Neural Voice Cloning with a Few Samples

Sercan Ö. Arık∗ Jitong Chen∗ Kainan Peng∗

[email protected] [email protected] [email protected]

Wei Ping∗ Yanqi Zhou

[email protected] [email protected]
arXiv:1802.06006v3 [cs.CL] 12 Oct 2018

Baidu Research
1195 Bordeaux Dr. Sunnyvale, CA 94089

Abstract
Voice cloning is a highly desired feature for personalized speech interfaces. We
introduce a neural voice cloning system that learns to synthesize a person’s voice
from only a few audio samples. We study two approaches: speaker adaptation
and speaker encoding. Speaker adaptation is based on fine-tuning a multi-speaker
generative model. Speaker encoding is based on training a separate model to
directly infer a new speaker embedding, which will be applied to a multi-speaker
generative model. In terms of naturalness of the speech and similarity to the original
speaker, both approaches can achieve good performance, even with a few cloning
audios. 2 While speaker adaptation can achieve slightly better naturalness and
similarity, cloning time and required memory for the speaker encoding approach
are significantly less, making it more favorable for low-resource deployment.

1 Introduction
Generative models based on deep neural networks have been successfully applied to many domains
such as image generation [e.g., Oord et al., 2016b, Karras et al., 2017], speech synthesis [e.g., Oord
et al., 2016a, Arik et al., 2017a, Wang et al., 2017], and language modeling [e.g., Jozefowicz et al.,
2016]. Deep neural networks are capable of modeling complex data distributions and can be further
conditioned on external inputs to control the content and style of generated samples.
In speech synthesis, generative models can be conditioned on text and speaker identity [e.g., Arik
et al., 2017b]. While text carries linguistic information and controls the content of the generated
speech, speaker identity captures characteristics such as pitch, speech rate and accent. One approach
for multi-speaker speech synthesis is to jointly train a generative model and speaker embeddings
on triplets of text, audio and speaker identity [e.g., Ping et al., 2018]. The idea is to encode the
speaker-dependent information with low-dimensional embeddings, while sharing the majority of the
model parameters across all speakers. One limitation of such methods is that they can only generate
speech for observed speakers during training. An intriguing task is to learn the voice of an unseen
speaker from a few speech samples, a.k.a. voice cloning, which corresponds to few-shot generative
modeling of speech conditioned on the speaker identity. While a generative model can be trained
from scratch with a large amount of audio samples 3 , we focus on voice cloning of a new speaker
with a few minutes or even few seconds data. It is challenging as the model has to learn the speaker
characteristics from very limited amount of data, and still generalize to unseen texts.
∗
Equal contribution
2
Cloned audio samples can be found in https://audiodemos.github.io
3
A single speaker model can require ∼20 hours of training data [e.g., Arik et al., 2017a, Wang et al., 2017],
while a multi-speaker model for 108 speakers [Arik et al., 2017b] requires about ∼20 minutes data per speaker.

32nd Conference on Neural Information Processing Systems (NIPS 2018), Montréal, Canada.
In this paper, we investigate voice cloning in sequence-to-sequence neural speech synthesis sys-
tems [Ping et al., 2018]. Our contributions are the following:

1. We demonstrate and analyze the strength of speaker adaption approaches for voice cloning,
based on fine-tuning a pre-trained multi-speaker model for an unseen speaker using a few
samples.
2. We propose a novel speaker encoding approach, which provides comparable naturalness
and similarity in subjective evaluations while yielding significantly less cloning time and
computational resource requirements.
3. We propose automated evaluation methods for voice cloning based on neural speaker classifi-
cation and speaker verification.
4. We demonstrate voice morphing for gender and accent transformation via embedding manip-
ulations.

2 Related Work
Our work builds upon the state-of-the-art in neural speech synthesis and few-shot generative modeling.
Neural speech synthesis: Recently, there is a surge of interest in speech synthesis with neu-
ral networks, including Deep Voice 1 [Arik et al., 2017a], Deep Voice 2 [Arik et al., 2017b],
Deep Voice 3 [Ping et al., 2018], WaveNet [Oord et al., 2016a], SampleRNN [Mehri et al., 2016],
Char2Wav [Sotelo et al., 2017], Tacotron [Wang et al., 2017] and VoiceLoop [Taigman et al., 2018].
Among these methods, sequence-to-sequence models [Ping et al., 2018, Wang et al., 2017, Sotelo
et al., 2017] with attention mechanism have much simpler pipeline and can produce more natural
speech [e.g., Shen et al., 2017]. In this work, we use Deep Voice 3 as the baseline multi-speaker
model, because of its simple convolutional architecture and high efficiency for training and fast model
adaptation. It should be noted that our techniques can be seamlessly applied to other neural speech
synthesis models.
Few-shot generative modeling: Humans can learn new generative tasks from only a few examples,
which motivates research on few-shot generative models. Early studies mostly focus on Bayesian
methods. For example, hierarchical Bayesian models are used to exploit compositionality and
causality for few-shot generation of characters [Lake et al., 2013, 2015] and words in speech [Lake
et al., 2014]. Recently, deep neural networks achieve great successes in few-shot density estimation
and conditional image generation [e.g., Rezende et al., 2016, Reed et al., 2017, Azadi et al., 2017],
because of the great potential for composition in their learned representation. In this work, we
investigate few-shot generative modeling of speech conditioned on a particular speaker. We train a
separate speaker encoding network to directly predict the parameters of multi-speaker generative
model by only taking unsubscribed audio samples as inputs.
Speaker-dependent speech processing: Speaker-dependent modeling has been widely studied for
automatic speech recognition (ASR), with the goal of improving the performance by exploiting
speaker characteristics. In particular, there are two groups of methods in neural ASR, in alignment
with our two voice cloning approaches. The first group is speaker adaptation for the whole-model [Yu
et al., 2013], a portion of the model [Miao and Metze, 2015, Cui et al., 2017], or merely to a speaker
embedding [Abdel-Hamid and Jiang, 2013, Xue et al., 2014]. Speaker adaptation for voice cloning is
in the same vein as these approaches, but differences arise when text-to-speech vs. speech-to-text
are considered [Yamagishi et al., 2009]. The second group is based on training ASR models jointly
with embeddings. Extraction of the embeddings can be based on i-vectors [Miao et al., 2015], or
bottleneck layers of neural networks trained with a classification loss [Li and Wu, 2015]. Although
the general idea of speaker encoding is also based on extracting the embeddings directly, as a
major distinction, our speaker encoder models are trained with an objective function that is directly
related to speech synthesis. Lastly, speaker-dependent modeling is essential for multi-speaker speech
synthesis. Using i-vectors to represent speaker-dependent characteristics is one approach [Wu et al.,
2015], however, they have the limitation of being separately trained, with an objective that is not
directly related to speech synthesis. Also they may not be accurately extracted with small amount of
audio [Miao et al., 2015]. Another approach for multi-speaker speech synthesis is using trainable
speaker embeddings [Arik et al., 2017b], which are randomly initialized and jointly optimized from a
generative loss function.

