S AMPLE E FFICIENT A DAPTIVE T EXT- TO -S PEECH

Yutian Chen, Yannis Assael, Brendan Shillingford, David Budden, Scott Reed,
Heiga Zen, Quan Wang, Luis C. Cobo, Andrew Trask, Ben Laurie,
Caglar Gulcehre, Aäron van den Oord, Oriol Vinyals, Nando de Freitas

DeepMind & Google

[email protected]
arXiv:1809.10460v1 [cs.LG] 27 Sep 2018

A BSTRACT

We present a meta-learning approach for adaptive text-to-speech (TTS) with few

data. During training, we learn a multi-speaker model using a shared conditional
WaveNet core and independent learned embeddings for each speaker. The aim
of training is not to produce a neural network with fixed weights, which is then
deployed as a TTS system. Instead, the aim is to produce a network that requires
few data at deployment time to rapidly adapt to new speakers. We introduce and
benchmark three strategies: (i) learning the speaker embedding while keeping the
WaveNet core fixed, (ii) fine-tuning the entire architecture with stochastic gradient
descent, and (iii) predicting the speaker embedding with a trained neural network
encoder. The experiments show that these approaches are successful at adapting
the multi-speaker neural network to new speakers, obtaining state-of-the-art results
in both sample naturalness and voice similarity with merely a few minutes of audio
data from new speakers.

1 I NTRODUCTION

Training a large model with lots of data and subsequently deploying this model to carry out clas-
sification or regression is an important and common methodology in machine learning. It has
been particularly successful in speech recognition (Hinton et al., 2012), machine translation (Wu
et al., 2016) and image recognition (Krizhevsky et al., 2012; Szegedy et al., 2015). In this text-
to-speech (TTS) work, we are instead interested in few-shot meta-learning. Here the objective of
training with many data is not to learn a fixed-parameter classifier, but rather to learn a “prior” neural
network. This prior TTS network can be adapted rapidly, using few data, to produce TTS systems for
new speakers at deployment time. That is, the intention is not to learn a fixed final model, but rather
to learn a model prior that harnesses few data at deployment time to learn new behaviours rapidly.
The output of training is not longer a fixed model, but rather a fast learner.
Biology provides motivation for this line of research. It may be argued that evolution is a slow
adaptation process that has resulted in biological machines with the ability to adapt rapidly to new
data during their lifetimes. These machines are born with strong priors that facilitate rapid learning.
We consider a meta-learning approach where the model has two types of parameters: task-dependent
parameters and task-independent parameters. During training, we learn all of these parameters but
discard the task-dependent parameters for deployment. The goal is to use few data to learn the
task-dependent parameters for new tasks rapidly.
Task-dependent parameters play a similar role to latent variables in classical probabilistic graphical
models. Intuitively, these variables introduce flexibility, thus making it easier to learn the task-
independent parameters. For example, in classical HMMs, knowing the latent variables results in a
simple learning problem of estimating the parameters of an exponential-family distribution. In neural
networks, this approach also facilitates learning when there is clear data diversity and categorization.
We show this for adaptive TTS (Dutoit, 1997; Taylor, 2009). In this setting, speakers correspond to
tasks. During training we have many speakers, and it is therefore helpful to have task-dependent
parameters to capture speaker-specific voice styles. At the same time, it is useful to have a large

1
 Embedding table WaveNet Stack
Output

Speaker index
Hidden Layer
e
Hidden Layer
Text l1:T
Linguistic features
Input
ff00 1:T
1:T t=1 t=2 … t=T
Fundamental freq

Waveform

Figure 1: Architecture of the WaveNet model for few-shot voice adaptation.

model with shared parameters to capture the generic process of mapping text to speech. To this end,
we employ the WaveNet model.
WaveNet (van den Oord et al., 2016) is an autoregressive generative model for audio waveforms that
has yielded state-of-art performance in speech synthesis. This model was later modified for real-time
speech generation via probability density distillation into a feed-forward model (van den Oord et al.,
2017). A fundamental limitation of WaveNet is the need for hours of training data for each speaker. In
this paper we describe a new WaveNet training procedure that facilitates adaptation to new speakers,
allowing the synthesis of new voices from no more than 10 minutes of data with high sample quality.
We propose several extensions of WaveNet for sample-efficient adaptive TTS. First, we present two
non-parametric adaptation methods that involve fine-tuning either the speaker embeddings only or
all the model parameters given few data from a new speaker. Second, we present a parametric text-
independent approach whereby an auxiliary network is trained to predict new speaker embeddings.
The experiments will show that all the proposed approaches, when provided with just a few seconds
or minutes of recording, can generate high-fidelity utterances that closely resemble the vocal tract
characteristics of a demonstration speaker, particularly when the entire model is fine-tuned end-to-end.
When fine-tuning by first estimating the speaker embedding and subsequently fine-tuning the entire
model, we achieve state-of-the-art results in terms of sample naturalness and voice similarity to target
speakers. These results are robust across speech datasets recorded under different conditions and,
moreover, we demonstrate that the generated samples are capable of confusing the state-of-the-art
text-independent speaker verification system (Wan et al., 2018).
TTS techniques require hours of high-quality recordings, collected in controlled environments, for
each new voice style. Given this high cost, reducing the length of the training dataset could be
valuable. For example, it is likely to be very beneficial when attempting to restore the voices of
patients who suffer from voice-impairing medical conditions. In these cases, long high quality
recordings are scarce.

