arXiv:2112.03557v1  [cs.CL]  7 Dec 2021Multi-speaker Emotional Text-to-speech Synthesizer
Sungjae Cho1,†, Soo-Young Lee2
1Korea Institute of Science and Technology, Republic of Kore a
2Korea Advanced Institute of Science and Technology, Republ ic of Korea
sj.cho@snu.ac.kr, sylee@kaist.ac.kr
Abstract
We present a methodology to train our multi-speaker emotion al
text-to-speech synthesizer that can express speech for 10 s peak-
ers’ 7 different emotions. All silences from audio samples a re
removed prior to learning. This results in fast learning by o ur
model. Curriculum learning is applied to train our model effi -
ciently. Our model is first trained with a large single-speak er
neutral dataset, and then trained with neutral speech from a ll
speakers. Finally, our model is trained using datasets of em o-
tional speech from all speakers. In each stage, training sam ples
of each speaker-emotion pair have equal probability to appe ar
in mini-batches. Through this procedure, our model can synt he-
size speech for all targeted speakers and emotions. Our synt he-
sized audio sets are available on our web page.
Index Terms : emotional speech synthesis, text-to-speech, ma-
chine learning, neural network, deep learning
1. Introduction
Emotional speech synthesis has been achieved through deep
neural networks [1, 2, 3, 4]. However, most studies have trai ned
models on a small number of speakers or balanced class distri -
butions because it is challenging to guarantee speech quali ty for
each speaker and emotion, given imbalanced data distributi ons
with respect to speakers and emotions. In this paper, we pres ent
a methodology for training our multi-speaker emotional tex t-to-
speech (TTS) synthesizer capable of generating speech for a ll
targeted speakers’ voices and emotions. The main methods ar e
silence removal, curriculum learning [5], and oversamplin g [6].
The synthesized audios are demonstrated through a web page.
2. Datasets
4 datasets were used to train the multi-speaker emotional TT S
synthesizer. The first dataset, the Korean single speaker sp eech
(KSS) dataset [7], is publicly available and contains speec h
samples of a single female speaker: kss-f. We labeled their
emotion as neutral. The remaining 3 datasets consist of spee ch
of the Ekman’s 7 basic emotions [8]: neutral, anger, disgust ,
fear, happiness, sadness, and surprise.
The first Korean emotional TTS (KETTS) dataset consists
of 1 female and 1 male speaker: ketts-30f and ketts-30m, whic h
are abbreviations of a 30’s female and male in KETTS. The 2
speakers were assigned to different sets of sentences; howe ver,
the same sentences were recorded across 7 emotions for a sing le
speaker. In the female case, only happy speech samples have a
different set of sentences. KETTS is balanced with respect t o
speakers and emotions, except for the female’s happy speech
subset (Table 1).
The second Korean emotional TTS (KETTS2) dataset con-
sists of 3 female and 3 male speakers, totally 6 speakers: ket t2-
†work done at KAISTTable 1: Hours of preprocessed training datasets
Speaker all neu ang dis fea hap sad sur
kss-f 12.59 12.59
ketts-30f 26.61 3.52 3.46 3.51 3.68 5.13 3.75 3.56
ketts-30m 24.12 3.37 3.29 3.31 3.51 3.50 3.73 3.40
ketts2-20m 5.09 0.72 0.72 0.74 0.76 0.69 0.75 0.70
ketts2-30f 4.69 0.66 0.65 0.67 0.65 0.70 0.68 0.68
ketts2-40m 4.98 0.73 0.69 0.70 0.75 0.69 0.74 0.69
ketts2-50f 4.98 0.73 0.71 0.71 0.70 0.72 0.71 0.69
ketts2-50m 4.73 0.68 0.68 0.69 0.67 0.68 0.68 0.65
ketts2-60f 4.90 0.77 0.68 0.67 0.68 0.72 0.72 0.67
ketts3-f 9.64 3.96 1.34 1.27 1.44 1.64
ketts3-m 9.38 3.90 1.43 1.18 1.39 1.48
all 111.70 31.63 13.65 11.01 13.85 15.64 14.87 11.05
20m, ketts2-30f, ketts2-40m, ketts-50f, ketts2-50m, and k etts2-
60f. The same sentences were recorded across 7 emotions and 6
speakers. Hence, KETTS2 is balanced with respect to speaker s
and emotions (Table 1).
