Berrak Sisman
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Berrak Sisman has a total of 61 co-authors. Here are the top co-authors by number of shared publications:
- Haizhou Li
The Chinese University of Hong Kong, Shenzhen, China
30 shared publications - Kun Zhou
Department of Electrical and Computer Engineering, National University of Singapore, Singapore
9 shared publications - Rui Liu
Inner Mongolia University, Hohhot, China
8 shared publications - Björn Schuller
Technische Universität München, Munich, Germany
5 shared publications Mingyang Zhang
Department of Electrical and Computer Engineering National University of Singapore Singapore
5 shared publicationsGuanglai Gao
Department of Computer Science, Inner Mongolia University, Huhhot, China
4 shared publications
Latest Publications
- Open Access
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Signal Processing
- Vol. 7, 247-256
Objective Evaluation of Prosody and Intelligibility in Speech Synthesis via Conditional Prediction of Discrete Tokens
Objective evaluation of synthesized speech is critical for advancing speech generation systems, yet existing metrics for intelligibility and prosody remain limited in scope and weakly correlated with human perception. Word Error Rate (WER) provides only a coarse text-based measure of intelligibility, while F0-RMSE and related pitch-based metrics offer a narrow, reference-dependent view of prosody. To address these limitations, we propose TTScore, a targeted and reference-free evaluation framework based on conditional prediction of discrete speech tokens. TTScore employs two sequence-to-sequence predictors conditioned on input text: TTScore-int, which measures intelligibility through content tokens, and TTScore-pro, which evaluates prosody from the perspective of pitch, through prosody tokens. For each synthesized utterance, the predictors compute the likelihood of the corresponding token sequences, yielding interpretable scores that capture alignment with intended linguistic content and prosodic structure. Experiments on the SOMOS, VoiceMOS, and TTSArena benchmarks demonstrate that TTScore-int and TTScore-pro provide reliable, aspect-specific evaluation and achieve stronger correlations with human judgments of overall quality than existing intelligibility and prosody-focused metrics.
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Open Journal of Signal Processing
- Vol. 7, 247-256
Open AccessObjective Evaluation of Prosody and Intelligibility in Speech Synthesis via Conditional Prediction of Discrete Tokens
Objective evaluation of synthesized speech is critical for advancing speech generation systems, yet existing metrics for intelligibility and prosody remain limited in scope and weakly correlated with human perception. Word Error Rate (WER) provides only a coarse text-based measure of intelligibility, while F0-RMSE and related pitch-based metrics offer a narrow, reference-dependent view of prosody. To address these limitations, we propose TTScore, a targeted and reference-free evaluation framework based on conditional prediction of discrete speech tokens. TTScore employs two sequence-to-sequence predictors conditioned on input text: TTScore-int, which measures intelligibility through content tokens, and TTScore-pro, which evaluates prosody from the perspective of pitch, through prosody tokens. For each synthesized utterance, the predictors compute the likelihood of the corresponding token sequences, yielding interpretable scores that capture alignment with intended linguistic content and prosodic structure. Experiments on the SOMOS, VoiceMOS, and TTSArena benchmarks demonstrate that TTScore-int and TTScore-pro provide reliable, aspect-specific evaluation and achieve stronger correlations with human judgments of overall quality than existing intelligibility and prosody-focused metrics.
- 13 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 32 (10709908) , 776-780
PRESENT: Zero-Shot Text-to-Prosody Control
Current strategies for achieving fine-grained prosody control in speech synthesis entail extracting additional style embeddings or adopting more complex architectures. To enable zero-shot application of pretrained text-to-speech (TTS) models, we present PRESENT (PRosody Editing without Style Embeddings or New Training), which exploits explicit prosody prediction in FastSpeech2-based models by modifying the inference process directly. We apply our text-to-prosody framework to zero-shot language transfer using a JETS model exclusively trained on English LJSpeech data. We obtain character error rates (CER) of 12.8%, 18.7% and 5.9% for German, Hungarian and Spanish respectively, beating the previous state-of-the-art CER by over 2× for all three languages. Furthermore, we allow subphoneme-level control, a first in this field. To evaluate its effectiveness, we show that PRESENT can improve the prosody of questions, and use it to generate Mandarin, a tonal language where vowel pitch varies at subphoneme level. We attain 25.3% hanzi CER and 13.0% pinyin CER with the JETS model. All our code and audio samples 1 1https://github.com/iamanigeeit/present and https://present2024.web.app/ are available online.