2
Voice conversion: A closely related task of voice cloning is voice conversion. The goal of voice
conversion is to modify an utterance from source speaker to make it sound like the target speaker,
while keeping the linguistic contents unchanged. Unlike voice cloning, voice conversion systems
do not need to generalize to unseen texts. One common approach is dynamic frequency warping,
to align spectra of different speakers. Agiomyrgiannakis and Roupakia [2016] proposes a dynamic
programming algorithm that simultaneously estimates the optimal frequency warping and weighting
transform while matching source and target speakers using a matching-minimization algorithm. Wu
et al. [2016] uses a spectral conversion approach integrated with the locally linear embeddings for
manifold learning. There are also approaches to model spectral conversion using neural networks
[Desai et al., 2010, Chen et al., 2014, Hwang et al., 2015]. Those models are typically trained with a
large amount of audio pairs of target and source speakers.

3 From Multi-Speaker Generative Modeling to Voice Cloning

We consider a multi-speaker generative model, f (ti,j , si ; W, esi ), which takes a text ti,j and a
speaker identity si . The trainable parameters in the model is parameterized by W , and esi . The
latter denotes the trainable speaker embedding corresponding to si . Both W and esi are optimized
by minimizing a loss function L that penalizes the difference between generated and ground-truth
audios (e.g. a regression loss for spectrograms):
min E si ∼S, {L (f (ti,j , si ; W, esi ), ai,j )} (1)
W,e (ti,j ,ai,j )∼Tsi
where S is a set of speakers, Tsi is a training set of text-audio pairs for speaker si , and ai,j is the
ground-truth audio for ti,j of speaker si . The expectation is estimated over text-audio pairs of all
training speakers. We use W c and b e to denote the trained parameters and embeddings. Speaker
embeddings have been shown to effectively capture speaker characteristics with low-dimensional vec-
tors [Arik et al., 2017b, Ping et al., 2018]. Despite training with only a generative loss, discriminative
properties (e.g. gender and accent) are observed in the speaker embedding space [Arik et al., 2017b].
For voice cloning, we extract the speaker characteristics for an unseen speaker sk from a set of cloning
audios Ask , and generate an audio given any text for that speaker. The two performance metrics for
the generated audio are speech naturalness and speaker similarity (i.e., whether the generated audio
sounds like it is pronounced by the target speaker). The two approaches for neural voice cloning are
summarized in Fig. 1 and explained in the following sections.

3.1 Speaker adaptation

The idea of speaker adaptation is to fine-tune a trained multi-speaker model for an unseen speaker
using a few audio-text pairs. Fine-tuning can be applied to either the speaker embedding [Taigman
et al., 2018] or the whole model. For embedding-only adaptation, we have the following objective:
n  o
min E(tk,j ,ak,j )∼Tsk L f (tk,j , sk ; W
c , es ), ak,j
k
, (2)
esk

where Tsk is a set of text-audio pairs for the target speaker sk . For whole model adaptation, we have
the following objective:
min E(tk,j ,ak,j )∼Tsk {L (f (tk,j , sk ; W, esk ), ak,j )} . (3)
W,esk

Although the entire model provides more degrees of freedom for speaker adaptation, its optimization
is challenging for small amount of cloning data. Early stopping is required to avoid overfitting.

3.2 Speaker encoding

We propose a speaker encoding method to directly estimate the speaker embedding from audio
samples of an unseen speaker. Such a model does not require any fine-tuning during voice cloning.
Thus, the same model can be used for all unseen speakers. The speaker encoder, g(Ask ; Θ), takes a
set of cloning audio samples Ask and estimates esk for speaker sk . The model is parametrized by
Θ. Ideally, the speaker encoder can be jointly trained with the multi-speaker generative model from
scratch, with a loss function defined for the generated audio:
min E si ∼S, {L (f (ti,j , si ; W, g(Asi ; Θ)), ai,j )} . (4)
W,Θ (ti,j ,ai,j )∼Tsi

3
 Trainable Speaker encoding
Speaker adaptation
Fixed

Text Multi-speaker Audio

Speaker embedding generative model

Training
Multi-speaker Text Audio
Audio Multi-speaker Speaker encoder Speaker embedding
Text
generative model generative model model
Speaker encoder
Speaker embedding Speaker embedding model
Cloning audio

Cloning audio

Cloning Multi-speaker Cloning

text generative model audio
Speaker embedding
Cloning

Speaker embedding

or Speaker encoder
model

Cloning Multi-speaker Cloning

Cloning audio
text generative model audio

Speaker embedding
Audio generation

Multi-speaker Multi-speaker
Text Audio Text Audio
generative model generative model

Speaker embedding Speaker embedding

Figure 1: Illustration of speaker adaptation and speaker encoding approaches for voice cloning.

