_thesis_
1819.pdf
31. Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z.,
Karpathy, A., Khosla, A., Bernstein, M., et al.: Imagenet large scale visual recog-
nition challenge. International journal of computer vision 115(3), 211–252 (2015)
32. Shaya, S.A., Matheson, N., Singarayar, J.A., Kollias, N., Bloom, J.A.: Intelligent
performance-based product recommendation system (2010), uS Patent 7,809,601
33. Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G.,
Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., et al.: Master-
ing the game of go with deep neural networks and tree search. nature 529(7587),
484 (2016)
34. Steeneken, H.: Potentials of speech and language technology systems for military
use: An application and technology-oriented survey. Tech. rep., Human Factors
Research Institute, TNO Soesterberg (The Nederlands) (1996)
35. Van Eerten, L.: Over het corpus gesproken nederlands. Nederlandse Taalkunde
12(3), 194–215 (2007)
36. West, J., Bhattacharya, M.: Intelligent financial fraud detection: a comprehensive
review. Computers & security 57, 47–66 (2016)