- 13 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Signal Processing Letters
- Vol. 32 (10709908) , 776-780
PRESENT: Zero-Shot Text-to-Prosody Control
Current strategies for achieving fine-grained prosody control in speech synthesis entail extracting additional style embeddings or adopting more complex architectures. To enable zero-shot application of pretrained text-to-speech (TTS) models, we present PRESENT (PRosody Editing without Style Embeddings or New Training), which exploits explicit prosody prediction in FastSpeech2-based models by modifying the inference process directly. We apply our text-to-prosody framework to zero-shot language transfer using a JETS model exclusively trained on English LJSpeech data. We obtain character error rates (CER) of 12.8%, 18.7% and 5.9% for German, Hungarian and Spanish respectively, beating the previous state-of-the-art CER by over 2× for all three languages. Furthermore, we allow subphoneme-level control, a first in this field. To evaluate its effectiveness, we show that PRESENT can improve the prosody of questions, and use it to generate Mandarin, a tonal language where vowel pitch varies at subphoneme level. We attain 25.3% hanzi CER and 13.0% pinyin CER with the JETS model. All our code and audio samples 1 1https://github.com/iamanigeeit/present and https://present2024.web.app/ are available online.
- Open Access
- 6 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 110492-110503
emoDARTS: Joint Optimization of CNN and Sequential Neural Network Architectures for Superior Speech Emotion Recognition
Speech Emotion Recognition (SER) is crucial for enabling computers to understand the emotions conveyed in human communication. With recent advancements in Deep Learning (DL), the performance of SER models has significantly improved. However, designing an optimal DL architecture requires specialised knowledge and experimental assessments. Fortunately, Neural Architecture Search (NAS) provides a potential solution for automatically determining the best DL model. The Differentiable Architecture Search (DARTS) is a particularly efficient method for discovering optimal models. This study presents emoDARTS, a DARTS-optimised joint CNN and Sequential Neural Network (SeqNN: LSTM, RNN) architecture that enhances SER performance. The literature supports the selection of CNN and LSTM coupling to improve performance. While DARTS has previously been used to choose CNN and LSTM operations independently, our technique adds a novel mechanism for selecting CNN and SeqNN operations in conjunction using DARTS. Unlike earlier work, we do not impose limits on the layer order of the CNN. Instead, we let DARTS choose the best layer order inside the DARTS cell. We demonstrate that emoDARTS outperforms conventionally designed CNN-LSTM models and surpasses the best-reported SER results achieved through DARTS on CNN-LSTM by evaluating our approach on the IEMOCAP, MSP-IMPROV, and MSP-Podcast datasets.
- 6 August 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 110492-110503
Open AccessemoDARTS: Joint Optimization of CNN and Sequential Neural Network Architectures for Superior Speech Emotion Recognition
Speech Emotion Recognition (SER) is crucial for enabling computers to understand the emotions conveyed in human communication. With recent advancements in Deep Learning (DL), the performance of SER models has significantly improved. However, designing an optimal DL architecture requires specialised knowledge and experimental assessments. Fortunately, Neural Architecture Search (NAS) provides a potential solution for automatically determining the best DL model. The Differentiable Architecture Search (DARTS) is a particularly efficient method for discovering optimal models. This study presents emoDARTS, a DARTS-optimised joint CNN and Sequential Neural Network (SeqNN: LSTM, RNN) architecture that enhances SER performance. The literature supports the selection of CNN and LSTM coupling to improve performance. While DARTS has previously been used to choose CNN and LSTM operations independently, our technique adds a novel mechanism for selecting CNN and SeqNN operations in conjunction using DARTS. Unlike earlier work, we do not impose limits on the layer order of the CNN. Instead, we let DARTS choose the best layer order inside the DARTS cell. We demonstrate that emoDARTS outperforms conventionally designed CNN-LSTM models and surpasses the best-reported SER results achieved through DARTS on CNN-LSTM by evaluating our approach on the IEMOCAP, MSP-IMPROV, and MSP-Podcast datasets.
