Berrak Sisman

Rui Liu, Berrak Sisman, Björn Schuller et al.

Emotion classification of speech and assessment of the emotion strength are required in applications such as emotional text-to-speech and voice conversion. The emotion attribute ranking function based on Support Vector Machine (SVM) was proposed to predict emotion strength for emotional speech corpus. However, the trained ranking function doesn't generalize to new domains, which limits the scope of applications, especially for out-of-domain or unseen speech. In this paper, we propose a data-driven deep learning model, i.e. StrengthNet, to improve the generalization of emotion strength assessment for seen and unseen speech. This is achieved by the fusion of emotional data from various domains. We follow a multi-task learning network architecture that includes an acoustic encoder, a strength predictor, and an auxiliary emotion predictor. Experiments show that the predicted emotion strength of the proposed StrengthNet is highly correlated with ground truth scores for both seen and unseen speech. We release the source codes at: https://github.com/ttslr/StrengthNet.

Kun Zhou, Berrak Sisman, Rajib Rana et al.

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.

Zongyang Du, Shreeram Suresh Chandra, Ismail Rasim Ulgen et al.

Everyday speech conveys far more than words, it reflects who we are, how we feel, and the circumstances surrounding our interactions. Yet, most existing speech datasets are acted, limited in scale, and fail to capture the expressive richness of real-life communication. With the rise of large neural networks, several large-scale speech corpora have emerged and been widely adopted across various speech processing tasks. However, the field of voice conversion (VC) still lacks large-scale, expressive, and real-life speech resources suitable for modeling natural prosody and emotion. To fill this gap, we release NaturalVoices (NV), the first large-scale spontaneous podcast dataset specifically designed for emotion-aware voice conversion. It comprises 5,049 hours of spontaneous podcast recordings with automatic annotations for emotion (categorical and attribute-based), speech quality, transcripts, speaker identity, and sound events. The dataset captures expressive emotional variation across thousands of speakers, diverse topics, and natural speaking styles. We also provide an open-source pipeline with modular annotation tools and flexible filtering, enabling researchers to construct customized subsets for a wide range of VC tasks. Experiments demonstrate that NaturalVoices supports the development of robust and generalizable VC models capable of producing natural, expressive speech, while revealing limitations of current architectures when applied to large-scale spontaneous data. These results suggest that NaturalVoices is both a valuable resource and a challenging benchmark for advancing the field of voice conversion. Dataset is available at: https://huggingface.co/JHU-SmileLab

Rui Liu, Berrak Sisman, Guanglai Gao et al.

Accented text-to-speech (TTS) synthesis seeks to generate speech with an accent (L2) as a variant of the standard version (L1). Accented TTS synthesis is challenging as L2 is different from L1 in both in terms of phonetic rendering and prosody pattern. Furthermore, there is no easy solution to the control of the accent intensity in an utterance. In this work, we propose a neural TTS architecture, that allows us to control the accent and its intensity during inference. This is achieved through three novel mechanisms, 1) an accent variance adaptor to model the complex accent variance with three prosody controlling factors, namely pitch, energy and duration; 2) an accent intensity modeling strategy to quantify the accent intensity; 3) a consistency constraint module to encourage the TTS system to render the expected accent intensity at a fine level. Experiments show that the proposed system attains superior performance to the baseline models in terms of accent rendering and intensity control. To our best knowledge, this is the first study of accented TTS synthesis with explicit intensity control.

Junchen Lu, Berrak Sisman, Mingyang Zhang et al.

The goal of Automatic Voice Over (AVO) is to generate speech in sync with a silent video given its text script. Recent AVO frameworks built upon text-to-speech synthesis (TTS) have shown impressive results. However, the current AVO learning objective of acoustic feature reconstruction brings in indirect supervision for inter-modal alignment learning, thus limiting the synchronization performance and synthetic speech quality. To this end, we propose a novel AVO method leveraging the learning objective of self-supervised discrete speech unit prediction, which not only provides more direct supervision for the alignment learning, but also alleviates the mismatch between the text-video context and acoustic features. Experimental results show that our proposed method achieves remarkable lip-speech synchronization and high speech quality by outperforming baselines in both objective and subjective evaluations. Code and speech samples are publicly available.