The third Korean emotional TTS (KETTS3) dataset con-
sists of 1 female and 1 male speaker: ketts3-f and ketts3-m. I t
includes 5 emotions, excluding disgust and surprise. The sa me
sentences were recorded across 2 speakers; however, differ ent
sentences were spoken for the 5 emotions. KETTS3 is balanced
for speakers but not for emotions. Therefore, the whole trai ning
dataset is balanced for neither speakers nor emotions (Tabl e 1).
3. Methodology
3.1. Preprocessing
The WebRTC voice activity detector, py-webrtcvad1, is utilized
to remove unvoiced segments in audios, with its settings of a n
aggressiveness level of 3, frame duration 30ms, and padding
duration 150ms. These settings remove silences at the start , end,
and middle of speech. However, the amount of silence removed
does not distort emotional expression. All audios are resam pled
to sampling rate 22,050Hz. Mel spectrograms are computed
through a short-time Fourier transform (STFT) using frame s ize
1024, hop size 256, window size 1024, and a Hann window
function. The STFT magnitudes are transformed to the libros a
Slaney mel scale using an 80-channel mel filterbank spanning
0Hz to 8kHz, and the results are then clipped to a minimum
value of10−5, followed by log dynamic range compression.
Every Korean character in an input sentence is decomposed
into 3 elements: an onset, nucleus, and coda. In total, 19 ons ets,
21 nuclei, and 28 codas including the empty coda are employed
as defined by Unicode. A sequence of these elements becomes
a grapheme sequence taken as input by our synthesizer.
1https://github.com/wiseman/py-webrtcvad3.2. Model
Our multi-speaker emotional TTS synthesizer takes 3 inputs —
the grapheme sequence of a Korean sentence, 1 of 10 speakers
(5 females, 5 males), and 1 of the 7 Ekman’s emotion classes. I t
then generates a waveform in which the speaker utters the inp ut
sentence with the given emotion. Our synthesizer consists o f 2
sub-models: Tacotron 2 [9], mapping a grapheme sequence to
a mel spectrogram, and WaveGlow [10], transforming the mel
spectrogram to a waveform. Tacotron 2 is an auto-regressive
sequence-to-sequence neural network with a location-sens itive
attention mechanism. WaveGlow is a flow-based generative
neural network without auto-regression. We adapted NVIDIA
Tacotron 2 and WaveGlow repositories2,3to synthesize speech
for multiple speakers and emotions. The WaveGlow model
was utilized without modification but the Tacotron 2 model wa s
modified as outlined in the following paragraph.
Speaker identity is represented as a 5-dimensional train-
able speaker vector . Emotion identity is represented as a 3-
dimensional trainable emotion vector , except for the neutral
emotion vector, which is a non-trainable zero vector. To syn -
thesize speech of a given speaker and emotion, in the decoder
of Tacotron 2, speaker and emotion vectors are concatenated to
attention context vectors taken by the first and second LSTM
layers and the linear layer estimating a mel spectrogram.
3.3. Training
Tacotron 2 was trained with a batch size of 64 equally dis-
tributed to 4 GPUs. The Adam optimizer [11] of the default
settings (β1= 0.9,β2= 0.999,ǫ= 10−6) was used with a
learning rate of 10−3andL2regularization with weight 10−6.
If the norm of gradients exceeded 1, their norm was normalize d
to 1 to ensure stable learning.
Using a curriculum learning [5] strategy, Tacotron 2 was
trained to learn single-speaker neutral speech, multi-spe aker
neutral speech, and multi-speaker emotional speech in this or-
der. More specifically, the model was trained with the KSS
dataset for 20,000 iterations, then additionally with all d atasets
of neutral speech for 30,000 iterations, and finally with all train-
ing datasets for 65,000 iterations. Transitioning to train ing on
the next dataset was done when the model stably pronounced
given whole sentences for all training speaker-emotion pai rs.
In each training stage, we oversampled [6] the training set
with respect to speaker-emotion pairs, which means samples of
each speaker-emotion pair appear in a mini-batch with equal
probability. For example, samples of (ketts-30f, neutral) and
those of (ketts2-20m, happy) appear in a mini-batch with equ al
probability. This helped overcome difficulty in learning to syn-
thesize speech of speaker-emotion pairs with relatively sc arce
samples.