2 WAVE N ET ARCHITECTURE

WaveNet is an autoregressive model that factorizes the joint probability distribution of a waveform,
x = {x1 , . . . , xT }, into a product of conditional distributions using the probabilistic chain rule:
T
Y
p(x|h; w) = p(xt |x1:t−1 , h; w),
t=1

where xt is the t-th timestep sample, and h and w are respectively the conditioning inputs and
parameters of the model. To train a multi-speaker WaveNet, the conditioning inputs h consist of the
speaker identity s, the linguistic features l, and the logarithmic fundamental frequency f0 values. l
encodes the sequence of phonemes derived from the input text, and f0 controls the dynamics of the
pitch in the generated utterance. Given the speaker identity s for each utterance in the dataset, the

2
 Log likelihood Log likelihood Sample Wave

Embedding vector WaveNet Stack Embedding vector WaveNet Stack Embedding vector WaveNet Stack
Text ID

Text
Linguistic features Linguistic features Linguistic features

Wave Text
Fundamental freq
Fundamental freq Fundamental freq Predictor
Wave

Wave
(a) Training (b) Adaptation (c) Inference

Figure 2: Training (slow, lots of data), adaptation (fast, few data) and inference stages for the SEA-
A LL architecture. The components with bold pink outlines are fine-tuned during the adaptation phase.
The purpose of training is to produce a prior. This prior is combined with few data during adaptation
to solve a new task. This adapted model is then deployed in the final inference stage.

model is expressed as:

T
Y
p(x|l, f0 ; es , w) = p(xt |x1:t−1 , l, f0 ; es , w),
t=1

where a table of speaker embedding vectors es (Embedding in Figure 1) is learned alongside the
standard WaveNet parameters. These vectors capture salient voice characteristics across individual
speakers, and provide a convenient mechanism for generalizing WaveNet to the few-shot adaptation
setting in this paper. The linguistic features l and fundamental frequency values f0 are both time-series
with a lower sampling frequency than the waveform. Thus, to be used as local conditioning variables
they are upsampled by a transposed convolutional network. During training, l and f0 are extracted by
signal processing methods from pairs of training utterance and transcript, and during testing, those
values are predicted from text by existing models (Zen et al., 2016).

3 F EW- SHOT ADAPTATION WITH WAVE N ET

In recent years, a large body of literature uses large datasets to train models to learn an input-output
mapping that is then used for inference. In contrast, few-shot meta-learning introduces an additional
step, adaptation. In this meta-learning setting, the purpose of training becomes to learn a prior. During
adaptation, this prior is combined with few data to rapidly learn a new skill; in this case adapting to a
new speakers’ voice style. Finally, the new skill is deployed, which in this paper we are referring to
as inference. These three stages — training, adaptation and inference — are illustrated in Figure 2.
We present two multi-speaker WaveNet extensions for few-shot voice adaptation. First, we introduce
a non-parametric model fine-tuning approach, which involves adapting either the speaker embeddings
or all the model parameters using held-aside demonstration data. Second, and for comparison
purposes, we use a parametric approach whereby an auxiliary network is trained to predict the
embedding vector of a new speaker using the demonstration data.

3.1 N ON - PARAMETRIC FEW- SHOT ADAPTATION VIA FINE - TUNING

Inspired by few-shot learning we first pre-train a multi-speaker conditional WaveNet model on a large
and diverse dataset, as described in Section 2. Subsequently, we fine-tune the model parameters by
retraining with respect to held-aside adaptation data.
Training this WaveNet model to maximize the conditional log-likelihood of the generated audio jointly
optimizes both the set of speaker parameters {es } and the shared WaveNet core parameters w. Next,
we extend this method to a new speaker by extracting the l and f0 features from their adaptation data
and randomly initializing a new embedding vector e. We then optimize e such that the demonstration
(1) (n) (1) (1) (n) (n)
waveforms, {xdemo , . . . , xdemo }, paired with features {(ldemo , f0,demo ), . . . , (ldemo , f0,demo )}, are likely
under the model with w fixed (SEA-E MB):
(i) (i) (i)
X
edemo = arg max log p(xdemo |ldemo , f0,demo ; e, w).
e
i

3
Alternatively, all of the model parameters may be additionally fine-tuned (SEA-A LL):
(i) (i) (i)
X
(edemo , wfinetuned ) = arg max log p(xdemo |ldemo , f0,demo ; e, w).
e,w
i

Both methods are non-parametric approaches to few-shot voice adaptation as the number of em-
bedding vectors scales with the number of speakers. However, the training processes are slightly
different. Because the SEA-E MB method optimizes only a low-dimensional vector, it is far less prone
to overfitting, and we are therefore able to retrain the model to convergence even with mere seconds
of adaptation data. By contrast, the SEA-A LL has many more parameters that might overfit to the
adaptation data. We therefore hold out 10% of our demonstration data for calculating a standard early
termination criterion. We also initialize e with the optimal value from the SEA-E MB method, and we
find this initialization significantly improves the generalization performance even with a few seconds
of adaptation data.

3.2 PARAMETRIC FEW- SHOT ADAPTATION USING AN EMBEDDING ENCODER

In contrast to the non-parametric approach, whereby a different embedding vector is fitted for each
speaker, one can train an auxiliary encoder network to predict an embedding vector for a new speaker
given their demonstration data. Specifically, we model:
T
Y
p(x|l, f0 , xdemo , ldemo , f0,demo ; w) = p(xt |x1:t−1 , l, f0 ; e(xdemo , ldemo , f0,demo ), w),
t=1

where for each training example, we include a randomly selected demonstration utterance from that
speaker in addition to the regular conditioning inputs. The full WaveNet model and the encoder
network e(·) are trained together from scratch. We refer the reader to the Appendix for further
architectural details. This approach (SEA-E NC) exhibits the advantage of being trained in a transcript-
independent setting given only the input waveform, e(xdemo ), and requires negligible computation at
adaptation time. However, the learned encoder can also introduce bias when fitting an embedding
due to its limited network capacity. As an example, Li et al. (2017) demonstrated a typical scenario
whereby speaker identity information can be very quickly extracted with deep models from audio
signals. Nonetheless, that the model is less capable of effectively leveraging additional training than
approaches based on statistical methods.