Jan Melechovsky, Ambuj Mehrish, Berrak Sisman et al.

Accent plays a significant role in speech communication, influencing one's capability to understand as well as conveying a person's identity. This paper introduces a novel and efficient framework for accented Text-to-Speech (TTS) synthesis based on a Conditional Variational Autoencoder. It has the ability to synthesize a selected speaker's voice, and convert this to any desired target accent. Our thorough experiments validate the effectiveness of the proposed framework using both objective and subjective evaluations. The results also show remarkable performance in terms of the model's ability to manipulate accents in the synthesized speech. Overall, our proposed framework presents a promising avenue for future accented TTS research.

Perry Lam, Huayun Zhang, Nancy F. Chen et al.

Neural models are known to be over-parameterized, and recent work has shown that sparse text-to-speech (TTS) models can outperform dense models. Although a plethora of sparse methods has been proposed for other domains, such methods have rarely been applied in TTS. In this work, we seek to answer the question: what are the characteristics of selected sparse techniques on the performance and model complexity? We compare a Tacotron2 baseline and the results of applying five techniques. We then evaluate the performance via the factors of naturalness, intelligibility and prosody, while reporting model size and training time. Complementary to prior research, we find that pruning before or during training can achieve similar performance to pruning after training and can be trained much faster, while removing entire neurons degrades performance much more than removing parameters. To our best knowledge, this is the first work that compares sparsity paradigms in text-to-speech synthesis.

Philip H. Lee, Ismail Rasim Ulgen, Berrak Sisman

Voice conversion (VC) aims to modify the speaker's identity while preserving the linguistic content. Commonly, VC methods use an encoder-decoder architecture, where disentangling the speaker's identity from linguistic information is crucial. However, the disentanglement approaches used in these methods are limited as the speaker features depend on the phonetic content of the utterance, compromising disentanglement. This dependency is amplified with attention-based methods. To address this, we introduce a novel masking mechanism in the input before speaker encoding, masking certain discrete speech units that correspond highly with phoneme classes. Our work aims to reduce the phonetic dependency of speaker features by restricting access to some phonetic information. Furthermore, since our approach is at the input level, it is applicable to any encoder-decoder based VC framework. Our approach improves disentanglement and conversion performance across multiple VC methods, showing significant effectiveness, particularly in attention-based method, with 44% relative improvement in objective intelligibility.

Perry Lam, Huayun Zhang, Nancy F. Chen et al.

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 2x 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 are available online.

Ismail Rasim Ulgen, Shreeram Suresh Chandra, Junchen Lu et al.

Synthesizing the voices of unseen speakers remains a persisting challenge in multi-speaker text-to-speech (TTS). Existing methods model speaker characteristics through speaker conditioning during training, leading to increased model complexity and limiting reproducibility and accessibility. A low-complexity alternative would broaden the reach of speech synthesis research, particularly in settings with limited computational and data resources. To this end, we propose SelectTTS, a simple and effective alternative. SelectTTS selects appropriate frames from the target speaker and decodes them using frame-level self-supervised learning (SSL) features. We demonstrate that this approach can effectively capture speaker characteristics for unseen speakers and achieves performance comparable to state-of-the-art multi-speaker TTS frameworks on both objective and subjective metrics. By directly selecting frames from the target speaker's speech, SelectTTS enables generalization to unseen speakers with significantly lower model complexity. Experimental results show that the proposed approach achieves performance comparable to state-of-the-art systems such as XTTS-v2 and VALL-E, while requiring over 8x fewer parameters and 270x less training data. Moreover, it demonstrates that frame selection with SSL features offers an efficient path to low-complexity, high-quality multi-speaker TTS.