- Open Access
- 25 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 16 (1) , 306-318
Versatile Audio-Visual Learning for Emotion Recognition
Most current audio-visual emotion recognition models lack the flexibility needed for deployment in practical applications. We envision a multimodal system that works even when only one modality is available and can be implemented interchangeably for either predicting emotional attributes or recognizing categorical emotions. Achieving such flexibility in a multimodal emotion recognition system is difficult due to the inherent challenges in accurately interpreting and integrating varied data sources. It is also a challenge to robustly handle missing or partial information while allowing direct switch between regression or classification tasks. This study proposes a versatile audio-visual learning (VAVL) framework for handling unimodal and multimodal systems for emotion regression or emotion classification tasks. We implement an audio-visual framework that can be trained even when audio and visual paired data is not available for part of the training set (i.e., audio only or only video is present). We achieve this effective representation learning with audio-visual shared layers, residual connections over shared layers, and a unimodal reconstruction task. Our experimental results reveal that our architecture significantly outperforms strong baselines on the CREMA-D, MSP-IMPROV, and CMU-MOSEI corpora. Notably, VAVL attains a new state-of-the-art performance in the emotional attribute prediction task on the MSP-IMPROV corpus.
- 25 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 16 (1) , 306-318
Open AccessVersatile Audio-Visual Learning for Emotion Recognition
Most current audio-visual emotion recognition models lack the flexibility needed for deployment in practical applications. We envision a multimodal system that works even when only one modality is available and can be implemented interchangeably for either predicting emotional attributes or recognizing categorical emotions. Achieving such flexibility in a multimodal emotion recognition system is difficult due to the inherent challenges in accurately interpreting and integrating varied data sources. It is also a challenge to robustly handle missing or partial information while allowing direct switch between regression or classification tasks. This study proposes a versatile audio-visual learning (VAVL) framework for handling unimodal and multimodal systems for emotion regression or emotion classification tasks. We implement an audio-visual framework that can be trained even when audio and visual paired data is not available for part of the training set (i.e., audio only or only video is present). We achieve this effective representation learning with audio-visual shared layers, residual connections over shared layers, and a unimodal reconstruction task. Our experimental results reveal that our architecture significantly outperforms strong baselines on the CREMA-D, MSP-IMPROV, and CMU-MOSEI corpora. Notably, VAVL attains a new state-of-the-art performance in the emotional attribute prediction task on the MSP-IMPROV corpus.
- 1 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 2188-2201
Controllable Accented Text-to-Speech Synthesis With Fine and Coarse-Grained Intensity Rendering
Accented text-to-speech (TTS) synthesis seeks to generate speech with an accent (L2) as a variant of the standard version (L1), which is challenging as L2 is different from L1 in terms of phonetic rendering and prosody pattern (pitch, energy, and duration variance, etc.). Accented TTS has several significant real-world applications, such as language learning, preserving and documenting endangered languages and dialects, etc. that make it an important area of research and development. Moreover, changing the accent intensity of any conversational AI system has the potential to allow specific users to understand its produced speech better. However, there is no intuitive solution for the control of the accent intensity for an utterance at both fine and coarse-grained levels, that are phoneme and utterance levels respectively. In this work, we propose a neural TTS architecture that allows us to control the accent style and its intensity. This is achieved through two novel mechanisms: 1) the front-end and back-end accent knowledge injection mechanism to enhance the accent interpretability of TTS modeling; and 2) an automatic speech recognition (ASR) based accent intensity modeling strategy to quantify the accent intensity in both L2 phoneme and utterance levels. In the front-end, a new accent variation adaptor seeks to project the accent-aware pitch, energy and duration features at a phoneme level, with the help of the fine-grained accent intensity information; In the back-end, a consistency constraint module that ensures the synthesized L2 speech manifests the expected accent intensity, is injected in the front-end, precisely. Experiments show that the proposed system attains superior performance to the baseline models in terms of accent rendering and intensity control. To our knowledge, this is the first study of accented TTS with explicit intensity control at both fine and coarse-grained levels.