Note that the speaker encoder is trained with the speakers for the multi-speaker generative model.
During training, a set of cloning audio samples Asi are randomly sampled for training speaker si .
During inference, audio samples from the target speaker sk , Ask , are used to compute g(Ask ; Θ).
We observed optimization challenges with joint training from scratch: the speaker encoder tends
to estimate an average voice to minimize the overall generative loss. One possible solution is to
introduce discriminative loss functions for intermediate embeddings4 or generated audios5 . In our
case, however, such approaches only slightly improve speaker differences. Instead, we propose a
separate training procedure for speaker encoder. Speaker embeddings b esi are extracted from a trained
multi-speaker generative model f (ti,j , si ; W, esi ). Then, the speaker encoder g(Ask ; Θ) is trained
with an L1 loss to predict the embeddings from sampled cloning audios:
min Esi ∼S {|g(Asi ; Θ) − b
esi )|} . (5)
Θ

Eventually, entire model can be jointly fine-tuned following Eq. 4, with pre-trained W
c and Θb as the
initialization. Fine-tuning encourages the generative model to compensate for embedding estimation
errors, and may reduce attention problems. However, generative loss still dominates learning and
speaker differences in generated audios may be slightly reduced as well (see Section 4.3 for details).

Mel Speaker
Prenet Conv. block Mean pooling Self attention
spectrograms embedding
× Nconv

Figure 2: Speaker encoder architecture. See Appendix A for details.

For speaker encoder g(Ask ; Θ), we propose a neural network architecture comprising three parts (as
shown in Fig. 2):
(i) Spectral processing: We input mel-spectrograms of cloning audio samples to prenet, which
contains fully-connected layers with exponential linear unit for feature transformation.
(ii) Temporal processing: To utilize long-term context, we use convolutional layers with gated linear
unit and residual connections, average pooling is applied to summarize the whole utterance.
4
We have experimented classification loss by mapping the embeddings to labels via a softmax layer.
5
We have experimented integrating a pre-trained classifier to encourage discrimination in generated audios.

4
(iii) Cloning sample attention: Considering that different cloning audios contain different amount of
speaker information, we use a multi-head self-attention mechanism [Vaswani et al., 2017] to
compute the weights for different audios and get aggregated embeddings.

3.3 Discriminative models for evaluation

Besides human evaluations, we propose two evaluation methods using discriminative models for
voice cloning performance.

3.3.1 Speaker classification

Speaker classifier determines which speaker an audio sample belongs to. For voice cloning evaluation,
a speaker classifier is trained with the set of speakers used for cloning. High-quality voice cloning
would result in high classification accuracy. The architecture is composed of similar spectral and
temporal processing layers in Fig. 6 and an additional embedding layer before the softmax function.

3.3.2 Speaker verification

Speaker verification is the task of authenticating the claimed identity of a speaker, based on a test
audio and enrolled audios from the speaker. In particular, it performs binary classification to identify
whether the test audio and enrolled audios are from the same speaker [e.g., Snyder et al., 2016].
We consider an end-to-end text-independent speaker verification model [Snyder et al., 2016] (see
Appendix C for more details of model architecture). The speaker verification model can be trained on
a multi-speaker dataset, and then used to verify if the cloned audio and the ground-truth audio are
from the same speaker. Unlike the speaker classification approach, speaker verification model does
not require training with the audios from the target speaker for cloning, hence it can be used for unseen
speakers with a few samples. As the quantitative performance metric, the equal error-rate (EER) 6
can be used to measure how close the cloned audios are to the ground truth audios.

4 Experiments
4.1 Datasets

In our first set of experiments (Sections 4.3 and 4.4), the multi-speaker generative model and speaker
encoder are trained using LibriSpeech dataset [Panayotov et al., 2015], which contains audios
(16 KHz) for 2484 speakers, totalling 820 hours. LibriSpeech is a dataset for automatic speech
recognition, and its audio quality is lower compared to speech synthesis datasets.7 Voice cloning
is performed on VCTK dataset [Veaux et al., 2017]. VCTK consists of audios sampled at 48 KHz
for 108 native speakers of English with various accents. To be consistent with LibriSpeech dataset,
VCTK audios are downsampled to 16 KHz. For a chosen speaker, a few cloning audios are randomly
sampled for each experiment. The sentences presented in Appendix B are used to generate audios for
evaluation. In our second set of experiments (Section 4.5), we aim to investigate the impact of the
training dataset. We split the VCTK dataset for training and testing: 84 speakers are used for training
the multi-speaker model, 8 speakers for validation, and 16 speakers for cloning.

4.2 Model specifications

Our multi-speaker generative model is based on the convolutional sequence-to-sequence architecture

proposed in Ping et al. [2018], with similar hyperparameters and Griffin-Lim vocoder. To get better
performance, we increase the time-resolution by reducing the hop length and window size parameters
to 300 and 1200, and add a quadratic loss term to penalize large amplitude components superlinearly.
For speaker adaptation experiments, we reduce the embedding dimensionality to 128, as it yields
less overfitting problems. Overall, the baseline multi-speaker generative model has around 25M
trainable parameters when trained for the LibriSpeech dataset. For the second set of experiments,
hyperparameters of the VCTK model is used from Ping et al. [2018] to train a multi-speaker model
for the 84 speakers of VCTK, with Griffin-Lim vocoder.
6
EER is the point when the false acceptance rate and false rejection rate are equal.
7
We designed a segmentation and denoising pipeline to process LibriSpeech, as in Ping et al. [2018].