WaveGlow was trained with a batch size of 24, equally dis-
tributed to 3 GPUs using 24 clips of 16,000 mel spectrogram
frames randomly chosen from each training sample. Training
samples shorter than 16,000 mel frames were excluded from th e
training set since these samples padded with zeros caused un sta-
ble learning such as exploding gradients. Similar to Tacotr on 2,
we oversampled the training set with respect to speaker-emo tion
pairs. The Adam optimizer was used with the default settings
and learning rate 10−4. Weight normalization was applied, as
described in the original paper [10]. To ensure stable learn ing,
if the norm of gradients exceeded 1, their norm was normalize d
2https://github.com/NVIDIA/tacotron2
3https://github.com/NVIDIA/waveglowto 1. The model was initialized with the pretrained weights4of-
fered in the WaveGlow repository. The network was trained fo r
400,000 iterations until its loss curve formed a plateau. Th ez
elements were sampled from Gaussians with standard deviati on
1 during training and 0.75 during inference.
4. Results and Discussion
Through this procedure, our speech synthesizer is able to sy n-
thesize speech for all available 10 speakers and 7 emotions. Un-
expectedly, disgusted and surprised expressions of the KET TS3
speakers can be synthesized even without training supervis ion.
Synthesized speech samples can be found on this web page5.
Although our model expresses speaker and emotion iden-
tities, there are some minor inconsistencies in the quality of
synthesized samples across speakers and emotions. Thus, in
production, it is reasonable to fine-tune for each speaker an d
respectively preserve the model parameters.
Our silence removal settings substantially accelerated th e
learning of Tacotron 2. This was probably because silence re -
moval at the start, end, and middle of speech resulted in the
linear relationship between text and speech, and this relat ion-
ship helped the location-sensitive attention network easi ly learn
text-to-speech alignments.
5. Acknowledgements
This work was supported by Ministry of Culture, Sports and
Tourism andKorea Creative Content Agency [R2019020013,
R2020040298].
6. References
[1] Y . Lee, A. Rabiee, and S.-Y . Lee, “Emotional end-to-end n eural
speech synthesizer,” ArXiv , vol. abs/1711.05447, 2017.
[2] H. Choi, S. Park, J. Park, and M. Hahn, “Multi-speaker emo tional
acoustic modeling for CNN-based speech synthesis,” in ICASSP ,
2019.
[3] S.-Y . Um, S. Oh, K. Byun, I. Jang, C. Ahn, and H.-G. Kang,
“Emotional speech synthesis with rich and granularized con trol,”
inICASSP , 2020.
[4] T.-H. Kim, S. Cho, S. Choi, S. Park, and S.-Y . Lee, “Emotio nal
voice conversion using multitask learning with text-to-sp eech,” in
ICASSP , 2020.
[5] Y . Bengio, J. Louradour, R. Collobert, and J. Weston, “Cu rriculum
learning,” in ICML , 2009.
[6] M. Buda, A. Maki, and M. A. Mazurowski, “A systematic stud y
of the class imbalance problem in convolutional neural netw orks,”
Neural Networks , vol. 106, 2018.
[7] K. Park, “KSS dataset: Korean single speaker speech data set,”
https://kaggle.com/bryanpark/korean-single-speaker- speech-dataset,
2018.
[8] P. Ekman and D. Cordaro, “What is meant by calling emotion s
basic,” Emotion Review , vol. 3, no. 4, 2011.
[9] J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Ya ng,
Z. Chen, Y . Zhang, Y . Wang, R.-S. Ryan, R. A. Saurous,
Y . Agiomyrgiannakis, and Y . Wu, “Natural TTS synthesis by co n-
ditioning wavenet on MEL spectrogram predictions,” in ICASSP ,
2018.
[10] R. Prenger, R. Valle, and B. Catanzaro, “Waveglow: A flow -based
generative network for speech synthesis,” in ICASSP , 2019.
[11] D. P. Kingma and J. Ba, “Adam: A method for stochastic opt i-
mization,” in ICLR , 2015.
4“waveglow_256channels_universal_v5.pt” was used.
5https://sungjae-cho.github.io/InterSpeech2021_STDem o/