3.3 R EMOVING IDENTITY- RELATED INFORMATION

The linguistic features and fundamental frequencies which are used as inputs contain information
specific to an individual speaker. As an example, the average voice pitch in the fundamental
frequency sequence is highly speaker-dependent. Instead, we would like these features to be as
speaker-independent as possible such that identity is modeled via global conditioning on the speaker
embedding. To achieve this, we normalize the fundamental frequency values to have zero mean and
unit variance separately for each speaker during training, denoted as f̂0 := (f − E[fs ])/std(fs ). As
mentioned earlier, at test time, we use an existing model (Zen et al., 2016) to predict (l, f̂0 ).

4 R ELATED WORK
Few-shot learning to build models, where one can rapidly learn using only a small amount of available
data, is one of the most important open challenges in machine learning. Recent studies have attempted
to address the problem of few-shot learning by using deep neural networks, and they have shown
promising results on classification tasks in vision (Santoro et al., 2016; Shyam et al., 2017) and
language (Vinyals et al., 2016). Few-shot learning can also be leveraged in reinforcement learning,
such as by imitating human Atari gameplay from a single recorded action sequence (Pohlen et al.,
2018) or online video (Aytar et al., 2018).
Meta-learning offers a sound framework for addressing few-shot learning. Here, an expensive learning
process results in machines with the ability to learn rapidly from few data. Meta-learning has a long
history (Harlow, 1949; Thrun and Pratt, 2012), and recent studies include efforts to learn optimization

4
processes (Andrychowicz et al., 2016; Chen et al., 2017) that have been shown to extend naturally to
the few-shot setting (Ravi and Larochelle, 2016). An alternative approach is model-agnostic meta
learning (MAML) (Finn et al., 2017a), which differs by using a fixed optimizer and learning a set of
base parameters that can be adapted to minimize any task loss by few steps of gradient descent. This
method has shown promise in robotics (Finn et al., 2017b; Yu et al., 2018).
In generative modeling, few-shot learning has been addressed from several perspectives, including
matching networks (Bartunov and Vetrov, 2017) and variable inference for memory addressing
(Bornschein et al., 2017). Rezende et al. (2016) developed a sequential generative model that
extended the Deep Recurrent Attention Writer (DRAW) model (Gregor et al., 2015), and Reed et al.
(2018) extended PixelCNN (Van Oord et al., 2016) with neural attention for few-shot auto-regressive
density modeling. Veness et al. (2017) presented a gated linear model able to model complex densities
from a single pass of a limited dataset.
Early attempts of few-shot adaptation involved the attention models of Reed et al. (2018) and
MAML (Finn et al., 2017a), but we found both of these strategies failed to learn informative speaker
embedding in our preliminary experiments.
There is growing interest in developing neural TTS models that can be trained end-to-end without
the need for hand-crafted representations. In this study we focus on extending the autoregressive
WaveNet model (van den Oord et al., 2016; 2017) to the few-shot learning setting to adapt to speakers
that were not presented at training time. Other recent neural TTS models include Tacotron 2 (Skerry-
Ryan et al., 2018) (building on (Wang et al., 2017)) which uses WaveNet as a vocoder to invert
mel-spectrograms generated by an attentive sequence-to-sequence model. DeepVoice 2 (Gibiansky
et al., 2017) (building on (Arık et al., 2017)) introduced a multi-speaker variation of Tacotron that
learns a low-dimensional embedding for each speaker, which was further extended in DeepVoice 3
(Ping et al., 2018) to a 2,400 multi-speaker scenario. Unlike WaveNet and DeepVoice, the Char2Wav
(Sotelo et al., 2017) and VoiceLoop (Taigman et al., 2018) models produce World Vocoder Features
(Morise et al., 2016) instead of generating raw audio signals.
Although many of these systems have produced high-quality samples for speakers present in the
training set, generalizing to new speakers given only a few seconds of audio remains a challenge.
There have been several concurrent works to address this few-shot learning problem. The VoiceLoop
model introduced a novel memory-based architecture that was extended by Nachmani et al. (2018)
to few-shot voice style adaptation, by introducing an auxiliary fitting network that predicts the
embedding of a new speaker. Jia et al. (2018) extended the Tacotron model for one-shot speaker
adaptation by conditioning on a speaker embedding vector extracted from a pretrained speaker identity
model of Wan et al. (2018). The most similar approached to our work was proposed by Arik et al.
(2018) for the DeepVoice 3 model. They considered both predicting the embedding with an encoding
network and fitting the embedding based on a small amount of adaptation data, but the adaptation
was applied to a prediction model for mel-spectrograms with a fixed vocoder.

5 E VALUATION
In this section, we evaluate the quality of samples of SEA-A LL, SEA-E MB and SEA-E NC. We first
measure the naturalness of the generated utterances using the standard Mean Opinion Score (MOS)
procedure. Then, we evaluate the similarity of generated and real samples using the subjective MOS
test and objectively using a speaker verification system (Wan et al., 2018). Finally, we study these
results varying the size of the adaptation dataset.

5.1 E XPERIMENTAL SETUP

We train a WaveNet model for each of our three methods using the same dataset, which combines the
high-quality LibriSpeech audiobook corpus (Panayotov et al., 2015) and a proprietary speech corpus.
The LibriSpeech dataset consists of 2302 speakers from the train speaker subsets and approximately
500 hours of utterances, sampled at a frequency of 16 kHz. The proprietary speech corpus consists
of 10 American English speakers and approximately 300 hours of utterances, and we down-sample
the recording frequency to 16 kHz to match LibriSpeech. The multi-speaker WaveNet model has
the same architecture as van den Oord et al. (2016) except that we use a 200-dimensional speaker
embedding space to model the large diversity of voices.