5
We train speaker encoders for different number of cloning audios separately. Initially, cloning audios
are converted to log-mel spectrograms with 80 frequency bands, with a hop length of 400, a window
size of 1600. Log-mel spectrograms are fed to spectral processing layers, which are composed of
2-layer prenet of size 128. Then, temporal processing is applied with two 1-D convolutional layers
with a filter width of 12. Finally, multi-head attention is applied with 2 heads and a unit size of 128 for
keys, queries and values. The final embedding size is 512. Validation set consists 25 held-out speakers.
A batch size of 64 is used, with an initial learning rate of 0.0006 with annealing rate of 0.6 applied
every 8000 iterations. Mean absolute error for the validation set is shown in Fig. 11 in Appendix
D. More cloning audios leads to more accurate speaker embedding estimation, especially with the
attention mechanism (see Appendix D for more details about the learned attention coefficients).
We train a speaker classifier using VCTK dataset to classify which of the 108 speakers an audio
sample belongs to. Speaker classifier has a fully-connected layer of size 256, 6 convolutional layers
with 256 filters of width 4, and a final embedding layer of size 32. The model achieves 100% accuracy
for validation set of size 512.
We train a speaker verification model using LibriSpeech dataset. Validation sets consists 50 held-out
speakers from Librispeech. EERs are estimated by randomly pairing up utterances from the same
or different speakers (50% for each case) in test set. We perform 40960 trials for each test set. We
describe the details of speaker verification model in Appendix C.

4.3 Voice cloning performance

Speaker adaptation Speaker encoding

Approaches Embedding-only Whole-model Without fine-tuning With fine-tuning
Data Text and audio Audio
Cloning time ∼ 8 hours ∼ 0.5 − 5 mins ∼ 1.5 − 3.5 secs ∼ 1.5 − 3.5 secs
Inference time ∼ 0.4 − 0.6 secs
Parameters per speaker 128 ∼ 25 million 512 512
Table 1: Comparison of speaker adaptation and speaker encoding approaches.

For speaker adaptation approach, we pick the optimal number of iterations using speaker classification
accuracy. For speaker encoding, we consider voice cloning with and without joint fine-tuning of the
speaker encoder and multi-speaker generative model.8 Table 1 summarizes the approaches and lists
the requirements for training, data, cloning time and memory footprint.

Figure 3: Performance of whole-model adaptation and speaker embedding adaptation for voice
cloning in terms of speaker classification accuracy for 108 VCTK speakers.

For speaker adaptation, Fig. 3 shows the speaker classification accuracy vs. the number of iterations.
For both, the classification accuracy significantly increases with more samples, up to ten samples. In
the low sample count regime, adapting the speaker embedding is less likely to overfit the samples
8
The learning rate and annealing parameters are optimized for joint fine-tuning.

6
 LibriSpeech (unseen speakers)
14 VCTK
speaker adaption: embedding-only
speaker adaption: whole-model
12 speaker encoding: without fine-tuning
speaker encoding: with fine-tuning

Equal error rate (in %)

10

0
1 2 3 5 10 20 50 100
Number of samples

(a) (b)
Figure 4: (a) Speaker classification accuracy with different numbers of cloning samples. (b) EER
(using 5 enrollment audios) for different numbers of cloning samples. LibriSpeech (unseen speakers)
and VCTK represent EERs estimated from random pairing of utterances from ground-truth datasets.

than adapting the whole model. The two methods also require different numbers of iterations to
converge. Compared to whole-model adaptation (which converges around 1000 iterations for even
100 cloning audio samples), embedding adaptation takes significantly more iterations to converge,
thus it results in much longer cloning time.
Figs. 4a and 4b show the classification accuracy and EER, obtained by speaker classification and
speaker verification models. Both speaker adaptation and speaker encoding benefit from more cloning
audios. When the number of cloning audio samples exceed five, whole-model adaptation outperforms
other techniques. Speaker encoding yields a lower classification accuracy compared to embedding
adaptation, but they achieve a similar speaker verification performance.
Besides evaluations by discriminative models, we conduct subject tests on Amazon Mechanical
Turk framework. For assessment of the naturalness, we use the 5-scale mean opinion score (MOS).
For assessment of how similar the generated audios are to the ground-truth audios from target
speakers, we use the 4-scale similarity score with the question and categories in [Wester et al., 2016].9
Tables 2 and 3 show the results of human evaluations. Higher number of cloning audios improve
both metrics. The improvement is more significant for whole model adaptation, due to the more
degrees of freedom provided for an unseen speaker. Indeed, for high sample counts, the naturalness
significantly exceeds the baseline model, due to the dominance of better quality adaptation samples
over training data. Speaker encoding achieves naturalness similar or better than the baseline model.
The naturalness is even further improved with fine-tuning since it allows the generative model to
learn how to compensate for the errors of the speaker encoder. Similarity scores slightly improve with
higher sample counts for speaker encoding, and match the scores for speaker embedding adaptation.

Approach Sample count

1 2 3 5 10
Ground-truth (16 KHz sampling rate) 4.66±0.06
Multi-speaker generative model 2.61±0.10
Speaker adaptation (embedding-only) 2.27±0.10 2.38±0.10 2.43±0.10 2.46±0.09 2.67±0.10
Speaker adaptation (whole-model) 2.32±0.10 2.87±0.09 2.98±0.11 2.67±0.11 3.16±0.09
Speaker encoding (without fine-tuning) 2.76±0.10 2.76±0.09 2.78±0.10 2.75±0.10 2.79±0.10
Speaker encoding (with fine-tuning) 2.93±0.10 3.02±0.11 2.97±0.1 2.93±0.10 2.99±0.12
Table 2: Mean Opinion Score (MOS) evaluations for naturalness with 95% confidence intervals
(training with LibriSpeech speakers and cloning with 108 VCTK speakers).