Rixi Xu, Qingyu Liu, Haitao Li et al.

In this paper, we present X-Voice, a 0.4B multilingual zero-shot voice cloning model that clones arbitrary voices and enables everyone to speak 30 languages. X-Voice is trained on a 420K-hour multilingual corpus using the International Phonetic Alphabet (IPA) as a unified representation. To eliminate the reliance on prompt text without complex preprocessing like forced alignment, we design a two-stage training paradigm. In Stage 1, we establish X-Voice$_{\text{s1}}$ through standard conditional flow-matching training and use it to synthesize 10K hours of speaker-consistent segments as audio prompts. In Stage 2, we fine-tune on these audio pairs with prompt text masked to derive X-Voice$_{\text{s2}}$, which enables zero-shot voice cloning without requiring transcripts of audio prompts. Architecturally, we extend F5-TTS by implementing a dual-level injection of language identifiers and decoupling and scheduling of Classifier-Free Guidance to facilitate multilingual speech synthesis. Subjective and objective evaluation results demonstrate that X-Voice outperforms existing flow-matching based multilingual systems like LEMAS-TTS and achieves zero-shot cross-lingual cloning capabilities comparable to billion-scale models such as Qwen3-TTS. To facilitate research transparency and community advancement, we open-source all related resources.

Xiutian Zhao, Ismail Rasim Ulgen, Philipp Koehn et al.

Large audio-language models (LALMs) can produce expressive speech, yet reliable emotion control remains elusive: conversions often miss the target affect and may degrade linguistic fidelity through refusals, hallucinations, or paraphrase. We present, to our knowledge, the first neuron-level study of emotion control in speech-generative LALMs and demonstrate that compact emotion-sensitive neurons (ESNs) are causally actionable, enabling training-free emotion steering at inference time. ESNs are identified via success-filtered activation aggregation enforcing both emotion realization and content preservation. Across three LALMs (Qwen2.5-Omni-7B, MiniCPM-o 4.5, Kimi-Audio), ESN interventions yield emotion-specific gains that generalize to unseen speakers and are supported by automatic and human evaluation. Controllability depends on selector design, mask sparsity, filtering, and intervention strength. Our results establish a mechanistic framework for training-free emotion control in speech generation.

Ismail Rasim Ulgen, John H. L. Hansen, Carlos Busso et al.

Modeling the rich prosodic variations inherent in human speech is essential for generating natural-sounding speech. While speaker embeddings are commonly used as conditioning inputs in personalized speech generation, they are typically optimized for speaker recognition, which encourages the loss of intra-speaker variation. This strategy makes them suboptimal for speech generation in terms of modeling the rich variations at the output speech distribution. In this work, we propose a novel speaker embedding network that employs multiple sub-centers per speaker class during training, instead of a single center as in conventional approaches. This sub-center modeling allows the embedding to capture a broader range of speaker-specific variations while maintaining speaker classification performance. We demonstrate the effectiveness of the proposed embeddings on a voice conversion task, showing improved naturalness and prosodic expressiveness in the synthesized speech.

Zhenqi Jia, Rui Liu, Berrak Sisman et al.

Conversational Speech Synthesis (CSS) aims to generate speech with natural prosody by understanding the multimodal dialogue history (MDH). The latest work predicts the accurate prosody expression of the target utterance by modeling the utterance-level interaction characteristics of MDH and the target utterance. However, MDH contains fine-grained semantic and prosody knowledge at the word level. Existing methods overlook the fine-grained semantic and prosodic interaction modeling. To address this gap, we propose MFCIG-CSS, a novel Multimodal Fine-grained Context Interaction Graph-based CSS system. Our approach constructs two specialized multimodal fine-grained dialogue interaction graphs: a semantic interaction graph and a prosody interaction graph. These two interaction graphs effectively encode interactions between word-level semantics, prosody, and their influence on subsequent utterances in MDH. The encoded interaction features are then leveraged to enhance synthesized speech with natural conversational prosody. Experiments on the DailyTalk dataset demonstrate that MFCIG-CSS outperforms all baseline models in terms of prosodic expressiveness. Code and speech samples are available at https://github.com/AI-S2-Lab/MFCIG-CSS.