5
 Dataset LibriSpeech VCTK
Real utterance 4.38 ± 0.04 4.45 ± 0.04
van den Oord et al. (2016) 4.21 ± 0.081
Nachmani et al. (2018) 2.53 ± 1.11 3.66 ± 0.84
Arik et al. (2018)
adapt embedding - 2.67 ± 0.10
adapt whole-model - 3.16 ± 0.09
encoding + fine-tuning - 2.99 ± 0.12
Jia et al. (2018)
trained on LibriSpeech 4.12 ± 0.05 4.01 ± 0.06
Adaptation data size 10s <5m 10s <10m
SEA-A LL (ours) 3.94 ± 0.08 4.13 ± 0.06 3.92 ± 0.07 3.92 ± 0.07
SEA-E MB (ours) 3.86 ± 0.07 3.95 ± 0.07 3.81 ± 0.07 3.82 ± 0.07
SEA-E NC (ours) 3.61 ± 0.06 3.56 ± 0.06 3.65 ± 0.06 3.58 ± 0.06

Table 1: Naturalness of the adapted voices using a 5-scale MOS score (higher is better) with 95%
confidence interval on the LibriSpeech and VCTK held-out adaptation datasets. Numbers in bold are
the best few-shot learning results on each dataset without statistically significant difference. van den
Oord et al. (2016) was trained with 24-hour production quality data, Nachmani et al. (2018) used all
samples of each new speaker, Arik et al. (2018) used 10 samples, and Jia et al. (2018) used 5 seconds.

Our few-shot model performance is evaluated using two hold-out datasets. First, the LibriSpeech
test corpus consists of 39 speakers, with an average of approximately 52 utterances and 5 minutes
of audio per speaker. For every test speaker, we randomly split their demonstration utterances into
an adaptation set for adapting our WaveNet models and a test set for evaluation. The subset of
utterances used for early termination in Section 3.1 is chosen from the adaptation set. There are about
4.2 utterances on average per speaker in the test set and the rest in the adaptation set. Second, we
consider a subset of the CSTR VCTK corpus (Veaux et al., 2017) consisting of 21 American English
speakers, with approximately 368 utterances and 12 minutes of audio per speaker. We also apply the
adaptation/test split with 10 utterances per speaker for test. We emphasize that no data from VCTK
was presented to the model at training time. Since our underlying WaveNet model was trained on
data largely from LibriSpeech (which was recorded under noisier conditions than VCTK), one might
expect that the generated samples on the VCTK dataset contain characteristic artifacts that make
generated samples easier to distinguish from real utterances. However, our evaluation using VCTK
indicates that our model generalizes effectively and that such artifacts are not detectable. Synthetic
utterances are provided on our demo webpage1 .
It is worth mentioning, that SEA-E NC requires no adaptation time. Where for SEA-E MB, it takes
5 ∼ 10k optimizing steps to fit the embedding vector, and an additional 100 ∼ 200 steps to fine-tune
the entire model using early stopping for SEA-A LL.

5.2 NATURALNESS OF THE GENERATED SAMPLES (MOS)

We measure the quality of the generated samples by conducting a MOS test, whereby subjects are
asked to rate the naturalness of generated utterances on a five-point Likert Scale (1: Bad, 2: Poor,
3: Fair, 4: Good, 5: Excellent). Furthermore, we compare with other published few-shot TTS
systems systems, that were developed in parallel to this work. However, the literature uses varying
combinations of training data and evaluation splits making comparison difficult. The results presented
are from the closest experimental setups to ours.
Table 1 presents MOS for the adaptation models compared to real utterances. Two different adaptation
dataset sizes are considered; T = 10 seconds, and T ≤ 5 minutes for LibriSpeech (T ≤ 10 minutes
for VCTK). For reference on 16 kHz data, WaveNet trained on a 24-hour production quality speech
dataset (van den Oord et al., 2016) achieves a score of 4.21, while for LibriSpeech our best few-shot
model attains an MOS score of 4.13 using only 5 minutes of data given a pre-trained multi-speaker
model. We note that both fine-tuning models produce overall “good” samples for both the LibriSpeech
and VCTK test sets, with SEA-A LL outperforming SEA-E MB in all cases. SEA-A LL is on par
1
https://sample-efficient-adaptive-tts.github.io/demo

6
 Dataset LibriSpeech VCTK
Real utterance 4.30 ± 0.08 4.59 ± 0.06
Jia et al. (2018)
trained on LibriSpeech 3.03 ± 0.09 2.77 ± 0.08
Adaptation data size 10s <5m 10s <10m
SEA-A LL (ours) 3.41 ± 0.10 3.75 ± 0.09 3.51 ± 0.10 3.97 ± 0.09
SEA-E MB (ours) 3.42 ± 0.10 3.56 ± 0.10 3.07 ± 0.10 3.18 ± 0.10
SEA-E NC (ours) 2.47 ± 0.09 2.59 ± 0.09 2.07 ± 0.08 2.19 ± 0.09

Table 2: Voice similarity of generated voices using a 5-scale MOS score (higher is better) with 95%
confidence interval on the LibriSpeech and VCTK held-out adaptation datasets.

with the state-of-the-art performance on both datasets. The addition of extra adaptation data beyond
10 seconds of audio helps performance on LibriSpeech but not VCTK, and the gap between our
best model and the real utterance is also wider on VCTK, possibly due to the different recording
conditions.