9
We conduct each evaluation independently, so the cloned audios of two different models are not directly
compared during rating. Multiple votes on the same sample are aggregated by a majority voting rule.

7
 Approach Sample count
1 2 3 5 10
Ground-truth (same speaker) 3.91±0.03
Ground-truth (different speakers) 1.52±0.09
Speaker adaptation (embedding-only) 2.66±0.09 2.64±0.09 2.71±0.09 2.78±0.10 2.95±0.09
Speaker adaptation (whole-model) 2.59±0.09 2.95±0.09 3.01±0.10 3.07±0.08 3.16±0.08
Speaker encoding (without fine-tuning) 2.48±0.10 2.73±0.10 2.70±0.11 2.81±0.10 2.85±0.10
Speaker encoding (with fine-tuning) 2.59±0.12 2.67±0.12 2.73±0.13 2.77±0.12 2.77±0.11
Table 3: Similarity score evaluations with 95% confidence intervals (training with LibriSpeech
speakers and cloning with 108 VCTK speakers).

4.4 Voice morphing via embedding manipulation

As shown in Fig. 5, speaker encoder maps speakers into a meaningful latent space. Inspired by word
embedding manipulation (e.g. to demonstrate the existence of simple algebraic operations as king
- queen = male - female), we apply algebraic operations to inferred embeddings to transform their
speech characteristics. To transform gender, we estimate the averaged speaker embeddings for each
gender, and add their difference to a particular speaker. For example, BritishMale + AveragedFemale -
AveragedMale yields a British female speaker. Similarly, we consider region of accent transformation
via BritishMale + AveragedAmerican - AveragedBritish to obtain an American male speaker. Our
results demonstrate high quality audios with specific gender and accent characteristics.10

0.3 0.3
Average North American male Average North American male
0.2 Average North American female 0.2 Average North American female
Average North American
0.1 0.1
Average male

0.0 Average female 0.0

-0.1 -0.1

-0.2 Average British male -0.2 Average British Average British male
Average British female Average British female

-0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6

Figure 5: Visualization of estimated speaker embeddings by speaker encoder. The first two principal
components of speaker embeddings (averaged across 5 samples for each speaker). Only British and
North American regional accents are shown as they constitute the majority of the labeled speakers in
the VCTK dataset. Please see Appendix E for more detailed analysis.

4.5 Impact of training dataset

Approach Sample count

1 5 10 20 100
Speaker adaptation (embedding-only) 3.01±0.11 - 3.13±0.11 - 3.13±0.11
Speaker adaptation (whole-model) 2.34±0.13 2.99±0.10 3.07±0.09 3.40±0.10 3.38±0.09
Table 4: Mean Opinion Score (MOS) evaluations for naturalness with 95% confidence intervals
(training with 84 VCTK speakers and cloning with 16 VCTK speakers).

Approach Sample count

1 5 10 20 100
Speaker adaptation (embedding-only) 2.42±0.13 - 2.37±0.13 - 2.37±0.12
Speaker adaptation (whole-model) 2.55±0.11 2.93±0.11 2.95±0.10 3.01±0.10 3.14±0.10
Table 5: Similarity score evaluations with 95% confidence intervals (training with 84 VCTK speakers
and cloning with 16 VCTK speakers).

To evaluate the impact of the dataset, we consider training with a subset of the VCTK containing
84 speakers, and cloning on another 16 speakers. Tables 4 and 5 present the human evaluations
10
https://audiodemos.github.io/

8
for speaker adaptation.11 Speaker verification results are given in Appendix C. One the one hand,
compared to LibriSpeech, cleaner VCTK data improves the multi-speaker generative model, leading
to better whole-model adaptation results. On the other hand, embedding-only adaptation significantly
underperforms whole-model adaptation due to the limited speaker diversity in VCTK dataset.

5 Conclusions
We study two approaches for neural voice cloning: speaker adaptation and speaker encoding. We
demonstrate that both approaches can achieve good cloning quality even with only a few cloning
audios. For naturalness, we show that both speaker adaptation and speaker encoding can achieve
an MOS similar to the baseline multi-speaker generative model. Thus, the proposed techniques can
potentially be improved with better multi-speaker models in the future (such as replacing Griffin-
Lim with WaveNet vocoder). For similarity, we demonstrate that both approaches benefit from a
larger number of cloning audios. The performance gap between whole-model and embedding-only
adaptation indicates that some discriminative speaker information still exists in the generative model
besides speaker embeddings. The benefit of compact representation via embeddings is fast cloning
and small footprint per speaker. We observe drawbacks of training the multi-speaker generative
model using a speech recognition dataset with low-quality audios and limited speaker diversity.
Improvements in the quality of dataset would result in higher naturalness. We expect our techniques
to benefit significantly from a large-scale and high-quality multi-speaker dataset.