Huhong Xian, Rui Liu, Berrak Sisman et al.

Different from traditional sentence-level audio deepfake detection (ADD), partial audio deepfake detection (PADD) requires frame-level positioning of the location of fake speech. While some progress has been made in this area, leveraging semantic information from audio, especially named entities, remains an underexplored aspect. To this end, we propose NE-PADD, a novel method for Partial Audio Deepfake Detection (PADD) that leverages named entity knowledge through two parallel branches: Speech Name Entity Recognition (SpeechNER) and PADD. The approach incorporates two attention aggregation mechanisms: Attention Fusion (AF) for combining attention weights and Attention Transfer (AT) for guiding PADD with named entity semantics using an auxiliary loss. Built on the PartialSpoof-NER dataset, experiments show our method outperforms existing baselines, proving the effectiveness of integrating named entity knowledge in PADD. The code is available at https://github.com/AI-S2-Lab/NE-PADD.

Rui Liu, Pu Gao, Jiatian Xi et al.

Text-based speech editing (TSE) modifies speech using only text, eliminating re-recording. However, existing TSE methods, mainly focus on the content accuracy and acoustic consistency of synthetic speech segments, and often overlook the emotional shifts or inconsistency issues introduced by text changes. To address this issue, we propose EmoCorrector, a novel post-correction scheme for TSE. EmoCorrector leverages Retrieval-Augmented Generation (RAG) by extracting the edited text's emotional features, retrieving speech samples with matching emotions, and synthesizing speech that aligns with the desired emotion while preserving the speaker's identity and quality. To support the training and evaluation of emotional consistency modeling in TSE, we pioneer the benchmarking Emotion Correction Dataset for TSE (ECD-TSE). The prominent aspect of ECD-TSE is its inclusion of $<$text, speech$>$ paired data featuring diverse text variations and a range of emotional expressions. Subjective and objective experiments and comprehensive analysis on ECD-TSE confirm that EmoCorrector significantly enhances the expression of intended emotion while addressing emotion inconsistency limitations in current TSE methods. Code and audio examples are available at https://github.com/AI-S2-Lab/EmoCorrector.

Rui Liu, Berrak Sisman, Haizhou Li

Recently, emotional speech synthesis has achieved remarkable performance. The emotion strength of synthesized speech can be controlled flexibly using a strength descriptor, which is obtained by an emotion attribute ranking function. However, a trained ranking function on specific data has poor generalization, which limits its applicability for more realistic cases. In this paper, we propose a deep learning based emotion strength assessment network for strength prediction that is referred to as StrengthNet. Our model conforms to a multi-task learning framework with a structure that includes an acoustic encoder, a strength predictor and an auxiliary emotion predictor. A data augmentation strategy was utilized to improve the model generalization. Experiments show that the predicted emotion strength of the proposed StrengthNet are highly correlated with ground truth scores for seen and unseen speech. Our codes are available at: https://github.com/ttslr/StrengthNet.

Ismail Rasim Ulgen, Zexin Cai, Nicholas Andrews et al.

To preserve or not to preserve prosody is a central question in voice anonymization. Prosody conveys meaning and affect, yet is tightly coupled with speaker identity. Existing methods either discard prosody for privacy or lack a principled mechanism to control the utility-privacy trade-off, operating at fixed design points. We propose DiffAnon, a diffusion-based anonymization method with classifier-free guidance (CFG) that provides explicit, continuous inference-time control over prosody preservation. DiffAnon refines acoustic detail over semantic embeddings of an RVQ codec, enabling smooth interpolation between anonymization strength and prosodic fidelity within a single model. To the best of our knowledge, it is the first voice anonymization framework to provide structured, interpolatable inference-time prosody control. Experiments demonstrate structured trade-off behavior, achieving strong utility while maintaining competitive privacy across controllable operating points.