- 1 April 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 32 (23299290) , 2188-2201
Controllable Accented Text-to-Speech Synthesis With Fine and Coarse-Grained Intensity Rendering
Accented text-to-speech (TTS) synthesis seeks to generate speech with an accent (L2) as a variant of the standard version (L1), which is challenging as L2 is different from L1 in terms of phonetic rendering and prosody pattern (pitch, energy, and duration variance, etc.). Accented TTS has several significant real-world applications, such as language learning, preserving and documenting endangered languages and dialects, etc. that make it an important area of research and development. Moreover, changing the accent intensity of any conversational AI system has the potential to allow specific users to understand its produced speech better. However, there is no intuitive solution for the control of the accent intensity for an utterance at both fine and coarse-grained levels, that are phoneme and utterance levels respectively. In this work, we propose a neural TTS architecture that allows us to control the accent style and its intensity. This is achieved through two novel mechanisms: 1) the front-end and back-end accent knowledge injection mechanism to enhance the accent interpretability of TTS modeling; and 2) an automatic speech recognition (ASR) based accent intensity modeling strategy to quantify the accent intensity in both L2 phoneme and utterance levels. In the front-end, a new accent variation adaptor seeks to project the accent-aware pitch, energy and duration features at a phoneme level, with the help of the fine-grained accent intensity information; In the back-end, a consistency constraint module that ensures the synthesized L2 speech manifests the expected accent intensity, is injected in the front-end, precisely. Experiments show that the proposed system attains superior performance to the baseline models in terms of accent rendering and intensity control. To our knowledge, this is the first study of accented TTS with explicit intensity control at both fine and coarse-grained levels.
- Open Access
- 30 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 14 (4) , 3120-3134
Speech Synthesis With Mixed Emotions
Emotional speech synthesis aims to synthesize human voices with various emotional effects. The current studies are mostly focused on imitating an averaged style belonging to a specific emotion type. In this paper, we seek to generate speech with a mixture of emotions at run-time. We propose a novel formulation that measures the relative difference between the speech samples of different emotions. We then incorporate our formulation into a sequence-to-sequence emotional text-to-speech framework. During the training, the framework does not only explicitly characterize emotion styles but also explores the ordinal nature of emotions by quantifying the differences with other emotions. At run-time, we control the model to produce the desired emotion mixture by manually defining an emotion attribute vector. The objective and subjective evaluations have validated the effectiveness of the proposed framework. To our best knowledge, this research is the first study on modelling, synthesizing, and evaluating mixed emotions in speech.
- 30 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 14 (4) , 3120-3134
Open AccessSpeech Synthesis With Mixed Emotions
Emotional speech synthesis aims to synthesize human voices with various emotional effects. The current studies are mostly focused on imitating an averaged style belonging to a specific emotion type. In this paper, we seek to generate speech with a mixture of emotions at run-time. We propose a novel formulation that measures the relative difference between the speech samples of different emotions. We then incorporate our formulation into a sequence-to-sequence emotional text-to-speech framework. During the training, the framework does not only explicitly characterize emotion styles but also explores the ordinal nature of emotions by quantifying the differences with other emotions. At run-time, we control the model to produce the desired emotion mixture by manually defining an emotion attribute vector. The objective and subjective evaluations have validated the effectiveness of the proposed framework. To our best knowledge, this research is the first study on modelling, synthesizing, and evaluating mixed emotions in speech.