References
O. Abdel-Hamid and H. Jiang. Fast speaker adaptation of hybrid nn/hmm model for speech recogni-
tion based on discriminative learning of speaker code. In IEEE ICASSP, 2013.
Y. Agiomyrgiannakis and Z. Roupakia. Voice morphing that improves tts quality using an optimal
dynamic frequency warping-and-weighting transform. IEEE ICASSP, 2016.
D. Amodei, S. Ananthanarayanan, R. Anubhai, J. Bai, E. Battenberg, C. Case, J. Casper, B. Catanzaro,
Q. Cheng, G. Chen, et al. Deep speech 2: End-to-end speech recognition in english and mandarin.
In International Conference on Machine Learning, pages 173–182, 2016.
S. Ö. Arik, M. Chrzanowski, A. Coates, G. Diamos, A. Gibiansky, Y. Kang, X. Li, J. Miller, J. Raiman,
S. Sengupta, and M. Shoeybi. Deep Voice: Real-time neural text-to-speech. In ICML, 2017a.
S. Ö. Arik, G. F. Diamos, A. Gibiansky, J. Miller, K. Peng, W. Ping, J. Raiman, and Y. Zhou. Deep
Voice 2: Multi-speaker neural text-to-speech. In NIPS, pages 2966–2974, 2017b.
S. Azadi, M. Fisher, V. Kim, Z. Wang, E. Shechtman, and T. Darrell. Multi-content gan for few-shot
font style transfer. CoRR, abs/1708.02182, 2017.
L. H. Chen, Z. H. Ling, L. J. Liu, and L. R. Dai. Voice conversion using deep neural networks
with layer-wise generative training. IEEE/ACM Transactions on Audio, Speech, and Language
Processing, 2014.
X. Cui, V. Goel, and G. Saon. Embedding-based speaker adaptive training of deep neural networks.
arXiv preprint arXiv:1710.06937, 2017.
S. Desai, A. W. Black, B. Yegnanarayana, and K. Prahallad. Spectral mapping using artificial neural
networks for voice conversion. IEEE Transactions on Audio, Speech, and Language Processing,
2010.
H. T. Hwang, Y. Tsao, H. M. Wang, Y. R. Wang, and S. H. Chen. A probabilistic interpretation for
artificial neural network-based voice conversion. In 2015 Asia-Pacific Signal and Information
Processing Association Annual Summit and Conference (APSIPA), 2015.
R. Jozefowicz, O. Vinyals, M. Schuster, N. Shazeer, and Y. Wu. Exploring the limits of language
modeling. arXiv preprint arXiv:1602.02410, 2016.
11
The speaker encoder models generalize poorly for unseen speakers due to limited training speakers.

9
T. Karras, T. Aila, S. Laine, and J. Lehtinen. Progressive growing of gans for improved quality,
stability, and variation. CoRR, abs/1710.10196, 2017.
B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum. One-shot learning by inverting a compositional
causal process. In NIPS, 2013.
B. M. Lake, C. ying Lee, J. R. Glass, and J. B. Tenenbaum. One-shot learning of generative speech
concepts. In CogSci, 2014.
B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum. Human-level concept learning through proba-
bilistic program induction. Science, 2015.
X. Li and X. Wu. Modeling speaker variability using long short-term memory networks for speech
recognition. In INTERSPEECH, 2015.
S. Mehri, K. Kumar, I. Gulrajani, R. Kumar, S. Jain, J. Sotelo, A. Courville, and Y. Bengio. Samplernn:
An unconditional end-to-end neural audio generation model. arXiv preprint arXiv:1612.07837,
2016.
Y. Miao and F. Metze. On speaker adaptation of long short-term memory recurrent neural networks.
In Sixteenth Annual Conference of the International Speech Communication Association, 2015.
Y. Miao, H. Zhang, and F. Metze. Speaker adaptive training of deep neural network acoustic models
using i-vectors. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2015.
A. v. d. Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner,
A. Senior, and K. Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint
arXiv:1609.03499, 2016a.
A. v. d. Oord, N. Kalchbrenner, L. Espeholt, O. Vinyals, A. Graves, et al. Conditional image
generation with pixelcnn decoders. In Advances in Neural Information Processing Systems, 2016b.
V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an ASR corpus based on public
domain audio books. In IEEE ICASSP, 2015.
W. Ping, K. Peng, A. Gibiansky, S. Arik, A. Kannan, S. Narang, J. Raiman, and J. Miller. Deep Voice
3: Scaling text-to-speech with convolutional sequence learning. In ICLR, 2018.
S. Prince and J. Elder. Probabilistic linear discriminant analysis for inferences about identity. In
ICCV, 2007.
S. E. Reed, Y. Chen, T. Paine, A. van den Oord, S. M. A. Eslami, D. J. Rezende, O. Vinyals, and
N. de Freitas. Few-shot autoregressive density estimation: Towards learning to learn distributions.
CoRR, 2017.
D. Rezende, Shakir, I. Danihelka, K. Gregor, and D. Wierstra. One-shot generalization in deep
generative models. In ICML, 2016.
J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y. Zhang, Y. Wang, R. Skerry-
Ryan, et al. Natural tts synthesis by conditioning wavenet on mel spectrogram predictions. arXiv
preprint arXiv:1712.05884, 2017.
D. Snyder, P. Ghahremani, D. Povey, D. Garcia-Romero, Y. Carmiel, and S. Khudanpur. Deep neural
network-based speaker embeddings for end-to-end speaker verification. In IEEE Spoken Language
Technology Workshop (SLT), pages 165–170, 2016.
J. Sotelo, S. Mehri, K. Kumar, J. F. Santos, K. Kastner, A. Courville, and Y. Bengio. Char2wav:
End-to-end speech synthesis. 2017.
Y. Taigman, L. Wolf, A. Polyak, and E. Nachmani. Voiceloop: Voice fitting and synthesis via a
phonological loop. In ICLR, 2018.
A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polosukhin.
Attention is all you need. In NIPS. 2017.