5.3 VOICE SIMILARITY (MOS)

Beside naturalness, we also measure the similarity of the generated and real voices. The quality
of similarity is the main evaluation metric for the voice adaptation problem. We first follow the
experiment setup of Jia et al. (2018) to run a MOS test for a subjective assessment and then use a
speaker verification model for objective evaluation in the next section.
In every trial of this test a subject is presented with a pair of utterances consisting of a real utterance
and another real or generated utterance from the same speaker, and is asked to rate the similarity in
voice identity using a five-scale score (1: Not at all similar, 2: Slightly similar, 3: Moderately similar,
4: Very similar, 5: Extremely similar).
Table 2 shows the MOS for real utterances and all the adaptation models under two adaptation data
time settings on both datasets. Again, the SEA-A LL model outperforms the other two models, and
the improvement over SEA-E MB scales with the amount of adaptation data. Particularly, the learned
voices on the VCTK dataset achieve an average score of 3.97, demonstrating the generalization
performance on a different dataset. As a rough comparisson, because of varying training setups, the
state of the art system of Jia et al. (2018) achieves scores of 3.03 for LibriSpeech and 2.77 for VCTK
when trained on LibriSpeech. Their model computes the embedding based on the d-vector, similar
to our SEA-E NC approach, and performs competitively for the one-shot learning setting, but its
performance saturates with 5 seconds of adaptation data, as explained in Section 3.2. We note the
gap of similarity scores between SEA-A LL and real utterances, which suggests that although the
generated samples sound similar to the target speakers, humans can still tell the difference from real
utterances.

5.4 VOICE SIMILARITY ( SPEAKER VERIFICATION )

We also apply the state-of-the-art text independent speaker verification (TI-SV) model of (Wan et al.,
2018) to objectively assess whether the generated samples preserve the acoustic features of the
speakers. We calculate the TI-SV d-vector embeddings for generated and real voices. In Figure 3,
we visualize the 2-dimensional projection of the d-vectors for a SEA-A LL model trained on T ≤ 5
minutes of data on the LibriSpeech dataset, and T ≤ 10 minutes on VCTK. There are clear clusters
on both datasets, with a strikingly large inter-cluster distance and low intra-cluster separation. This
shows both (1) an ease of correctly identifying the speaker associated with a given generated utterance,
and (2) the difficulty in differentiating real from synthetic samples.
A similar figure is presented in (Jia et al., 2018), but there the generated and real samples do not
overlap. This indicates that the method presented in this paper generates voices that are more
indistinguishable from real ones, when measured with the same verification system. In the following
subsections, we further analyze these results.

7
 Generated Generated
Real Real

(a) LibriSpeech (b) VCTK

Figure 3: t-SNE visualization of the d-vector embeddings of real and SEA-A LL-generated utterances,
for both the LibriSpeech (T ≤ 5 mins) and VCTK (T ≤ 10 mins) evaluation datasets.

100 100

10 10
False Negative Rate (%)

False Negative Rate (%)

1 1

Real, EER: 2.52 Real, EER: 2.79

0.1 SEA-All, EER: 1.82 0.1 SEA-All, EER: 4.39
SEA-Emb, EER: 2.69 SEA-Emb, EER: 10.33
SEA-Enc, EER: 9.41 SEA-Enc, EER: 25.22
0.1 1 10 100 0.1 1 10 100
False Positive Rate (%) False Positive Rate (%)
(a) LibriSpeech (b) VCTK

Figure 4: Detection error trade-off (DET) curve for speaker verification in percentage, using the
TI-SV speaker verification model (Wan et al., 2018). The utterances were generated using T ≤ 5 and
T ≤ 10 minute samples from LibriSpeech and VCTK respectively. EER is marked with a dot.

5.4.1 D ISCERNING DIFFERENT SPEAKERS

We first quantify whether generated utterances are attributed to the correct speaker. Following
common practice in speaker verification (Wan et al., 2018), we select the hold-out test set of real
utterances from test speakers as the enrollment set and compute the centroid of the d-vectors for
each speaker ci . We then use the adaptation set of test speakers as the verification set. For every
verification utterance, we compute the cosine similarity between its d-vector v and a randomly chosen
centroid ci . The utterance is accepted as one from speaker i if the similarity is exceeds a given
threshold. We repeat the experiments with the same enrollment set and replace the verification set
with samples generated by each adaptation method under different data size settings.

Dataset LibriSpeech VCTK

Real utterance 2.47 2.79
Adaptation data size 10s 1m <5m 10s 1m <10m
SEA-A LL (ours) 3.17 2.47 1.85 7.34 5.02 4.33
SEA-E MB (ours) 3.26 2.92 2.74 10.18 9.91 10.24
SEA-E NC (ours) 10.73 9.77 9.42 27.20 25.34 25.23

Table 3: Equal error rate (EER) of real and few-shot adapted voice samples for evaluation of voice
similarity. Varying adaptation dataset sizes were considered.

8
 Same speaker (real) Same speaker (real)
Random speaker (real) Random speaker (real)
SEA-All 10m SEA-All 10m
- 1m - 1m
- 10s - 10s
SEA-Emb 10m SEA-Emb 10m
- 1m - 1m
- 10s - 10s
SEA-Enc 10m SEA-Enc 10m
- 1m - 1m
- 10s - 10s
0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00
(a) LibriSpeech (b) VCTK

Figure 5: Cosine similarity of real and generated utterances to the real enrollment set.