- Open Access
Speech Synthesis with Mixed Emotions
Emotional speech synthesis aims to synthesize human voices with various emotional effects. The current studies are mostly focused on imitating an averaged style belonging to a specific emotion type. In this paper, we seek to generate speech with a mixture of emotions at run-time. We propose a novel formulation that measures the relative difference between the speech samples of different emotions. We then incorporate our formulation into a sequence-to-sequence emotional text-to-speech framework. During the training, the framework does not only explicitly characterize emotion styles, but also explores the ordinal nature of emotions by quantifying the differences with other emotions. At run-time, we control the model to produce the desired emotion mixture by manually defining an emotion attribute vector. The objective and subjective evaluations have validated the effectiveness of the proposed framework. To our best knowledge, this research is the first study on modelling, synthesizing, and evaluating mixed emotions in speech.
Open AccessSpeech Synthesis with Mixed Emotions
Emotional speech synthesis aims to synthesize human voices with various emotional effects. The current studies are mostly focused on imitating an averaged style belonging to a specific emotion type. In this paper, we seek to generate speech with a mixture of emotions at run-time. We propose a novel formulation that measures the relative difference between the speech samples of different emotions. We then incorporate our formulation into a sequence-to-sequence emotional text-to-speech framework. During the training, the framework does not only explicitly characterize emotion styles, but also explores the ordinal nature of emotions by quantifying the differences with other emotions. At run-time, we control the model to produce the desired emotion mixture by manually defining an emotion attribute vector. The objective and subjective evaluations have validated the effectiveness of the proposed framework. To our best knowledge, this research is the first study on modelling, synthesizing, and evaluating mixed emotions in speech.
- Open Access
Emotion Intensity and its Control for Emotional Voice Conversion
Emotional voice conversion (EVC) seeks to convert the emotional state of an utterance while preserving the linguistic content and speaker identity. In EVC, emotions are usually treated as discrete categories overlooking the fact that speech also conveys emotions with various intensity levels that the listener can perceive. In this paper, we aim to explicitly characterize and control the intensity of emotion. We propose to disentangle the speaker style from linguistic content and encode the speaker style into a style embedding in a continuous space that forms the prototype of emotion embedding. We further learn the actual emotion encoder from an emotion-labelled database and study the use of relative attributes to represent fine-grained emotion intensity. To ensure emotional intelligibility, we incorporate emotion classification loss and emotion embedding similarity loss into the training of the EVC network. As desired, the proposed network controls the fine-grained emotion intensity in the output speech. Through both objective and subjective evaluations, we validate the effectiveness of the proposed network for emotional expressiveness and emotion intensity control.
Open AccessEmotion Intensity and its Control for Emotional Voice Conversion
Emotional voice conversion (EVC) seeks to convert the emotional state of an utterance while preserving the linguistic content and speaker identity. In EVC, emotions are usually treated as discrete categories overlooking the fact that speech also conveys emotions with various intensity levels that the listener can perceive. In this paper, we aim to explicitly characterize and control the intensity of emotion. We propose to disentangle the speaker style from linguistic content and encode the speaker style into a style embedding in a continuous space that forms the prototype of emotion embedding. We further learn the actual emotion encoder from an emotion-labelled database and study the use of relative attributes to represent fine-grained emotion intensity. To ensure emotional intelligibility, we incorporate emotion classification loss and emotion embedding similarity loss into the training of the EVC network. As desired, the proposed network controls the fine-grained emotion intensity in the output speech. Through both objective and subjective evaluations, we validate the effectiveness of the proposed network for emotional expressiveness and emotion intensity control.