10
C. Veaux, J. Yamagishi, and K. e. a. MacDonald. Cstr vctk corpus: English multi-speaker corpus for
cstr voice cloning toolkit, 2017.
Y. Wang, R. J. Skerry-Ryan, D. Stanton, Y. Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y. Xiao, Z. Chen,
S. Bengio, Q. V. Le, Y. Agiomyrgiannakis, R. Clark, and R. A. Saurous. Tacotron: A fully
end-to-end text-to-speech synthesis model. CoRR, abs/1703.10135, 2017.
M. Wester, Z. Wu, and J. Yamagishi. Analysis of the voice conversion challenge 2016 evaluation
results. In INTERSPEECH, pages 1637–1641, 09 2016.
Y.-C. Wu, H.-T. Hwang, C.-C. Hsu, Y. Tsao, and H.-m. Wang. Locally linear embedding for
exemplar-based spectral conversion. In INTERSPEECH, pages 1652–1656, 09 2016.
Z. Wu, P. Swietojanski, C. Veaux, S. Renals, and S. King. A study of speaker adaptation for dnn-based
speech synthesis. In INTERSPEECH, 2015.
S. Xue, O. Abdel-Hamid, H. Jiang, L. Dai, and Q. Liu. Fast adaptation of deep neural network based
on discriminant codes for speech recognition. IEEE/ACM Transactions on Audio, Speech, and
Language Processing, 2014.
J. Yamagishi, T. Kobayashi, Y. Nakano, K. Ogata, and J. Isogai. Analysis of speaker adaptation
algorithms for hmm-based speech synthesis and a constrained smaplr adaptation algorithm. IEEE
Transactions on Audio, Speech, and Language Processing, 2009.
D. Yu, K. Yao, H. Su, G. Li, and F. Seide. Kl-divergence regularized deep neural network adaptation
for improved large vocabulary speech recognition. In IEEE ICASSP, 2013.

11
Appendices
A Detailed speaker encoder architecture

Speaker embedding
[batch, dembedding]

[batch, Nsamples, dembedding]

×
[batch, Nsamples]

Normalize
[batch, Nsamples]

Cloning samples attention FC Softsign

[batch, Nsamples]

FC
[batch, Nsamples, dattn]

Multi-head attention

queries

values
keys
[batch, Nsamples, dattn]

ELU ELU ELU

FC FC FC

[batch, Nsamples, Fmapped]

Global mean pooling

[batch, Nsamples, T, Fmapped]

Temporal masking

[batch, Nsamples, T, Fmapped]

Temporal processing

√0.5
×

× Nconv
Gated linear unit

Convolution

[batch, Nsamples, T, Fmapped]

Spectral processing

ELU
× Nprenet
FC

[batch, Nsamples, T, Fmel]

Mel spectrograms

Figure 6: Speaker encoder architecture with intermediate state dimensions. (batch: batch size,
Nsamples : number of cloning audio samples |Ask |, T : number of mel spectrograms timeframes,
Fmel : number of mel frequency channels, Fmapped : number of frequency channels after prenet,
dembedding : speaker embedding dimension). Multiplication operation at the last layer represents
inner product along the dimension of cloning samples.

12
B Voice cloning test sentences

Prosecutors have opened a massive investigation/into allegations of/fixing games/and illegal betting%.
Different telescope designs/perform differently%and have different strengths/and weaknesses%.
We can continue to strengthen the education of good lawyers%.
Feedback must be timely/and accurate/throughout the project%.
Humans also judge distance/by using the relative sizes of objects%.
Churches should not encourage it%or make it look harmless%.
Learn about/setting up/wireless network configuration%.
You can eat them fresh cooked%or fermented%.
If this is true%then those/who tend to think creatively%really are somehow different%.
She will likely jump for joy%and want to skip straight to the honeymoon%.
The sugar syrup/should create very fine strands of sugar%that drape over the handles%.
But really in the grand scheme of things%this information is insignificant%.
I let the positive/overrule the negative%.
He wiped his brow/with his forearm%.
Instead of fixing it%they give it a nickname%.
About half the people%who are infected%also lose weight%.
The second half of the book%focuses on argument/and essay writing%.
We have the means/to help ourselves%.
The large items/are put into containers/for disposal%.
He loves to/watch me/drink this stuff%.
Still%it is an odd fashion choice%.
Funding is always an issue/after the fact%.
Let us/encourage each other%.

Figure 7: The sentences used to generate test samples for the voice cloning models. The white space
characters / and % follow the same definition as in Ping et al. [2018].

13
C Speaker verification model
Given a set of (e.g., 1∼5) enrollment audios 12 and a test audio, speaker verification model per-
forms a binary classification and tells whether the enrollment and test audios are from the same
speaker. Although using other speaker verification models [e.g., Snyder et al., 2016] would also
suffice, we choose to create our own speaker verification models using convolutional-recurrent
architecture [Amodei et al., 2016]. We note that our equal-error-rate results on test set of unseen
speakers are on par with the state-of-the-art speaker verification models. The architecture of our
model is illustrated in Figure 8. We compute mel-scaled spectrogram of enrollment audios and test
audio after resampling the input to a constant sampling frequency. Then, we apply two-dimensional
convolutional layers convolving over both time and frequency bands, with batch normalization and
ReLU non-linearity after each convolution layer. The output of last convolution layer is feed into
a recurrent layer (GRU). We then mean-pool over time (and enrollment audios if there are many),
then apply a fully connected layer to obtain the speaker encodings for both enrollment audios and
test audio. We use the probabilistic linear discriminant analysis (PLDA) for scoring the similarity
between the two encodings [Prince and Elder, 2007, Snyder et al., 2016]. The PLDA score [Snyder
et al., 2016] is defined as,
s(x, y) = w · x> y − x> Sx − y > Sy + b (6)
where x and y are speaker encodings of enrollment and test audios respectively after fully-connected
layer, w and b are scalar parameters, and S is a symmetric matrix. Then, s(x, y) is feed into a
sigmoid unit to obtain the probability that they are from the same speaker. The model is trained
using cross-entropy loss. Table 6 lists hyperparameters of speaker verification model for LibriSpeech
dataset.
In addition to speaker verification test results presented in main text (Figure 4b), we also include the
result using 1 enrollment audio when the multi-speaker generative model is trained on LibriSpeech.
When multi-speaker generative model is trained on VCTK, the results are in Figure 10. It should be
noted that, the EER on cloned audios could be potentially better than on ground truth VCTK, because
the speaker verification model is trained on LibriSpeech dataset.