In our setup we fix the enrollment set together with the speaker verification model from (Wan et al.,
2018), and study the performance of different verification sets that are either from real utterances or
generated by a TTS system. Table 3 lists the equal error rate (EER) of the verification model with real
and generated verification utterances, and Figure 4 shows the detection error trade-off (DET) curves
for a more thorough inspection. Figure 4 only shows the adaptation models with the maximum data
size setting (T ≤ 5 minutes for LibriSpeech and ≤ 10 minutes for VCTK). The results for other data
sizes are provided in Appendix B.
We find that SEA-A LL outperforms the other two approaches, and the error rate decreases clearly
with the size of demonstration data. Noticeably, the EER of SEA-A LL is even lower than the real
utterance on the LibriSpeech dataset with sufficient adaptation data. A possible explanation is that the
generated samples might be concentrated closer to the centroid of a speaker’s embeddings than real
speech with larger variance across utterances. Our SEA-E MB model performs better than SEA-E NC.
Additionally, the benefit of more demonstration data is less significant than for SEA-A LL in both of
these models.

5.4.2 D ISCERNING REAL FROM GENERATED UTTERANCES

In this section, we compare the generated samples and the real utterances of the speaker being
imitated. Figure 5 shows the box-plot of the cosine similarity between the embedding centroids of
test speakers’ enrollment set and (1) real utterances from the same speaker, (2) real utterances from
a different speaker, and (3) generated utterances adapted to the same speaker. Consistent with the
observations from the previous subsection, SEA-A LL performs best.
We further consider an adversarial scenario for speaker verification. In contrast to the previous
standard speaker verification setup where we now select a verification utterance with either a real
utterance from the same speaker or a synthetic sample from a model adapted to the same speaker.
Under this setup, the speaker verification system is challenged by synthetic samples and acts as a
classifier for real versus generated utterances. The ROC curve of this setup is shown in Figure 6
and the models are using the maximum data size setting. Other data size settings can be found in
Appendix C. If the generated samples are indistinguishable from real utterances, the ROC curve
approaches the diagonal line (that is, the verification system fails to separate real and generated
voices). Importantly, SEA-A LL manages to confuse the verification system especially for the VCTK
dataset where the ROC curve is almost inline with the diagonal line with an AUC of 0.56.

6 C ONCLUSION
This paper studied three variants of meta-learning for sample efficient adaptive TTS. The adaptation
method that fine-tunes the entire model, with the speaker embedding vector first optimized, shows
impressive performance even with only 10 seconds of audio from new speakers. When adapted with
a few minutes of data, our model matches the state-of-the-art performance in sample naturalness.
Moreover, it outperforms other recent works in matching the new speaker’s voice. We also demon-

9
 1.0 1.0

0.8 0.8

True Positive Rate

True Positive Rate

0.6 0.6

0.4 0.4

0.2 SEA-All, AUC: 0.63 0.2 SEA-All, AUC: 0.56

SEA-Emb, AUC: 0.70 SEA-Emb, AUC: 0.80
SEA-Enc, AUC: 0.90 SEA-Enc, AUC: 0.98
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate False Positive Rate
(a) LibriSpeech (b) VCTK

Figure 6: ROC curve for real versus generated utterance detection. The utterances were generated
using models with 5 and 10 minutes of training data per speaker from LibriSpeech and VCTK
respectively. Lower curve indicate that the verification system is having a harder time distinguishing
real from generated samples.

strated that the generated samples achieved a similar level of voice similarity to real utterances from
the same speaker, when measured by a text independent speaker verification model.
Our paper considers the adaptation to new voices with clean, high-quality training data collected in a
controlled environment. The few-shot learning of voices with noisy data is beyond the scope of this
paper and remains a challenging open research problem.
A requirement for less training data to adapt the model, however, increases the potential for both
beneficial and harmful applications of text-to-speech technologies such as the creation of synthesized
media. While the requirements for this particular model (including the high-quality training data
collected in a controlled environment and equally high quality data from the speakers to which we
adapt, as described in Section 5.1) present barriers to misuse, more research must be conducted to
mitigate and detect instances of misuse of text-to-speech technologies in general.

R EFERENCES
G. Hinton, L. Deng, D. Yu, G. E. Dahl, A.-r. Mohamed, N. Jaitly, A. Senior, V. Vanhoucke, P. Nguyen, T. N.
Sainath, et al. Deep neural networks for acoustic modeling in speech recognition: The shared views of four
research groups. IEEE Signal Processing Magazine, 29(6):82–97, 2012.

Y. Wu, M. Schuster, Z. Chen, Q. V. Le, M. Norouzi, W. Macherey, M. Krikun, Y. Cao, Q. Gao, K. Macherey,
et al. Google’s neural machine translation system: Bridging the gap between human and machine translation.
arXiv preprint arXiv:1609.08144, 2016.

A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks.
In Advances in Neural Information processing Systems, pages 1097–1105, 2012.

C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, A. Rabinovich, et al.
Going deeper with convolutions. In Computer Vision and Pattern Recognition, 2015.

T. Dutoit. An Introduction to Text-to-speech Synthesis. Kluwer Academic Publishers, Norwell, MA, USA, 1997.
ISBN 0-7923-4498-7.

P. Taylor. Text-to-Speech Synthesis. Cambridge University Press, New York, NY, USA, 1st edition, 2009. ISBN
0521899273, 9780521899277.

A. van den 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, 2016.

A. van den Oord, Y. Li, I. Babuschkin, K. Simonyan, O. Vinyals, K. Kavukcuoglu, G. v. d. Driessche, E. Lockhart,
L. C. Cobo, F. Stimberg, et al. Parallel WaveNet: Fast high-fidelity speech synthesis. arXiv preprint
arXiv:1711.10433, 2017.

10
L. Wan, Q. Wang, A. Papir, and I. L. Moreno. Generalized end-to-end loss for speaker verification. In
International Conference on Acoustics, Speech, and Signal Processing, pages 4879–4883. IEEE, 2018.

H. Zen, Y. Agiomyrgiannakis, N. Egberts, F. Henderson, and P. Szczepaniak. Fast, compact, and high quality
LSTM-RNN based statistical parametric speech synthesizers for mobile devices. In INTERSPEECH, pages
2273–2277, 2016.