Jan Melechovsky, Ambuj Mehrish, Berrak Sisman et al.

Recent advancements in Text-to-Speech (TTS) systems have enabled the generation of natural and expressive speech from textual input. Accented TTS aims to enhance user experience by making the synthesized speech more relatable to minority group listeners, and useful across various applications and context. Speech synthesis can further be made more flexible by allowing users to choose any combination of speaker identity and accent, resulting in a wide range of personalized speech outputs. Current models struggle to disentangle speaker and accent representation, making it difficult to accurately imitate different accents while maintaining the same speaker characteristics. We propose a novel approach to disentangle speaker and accent representations using multi-level variational autoencoders (ML-VAE) and vector quantization (VQ) to improve flexibility and enhance personalization in speech synthesis. Our proposed method addresses the challenge of effectively separating speaker and accent characteristics, enabling more fine-grained control over the synthesized speech. Code and speech samples are publicly available.

Aurosweta Mahapatra, Ismail Rasim Ulgen, Abinay Reddy Naini et al.

Traditional anti-spoofing focuses on models and datasets built on synthetic speech with mostly neutral state, neglecting diverse emotional variations. As a result, their robustness against high-quality, emotionally expressive synthetic speech is uncertain. We address this by introducing EmoSpoof-TTS, a corpus of emotional text-to-speech samples. Our analysis shows existing anti-spoofing models struggle with emotional synthetic speech, exposing risks of emotion-targeted attacks. Even trained on emotional data, the models underperform due to limited focus on emotional aspect and show performance disparities across emotions. This highlights the need for emotion-focused anti-spoofing paradigm in both dataset and methodology. We propose GEM, a gated ensemble of emotion-specialized models with a speech emotion recognition gating network. GEM performs effectively across all emotions and neutral state, improving defenses against spoofing attacks. We release the EmoSpoof-TTS Dataset: https://emospoof-tts.github.io/Dataset/

Shreeram Suresh Chandra, Zongyang Du, Berrak Sisman

Many frameworks for emotional text-to-speech (E-TTS) rely on human-annotated emotion labels that are often inaccurate and difficult to obtain. Learning emotional prosody implicitly presents a tough challenge due to the subjective nature of emotions. In this study, we propose a novel approach that leverages text awareness to acquire emotional styles without the need for explicit emotion labels or text prompts. We present TEMOTTS, a two-stage framework for E-TTS that is trained without emotion labels and is capable of inference without auxiliary inputs. Our proposed method performs knowledge transfer between the linguistic space learned by BERT and the emotional style space constructed by global style tokens. Our experimental results demonstrate the effectiveness of our proposed framework, showcasing improvements in emotional accuracy and naturalness. This is one of the first studies to leverage the emotional correlation between spoken content and expressive delivery for emotional TTS.

Shreeram Suresh Chandra, Lucas Goncalves, Junchen Lu et al.

Current emotion-based contrastive language-audio pretraining (CLAP) methods typically learn by naïvely aligning audio samples with corresponding text prompts. Consequently, this approach fails to capture the ordinal nature of emotions, hindering inter-emotion understanding and often resulting in a wide modality gap between the audio and text embeddings due to insufficient alignment. To handle these drawbacks, we introduce EmotionRankCLAP, a supervised contrastive learning approach that uses dimensional attributes of emotional speech and natural language prompts to jointly capture fine-grained emotion variations and improve cross-modal alignment. Our approach utilizes a Rank-N-Contrast objective to learn ordered relationships by contrasting samples based on their rankings in the valence-arousal space. EmotionRankCLAP outperforms existing emotion-CLAP methods in modeling emotion ordinality across modalities, measured via a cross-modal retrieval task.

Thejan Rajapakshe, Rajib Rana, Sara Khalifa et al.

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.