same or different speaker

PLDA score

SV module FC

pooling

recurrent layer

Conv. xN
block ReLU

batch-nom

2D convolution

Mel spectrograms Mel spectrogram

enroll audios test audio

Figure 8: Architecture of speaker verification model.

12
Enrollment audios are from the same speaker.

14
 Parameter
Audio resampling freq. 16 KHz
Bands of Mel-spectrogram 80
Hop length 400
Convolution layers, channels, filter, strides 1, 64, 20 × 5, 8 × 2
Recurrent layer size 128
Fully connected size 128
Dropout probability 0.9
Learning Rate 10−3
Max gradient norm 100
Gradient clipping max. value 5

Table 6: Hyperparameters of speaker verification model for LibriSpeech dataset.

25 LibriSpeech (unseen speakers)

VCTK
speaker adaption: embedding-only
speaker adaption: whole-model
speaker encoding: without fine-tuning
20 speaker encoding: with fine-tuning
Equal error rate (in %)

15

10

0
1 2 3 5 10 20 50 100
Number of samples

Figure 9: Speaker verification EER (using 1 enrollment audio) vs. number of cloning audio samples.
Multi-speaker generative model and speaker verification model are trained using LibriSpeech dataset.
Voice cloing is performed using VCTK dataset.

35 LibriSpeech (unseen speakers) 35 LibriSpeech (unseen speakers)

VCTK VCTK
speaker adaption: embedding-only speaker adaption: embedding-only
30 speaker adaption: whole-model 30 speaker adaption: whole-model

25 25
Equal error rate (in %)

20 20

15 15

10 10

5 5

0 0
1 5 10 20 50 100 1 5 10 20 50 100
Number of samples

(a) (b)
Figure 10: Speaker verification EER using (a) 1 enrollment audio (b) 5 enrollment audios vs. number
of cloning audio samples. Multi-speaker generative model is trained on a subset of VCTK dataset
including 84 speakers, and voice cloning is performed on other 16 speakers. Speaker verification
model is trained using the LibriSpeech dataset.

15
D Implications of attention
For a trained speaker encoder model, Fig. 12 exemplifies attention distributions for different audio
lengths. The attention mechanism can yield highly non-uniformly distributed coefficients while
combining the information in different cloning samples, and especially assigns higher coefficients to
longer audios, as intuitively expected due to the potential more information content in them.

Figure 11: Mean absolute error in embedding estimation vs. the number of cloning audios for a
validation set of 25 speakers, shown with the attention mechanism and without attention mechanism
(by simply averaging).

0.5 0.5 0.5 0.5

Coefficients

Coefficients

Coefficients

Coefficients
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Length (sec) Length (sec) Length (sec) Length (sec)

0.5 0.5 0.5 0.5

Coefficients

Coefficients

Coefficients

Coefficients
0.2 0.2 0.2 0.2
0.0 0.0 0.0 0.0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Length (sec) Length (sec) Length (sec) Length (sec)

0.5 0.5 0.5 0.5

Coefficients

Coefficients

Coefficients

Coefficients

0.2 0.2 0.2 0.2

0.0 0.0 0.0 0.0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Length (sec) Length (sec) Length (sec) Length (sec)

0.5 0.5 0.5 0.5

Coefficients

Coefficients

Coefficients

Coefficients

0.2 0.2 0.2 0.2

0.0 0.0 0.0 0.0
0 2 4 6 8 0 2 4 6 8 0 2 4 6 8 0 2 4 6 8
Length (sec) Length (sec) Length (sec) Length (sec)

Figure 12: Inferred attention coefficients for the speaker encoder model with Nsamples = 5 vs. lengths
of the cloning audio samples. The dashed line corresponds to the case of averaging all cloning audio
samples.

16
E Speaker embedding space learned by the encoder
To analyze the speaker embedding space learned by the trained speaker encoders, we apply principal
component analysis to the space of inferred embeddings and consider their ground truth labels for
gender and region of accent from the VCTK dataset. Fig. 13 shows visualization of the first two
principal components. We observe that speaker encoder maps the cloning audios to a latent space
with highly meaningful discriminative patterns. In particular for gender, a one dimensional linear
transformation from the learned speaker embeddings can achieve a very high discriminative accuracy
- although the models never see the ground truth gender label while training.

Female Sample count:1 Sample count:1 Great Britain

Male North American
Ireland
South Hemisphere
Asia

Sample count:2 Sample count:2

Sample count:3 Sample count:3

Sample count:5 Sample count:5

Sample count:10 Sample count:10

Figure 13: First two principal components of the inferred embeddings, with the ground truth labels
for gender and region of accent for the VCTK speakers as in Arik et al. [2017b].

17
F Similarity scores
For the result in Table 3, Fig. 14 shows the distribution of the scores given by MTurk users as in
[Wester et al., 2016]. For 10 sample count, the ratio of evaluations with the ‘same speaker’ rating
exceeds 70 % for all models.

Figure 14: Distribution of similarity scores for 1 and 10 sample counts.

18