L. Li, Y. Chen, Y. Shi, Z. Tang, and D. Wang. Deep speaker feature learning for text-independent speaker
verification. In INTERSPEECH, pages 1542–1546, 2017.

A. Santoro, S. Bartunov, M. Botvinick, D. Wierstra, and T. Lillicrap. Meta-learning with memory-augmented

neural networks. In International Conference on Machine Learning, pages 1842–1850, 2016.

P. Shyam, S. Gupta, and A. Dukkipati. Attentive recurrent comparators. In International Conference on Machine
Learning, pages 3173–3181, 2017.

O. Vinyals, C. Blundell, T. Lillicrap, D. Wierstra, et al. Matching networks for one shot learning. In Advances in
Neural Information Processing Systems, pages 3630–3638, 2016.

T. Pohlen, B. Piot, T. Hester, M. G. Azar, D. Horgan, D. Budden, G. Barth-Maron, H. van Hasselt, J. Quan,
M. Večerík, et al. Observe and look further: Achieving consistent performance on atari. arXiv preprint
arXiv:1805.11593, 2018.

Y. Aytar, T. Pfaff, D. Budden, T. L. Paine, Z. Wang, and N. de Freitas. Playing hard exploration games by
watching youtube. arXiv preprint arXiv:1805.11592, 2018.

H. F. Harlow. The formation of learning sets. Psychological review, 56(1):51, 1949.

S. Thrun and L. Pratt. Learning to learn. Springer Science & Business Media, 2012.

M. Andrychowicz, M. Denil, S. Gomez, M. W. Hoffman, D. Pfau, T. Schaul, B. Shillingford, and N. De Freitas.

Learning to learn by gradient descent by gradient descent. In Advances in Neural Information Processing
Systems, pages 3981–3989, 2016.

Y. Chen, M. W. Hoffman, S. G. Colmenarejo, M. Denil, T. P. Lillicrap, M. Botvinick, and N. Freitas. Learning

to learn without gradient descent by gradient descent. In International Conference on Machine Learning,
pages 748–756, 2017.

S. Ravi and H. Larochelle. Optimization as a model for few-shot learning. International Conference on Learning
Representations, 2016.

C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta-learning for fast adaptation of deep networks. In
International Conference on Machine Learning, pages 1126–1135, 2017a.

C. Finn, T. Yu, T. Zhang, P. Abbeel, and S. Levine. One-shot visual imitation learning via meta-learning. In
Conference on Robot Learning, pages 357–368, 2017b.

T. Yu, C. Finn, A. Xie, S. Dasari, T. Zhang, P. Abbeel, and S. Levine. One-shot imitation from observing humans
via domain-adaptive meta-learning. In International Conference on Learning Representations Workshop,
2018.

S. Bartunov and D. P. Vetrov. Fast adaptation in generative models with generative matching networks. In
International Conference on Learning Representations Workshop, 2017.

J. Bornschein, A. Mnih, D. Zoran, and D. J. Rezende. Variational memory addressing in generative models. In
Advances in Neural Information Processing Systems, pages 3923–3932, 2017.

D. J. Rezende, S. Mohamed, I. Danihelka, K. Gregor, and D. Wierstra. One-shot generalization in deep generative
models. In International Conference on Machine Learning, pages 1521–1529, 2016.

K. Gregor, I. Danihelka, A. Graves, D. Rezende, and D. Wierstra. DRAW: A recurrent neural network for image
generation. In International Conference on Machine Learning, pages 1462–1471, 2015.

S. Reed, Y. Chen, T. Paine, A. van den Oord, S. M. Eslami, D. Rezende, O. Vinyals, and N. de Freitas. Few-shot
autoregressive density estimation: Towards learning to learn distributions. In International Conference on
Learning Representations, 2018.

A. Van Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. In International
Conference on Machine Learning, pages 1747–1756, 2016.

11
J. Veness, T. Lattimore, A. Bhoopchand, A. Grabska-Barwinska, C. Mattern, and P. Toth. Online learning with
gated linear networks. arXiv preprint arXiv:1712.01897, 2017.

R. Skerry-Ryan, E. Battenberg, Y. Xiao, Y. Wang, D. Stanton, J. Shor, R. J. Weiss, R. Clark, and R. A.

Saurous. Towards end-to-end prosody transfer for expressive speech synthesis with tacotron. arXiv preprint
arXiv:1803.09047, 2018.

Y. Wang, R. Skerry-Ryan, D. Stanton, Y. Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y. Xiao, Z. Chen, S. Bengio,
Q. Le, Y. Agiomyrgiannakis, R. Clark, and R. A. Saurous. Tacotron: Towards end-to-end speech synthesis. In
INTERSPEECH, pages 4006–4010, 2017.

A. Gibiansky, S. Arik, G. Diamos, J. Miller, K. Peng, W. Ping, J. Raiman, and Y. Zhou. Deep voice 2: Multi-
speaker neural text-to-speech. In Advances in Neural Information Processing Systems, pages 2962–2970,
2017.

S. Ö. Arık, M. Chrzanowski, A. Coates, G. Diamos, A. Gibiansky, Y. Kang, X. Li, J. Miller, A. Ng, J. Raiman,
et al. Deep voice: Real-time neural text-to-speech. In International Conference on Machine Learning, pages
195–204, 2017.

W. Ping, K. Peng, A. Gibiansky, S. O. Arik, A. Kannan, S. Narang, J. Raiman, and J. Miller. Deep voice 3:
2000-speaker neural text-to-speech. In International Conference on Learning Representations, 2018.
J. Sotelo, S. Mehri, K. Kumar, J. F. Santos, K. Kastner, A. Courville, and Y. Bengio. Char2wav: End-to-end
speech synthesis. In International Conference on Learning Representations Workshop, 2017.