Aurosweta Mahapatra, Ismail Rasim Ulgen, Berrak Sisman

Current anti-spoofing systems remain vulnerable to expressive and emotional synthetic speech, since they rarely leverage prosody as a discriminative cue. Prosody is central to human expressiveness and emotion, and humans instinctively use prosodic cues such as F0 patterns and voiced/unvoiced structure to distinguish natural from synthetic speech. In this paper, we propose HuLA, a two-stage prosody-aware multi-task learning framework for spoof detection. In Stage 1, a self-supervised learning (SSL) backbone is trained on real speech with auxiliary tasks of F0 prediction and voiced/unvoiced classification, enhancing its ability to capture natural prosodic variation similar to human perceptual learning. In Stage 2, the model is jointly optimized for spoof detection and prosody tasks on both real and synthetic data, leveraging prosodic awareness to detect mismatches between natural and expressive synthetic speech. Experiments show that HuLA consistently outperforms strong baselines on challenging out-of-domain dataset, including expressive, emotional, and cross-lingual attacks. These results demonstrate that explicit prosodic supervision, combined with SSL embeddings, substantially improves robustness against advanced synthetic speech attacks.

Ismail Rasim Ulgen, Zongyang Du, Junchen Lu et al.

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 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.

Ahad Jawaid, Shreeram Suresh Chandra, Junchen Lu et al.

Recent advances in style transfer text-to-speech (TTS) have improved the expressiveness of synthesized speech. However, encoding stylistic information (e.g., timbre, emotion, and prosody) from diverse and unseen reference speech remains a challenge. This paper introduces StyleMoE, an approach that addresses the issue of learning averaged style representations in the style encoder by creating style experts that learn from subsets of data. The proposed method replaces the style encoder in a TTS framework with a Mixture of Experts (MoE) layer. The style experts specialize by learning from subsets of reference speech routed to them by the gating network, enabling them to handle different aspects of the style space. As a result, StyleMoE improves the style coverage of the style encoder for style transfer TTS. Our experiments, both objective and subjective, demonstrate improved style transfer for diverse and unseen reference speech. The proposed method enhances the performance of existing state-of-the-art style transfer TTS models and represents the first study of style MoE in TTS.

Jan Melechovsky, Ambuj Mehrish, Berrak Sisman et al.

With rapid globalization, the need to build inclusive and representative speech technology cannot be overstated. Accent is an important aspect of speech that needs to be taken into consideration while building inclusive speech synthesizers. Inclusive speech technology aims to erase any biases towards specific groups, such as people of certain accent. We note that state-of-the-art Text-to-Speech (TTS) systems may currently not be suitable for all people, regardless of their background, as they are designed to generate high-quality voices without focusing on accent. In this paper, we propose a TTS model that utilizes a Multi-Level Variational Autoencoder with adversarial learning to address accented speech synthesis and conversion in TTS, with a vision for more inclusive systems in the future. We evaluate the performance through both objective metrics and subjective listening tests. The results show an improvement in accent conversion ability compared to the baseline.

Ismail Rasim Ulgen, Zongyang Du, Carlos Busso et al.

Speaker embeddings carry valuable emotion-related information, which makes them a promising resource for enhancing speech emotion recognition (SER), especially with limited labeled data. Traditionally, it has been assumed that emotion information is indirectly embedded within speaker embeddings, leading to their under-utilization. Our study reveals a direct and useful link between emotion and state-of-the-art speaker embeddings in the form of intra-speaker clusters. By conducting a thorough clustering analysis, we demonstrate that emotion information can be readily extracted from speaker embeddings. In order to leverage this information, we introduce a novel contrastive pretraining approach applied to emotion-unlabeled data for speech emotion recognition. The proposed approach involves the sampling of positive and the negative examples based on the intra-speaker clusters of speaker embeddings. The proposed strategy, which leverages extensive emotion-unlabeled data, leads to a significant improvement in SER performance, whether employed as a standalone pretraining task or integrated into a multi-task pretraining setting.

Thejan Rajapakshe, Rajib Rana, Sara Khalifa et al.