- Top Cited PapersOpen Access
- 19 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 14 (1) , 31-48
Emotion Intensity and its Control for Emotional Voice Conversion
Emotional voice conversion (EVC) seeks to convert the emotional state of an utterance while preserving the linguistic content and speaker identity. In EVC, emotions are usually treated as discrete categories overlooking the fact that speech also conveys emotions with various intensity levels that the listener can perceive. In this work, we aim to explicitly characterize and control the intensity of emotion. We propose to disentangle the speaker style from linguistic content and encode the speaker style into a style embedding in a continuous space that forms the prototype of emotion embedding. We further learn the actual emotion encoder from an emotion-labelled database and study the use of relative attributes to represent fine-grained emotion intensity. To ensure emotional intelligibility, we incorporate emotion classification loss and emotion embedding similarity loss into the training of the EVC network. As desired, the proposed network controls the fine-grained emotion intensity in the output speech. Through both objective and subjective evaluations, we validate the effectiveness of the proposed network for emotional expressiveness and emotion intensity control.
- 19 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Affective Computing
- Vol. 14 (1) , 31-48
Top Cited PapersOpen AccessEmotion Intensity and its Control for Emotional Voice Conversion
Emotional voice conversion (EVC) seeks to convert the emotional state of an utterance while preserving the linguistic content and speaker identity. In EVC, emotions are usually treated as discrete categories overlooking the fact that speech also conveys emotions with various intensity levels that the listener can perceive. In this work, we aim to explicitly characterize and control the intensity of emotion. We propose to disentangle the speaker style from linguistic content and encode the speaker style into a style embedding in a continuous space that forms the prototype of emotion embedding. We further learn the actual emotion encoder from an emotion-labelled database and study the use of relative attributes to represent fine-grained emotion intensity. To ensure emotional intelligibility, we incorporate emotion classification loss and emotion embedding similarity loss into the training of the EVC network. As desired, the proposed network controls the fine-grained emotion intensity in the output speech. Through both objective and subjective evaluations, we validate the effectiveness of the proposed network for emotional expressiveness and emotion intensity control.
- Open Access
- 3 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 30 (23299290) , 1789-1802
Decoding Knowledge Transfer for Neural Text-to-Speech Training
Neural end-to-end text-to-speech (TTS) is superior to conventional statistical methods in many ways. However, the exposure bias problem, that arises from the mismatch between the training and inference process in autoregressive models, remains an issue. It often leads to performance degradation in face of out-of-domain test data. To address this problem, we study a novel decoding knowledge transfer strategy, and propose a multi-teacher knowledge distillation (MT-KD) network for Tacotron2 TTS model. The idea is to pre-train two Tacotron2 TTS teacher models in teacher forcing and scheduled sampling modes, and transfer the pre-trained knowledge to a student model that performs free running decoding. We show that the MT-KD network provides an adequate platform for neural TTS training, where the student model learns to emulate the behaviors of the two teachers, at the same time, minimizing the mismatch between training and run-time inference. Experiments on both Chinese and English data show that MT-KD system consistently outperforms the competitive baselines in terms of naturalness, robustness and expressiveness for in-domain and out-of-domain test data. Furthermore, we show that knowledge distillation outperforms adversarial learning and data augmentation in addressing the exposure bias problem.
- 3 May 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 30 (23299290) , 1789-1802
Open AccessDecoding Knowledge Transfer for Neural Text-to-Speech Training
Neural end-to-end text-to-speech (TTS) is superior to conventional statistical methods in many ways. However, the exposure bias problem, that arises from the mismatch between the training and inference process in autoregressive models, remains an issue. It often leads to performance degradation in face of out-of-domain test data. To address this problem, we study a novel decoding knowledge transfer strategy, and propose a multi-teacher knowledge distillation (MT-KD) network for Tacotron2 TTS model. The idea is to pre-train two Tacotron2 TTS teacher models in teacher forcing and scheduled sampling modes, and transfer the pre-trained knowledge to a student model that performs free running decoding. We show that the MT-KD network provides an adequate platform for neural TTS training, where the student model learns to emulate the behaviors of the two teachers, at the same time, minimizing the mismatch between training and run-time inference. Experiments on both Chinese and English data show that MT-KD system consistently outperforms the competitive baselines in terms of naturalness, robustness and expressiveness for in-domain and out-of-domain test data. Furthermore, we show that knowledge distillation outperforms adversarial learning and data augmentation in addressing the exposure bias problem.