Y. Taigman, L. Wolf, A. Polyak, and E. Nachmani. Voiceloop: Voice fitting and synthesis via a phonological
loop. In International Conference on Learning Representations, 2018.

M. Morise, F. Yokomori, and K. Ozawa. World: a vocoder-based high-quality speech synthesis system for
real-time applications. IEICE transactions on Information and Systems, 99(7):1877–1884, 2016.
E. Nachmani, A. Polyak, Y. Taigman, and L. Wolf. Fitting new speakers based on a short untranscribed sample.
arXiv preprint arXiv:1802.06984, 2018.

Y. Jia, Y. Zhang, R. J. Weiss, Q. Wang, J. Shen, F. Ren, Z. Chen, P. Nguyen, R. Pang, I. L. Moreno, et al. Transfer
learning from speaker verification to multispeaker text-to-speech synthesis. arXiv preprint arXiv:1806.04558,
2018.

S. O. Arik, J. Chen, K. Peng, W. Ping, and Y. Zhou. Neural voice cloning with a few samples. arXiv preprint
arXiv:1802.06006, 2018.
V. Panayotov, G. Chen, D. Povey, and S. Khudanpur. Librispeech: an asr corpus based on public domain audio
books. In International Conference on Acoustics, Speech and Signal Processing, pages 5206–5210. IEEE,
2015.
C. Veaux, J. Yamagishi, K. MacDonald, et al. CSTR VCTK corpus: English multi-speaker corpus for CSTR
Voice Cloning Toolkit, 2017.

12
A E MBEDDING ENCODER
Our encoding network is illustrated as the summation of two sub-network outputs in Figure 7. The first sub-
network is a pre-trained speaker verification model (TI-SV) (Wan et al., 2018), comprising 3 LSTM layers and
a single linear layer. This model maps a waveform sequence of arbitrary length to a fixed 256-dimensional
d-vector with a sliding window, and is trained from approximately 36M utterances from 18K speakers extracted
from anonymized voice search logs. On top of this we add a shallow MLP to project the output d-vector to
the speaker embedding space. The second sub-network comprises 16 1-D convolutional layers. This network
reduces the temporal resolution to 256 ms per frame (for 16 kHz audio), then averages across time and projects
into the speaker embedding space. The purpose of this network is to extract residual speaker information present
in the demonstration waveforms but not captured by the pre-trained TI-SV model.

Linear

Linear
ReLU
TI-SV
Demo waveform +

Global Mean Pool

Speaker
Dilated CNN embedding
Max Pool

Linear
ReLU

ReLU
CNN

x12 x4

Figure 7: Encoder network architecture for predicting speaker embeddings.

13
B DET CURVES VARYING TRAINING DATA SIZES
Here we provide the DET curves of speaker verification problem for models with different training data sizes in
addition to those shown in Section 5.4.1.

100 100

10 10
False Negative Rate (%)

False Negative Rate (%)

1 1

Real, EER: 2.52 Real, EER: 2.79

0.1 SEA-All, EER: 2.43 0.1 SEA-All, EER: 5.05
SEA-Emb, EER: 2.90 SEA-Emb, EER: 9.88
SEA-Enc, EER: 9.74 SEA-Enc, EER: 25.38
0.1 1 10 100 0.1 1 10 100
False Positive Rate (%) False Positive Rate (%)
(a) LibriSpeech, 1 min (b) VCTK, 1 min
100 100

10 10
False Negative Rate (%)

False Negative Rate (%)

1 1

Real, EER: 2.52 Real, EER: 2.79

0.1 SEA-All, EER: 3.19 0.1 SEA-All, EER: 7.33
SEA-Emb, EER: 3.23 SEA-Emb, EER: 10.15
SEA-Enc, EER: 10.69 SEA-Enc, EER: 27.20
0.1 1 10 100 0.1 1 10 100
False Positive Rate (%) False Positive Rate (%)
(c) LibriSpeech, 10 sec (d) VCTK, 10 sec

Figure 8: Detection error trade-off (DET) curve for speaker verification, using the TI-SV speaker
verification model (Wan et al., 2018). The utterances were generated using 1 minute or 10 seconds of
utterance from LibriSpeech and VCTK. EER is marked with a dot.

14
C ROC CURVES VARYING TRAINING DATA SIZES
We provide the ROC curves of the speaker verification problem with adversarial examples from adaptation
models with different training data sizes in addition to those shown in Section 5.4.2.

1.0 1.0

0.8 0.8
True Positive Rate

True Positive Rate

0.6 0.6

0.4 0.4

0.2 SEA-All, AUC: 0.69 0.2 SEA-All, AUC: 0.62

SEA-Emb, AUC: 0.71 SEA-Emb, AUC: 0.80
SEA-Enc, AUC: 0.90 SEA-Enc, AUC: 0.98
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate False Positive Rate
(a) LibriSpeech, 1 min (b) VCTK, 1 min
1.0 1.0

0.8 0.8
True Positive Rate

True Positive Rate

0.6 0.6

0.4 0.4

0.2 SEA-All, AUC: 0.72 0.2 SEA-All, AUC: 0.70

SEA-Emb, AUC: 0.73 SEA-Emb, AUC: 0.81
SEA-Enc, AUC: 0.91 SEA-Enc, AUC: 0.98
0.0 0.0
0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
False Positive Rate False Positive Rate
(c) LibriSpeech, 10 sec (d) VCTK, 10 sec

Figure 9: ROC curve for real vs. generated utterance detection. The utterances were generated using
1 minute or 10 seconds of utterance from LibriSpeech and VCTK. Lower curve suggests harder to
distinguish real from generated samples.

15