IEEE Circuits and Systems Magazine - Q4 2019 - 34

300

271

250
200

167

150

133

100

17

16

20

14

15

20

20

20

11

20

13

10

58

20

09

20

44

21

12

7

20

2

20

1

0

20

0

08

50

Figure 15. Citation report for total publications.

2,000
1,800
1,600
1,400
1,200
1,000
800
600
400
200
0

1,813
1,431

770

26

11
20
12
20
13
20
14
20
15
20
16
20
17

10

1

20

09

20

5

20

2

20

1

08

354
169

Figure 16. Citation report for sum of times cited per year.

Table II.
A comparison of deep learning based approaches
on speech generation.

34

Paper

Evaluation Results

Zen et al. [47]

39% (Subjective Preference)

Wu et al. [67]

74% (Subjective Preference)

Mehri et al. [69]

80% (Subjective Preference)

Fan et al. [70]

59% (Subjective Preference)

Ling et al. [45]

39% (Subjective Preference)

Zen et al. [64]

3.81 ± 0.11 (5-scale MOS)

Oord et al. [88]

4.21 ± 0.08 (5-scale MOS)

Oord et al. [89]

4.41 ± 0.08 (5-scale MOS)

Ö. Arık et al. [90]

3.94 ± 0.26 (5-scale MOS)

Ö. Arık et al. [91]

3.53 ± 0.12 (5-scale MOS)

Ping et al. [92]

3.78 ± 0.30 (5-scale MOS)

Wang et al. [93]

3.82 ± 0.09 (5-scale MOS)

Shen et al. [94]

4.53 ± 0.07 (5-scale MOS)

Wan et al. [96]

4.53 ± 0.07 (5-scale MOS)

IEEE CIRCUITS AND SYSTEMS MAGAZINE

a DCGAN-based model was used for generating more
samples to augment the DCASE 2017 dataset. Random
values were added to the generated samples to alleviate
the issue of sample bias and over-fitting. A support vector
machine (SVM) hyper-plane was trained to sift the samples of suitable quality from the generated feature pool. It
was confirmed that the usage of the augmented data can
improve the acoustic scene classification performance
on the DCASE development set.
V. Discussion
The latest developments and applications of deep learning algorithms are reviewed in this paper. Undoubtedly,
the introduction of deep generative and discriminative
methods in the field of speech/acoustic signal processing boosts the accuracy and the efficiency in recognition
and synthesis problems. Fig. 15 and Fig. 16 are the citation analysis from the database of the Web of Science.
The search keywords are 'speech synthesis' and 'deep
learning.' The bars in Fig. 15 denote the numbers of the
publications recorded in the database of the Web of Science from 2008 to 2017. The curve in Fig. 16 is the total
number of citations per year. It is obvious that speech
synthesis applications have gripped the deep learning
research community.
In Table II, evaluation results of several deep learning
based speech generation methods are given for a comparison. The data listed in the table are from the best
performance of the subjective experiments in each referenced paper. Given the different speech corpora used
for synthesis and different implementation protocols,
Table II basically demonstrates an increasing quality
of the generated speech samples by employing various
deep learning techniques.
In this section, the key reasons behind the success
of deep learning algorithms with audio generation are
discussed and several potential issues for further consideration are presented.
A. Benefits of Deep Learning
The introduction of deep learning is an important
breakthrough for the speech processing community.
More practical problems in audio generation can be
simplified and solved in a much more refined manner.
The accessibility and expandability of deep learning
techniques are convenient for users to implement audio synthesis systems, not only data experts but also
non-expert developers. Given enough data, computational resources and suitable algorithms, deep learning
is able to address problems from a variety of aspects in
acoustic signal generation.
Listed are several key advantages of deep learning
based audio generation systems:
FOURTH QUARTER 2019
IEEE Circuits and Systems Magazine - Q4 2019

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q4 2019