Speech Emotion Recognition (SER) is a critical enabler of emotion-aware communication in human-computer interactions. Recent advancements in Deep Learning (DL) have substantially enhanced the performance of SER models through increased model complexity. However, designing optimal DL architectures requires prior experience and experimental evaluations. Encouragingly, Neural Architecture Search (NAS) offers a promising avenue to determine an optimal DL model automatically. In particular, Differentiable Architecture Search (DARTS) is an efficient method of using NAS to search for optimised models. This paper proposes a DARTS-optimised joint CNN and LSTM architecture, to improve SER performance, where the literature informs the selection of CNN and LSTM coupling to offer improved performance. While DARTS has previously been applied to CNN and LSTM combinations, our approach introduces a novel mechanism, particularly in selecting CNN operations using DARTS. In contrast to previous studies, we refrain from imposing constraints on the order of the layers for the CNN within the DARTS cell; instead, we allow DARTS to determine the optimal layer order autonomously. Experimenting with the IEMOCAP and MSP-IMPROV datasets, we demonstrate that our proposed methodology achieves significantly higher SER accuracy than hand-engineering the CNN-LSTM configuration. It also outperforms the best-reported SER results achieved using DARTS on CNN-LSTM.

Lucas Goncalves, Seong-Gyun Leem, Wei-Cheng Lin et al.

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.

Kun Zhou, Berrak Sisman, Rajib Rana et al.

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.

Zongyang Du, Berrak Sisman, Kun Zhou et al.

Expressive voice conversion performs identity conversion for emotional speakers by jointly converting speaker identity and emotional style. Due to the hierarchical structure of speech emotion, it is challenging to disentangle the emotional style for different speakers. Inspired by the recent success of speaker disentanglement with variational autoencoder (VAE), we propose an any-to-any expressive voice conversion framework, that is called StyleVC. StyleVC is designed to disentangle linguistic content, speaker identity, pitch, and emotional style information. We study the use of style encoder to model emotional style explicitly. At run-time, StyleVC converts both speaker identity and emotional style for arbitrary speakers. Experiments validate the effectiveness of our proposed framework in both objective and subjective evaluations.

Sergey Nikonorov, Berrak Sisman, Mingyang Zhang et al.

Conventional vocoders are commonly used as analysis tools to provide interpretable features for downstream tasks such as speech synthesis and voice conversion. They are built under certain assumptions about the signals following signal processing principle, therefore, not easily generalizable to different audio, for example, from speech to singing. In this paper, we propose a deep neural analyzer, denoted as DeepA - a neural vocoder that extracts F0 and timbre/aperiodicity encoding from the input speech that emulate those defined in conventional vocoders. Therefore, the resulting parameters are more interpretable than other latent neural representations. At the same time, as the deep neural analyzer is learnable, it is expected to be more accurate for signal reconstruction and manipulation, and generalizable from speech to singing. The proposed neural analyzer is built based on a variational autoencoder (VAE) architecture. We show that DeepA improves F0 estimation over the conventional vocoder (WORLD). To our best knowledge, this is the first study dedicated to the development of a neural framework for extracting learnable vocoder-like parameters.

Junchen Lu, Berrak Sisman, Rui Liu et al.

In this paper, we formulate a novel task to synthesize speech in sync with a silent pre-recorded video, denoted as automatic voice over (AVO). Unlike traditional speech synthesis, AVO seeks to generate not only human-sounding speech, but also perfect lip-speech synchronization. A natural solution to AVO is to condition the speech rendering on the temporal progression of lip sequence in the video. We propose a novel text-to-speech model that is conditioned on visual input, named VisualTTS, for accurate lip-speech synchronization. The proposed VisualTTS adopts two novel mechanisms that are 1) textual-visual attention, and 2) visual fusion strategy during acoustic decoding, which both contribute to forming accurate alignment between the input text content and lip motion in input lip sequence. Experimental results show that VisualTTS achieves accurate lip-speech synchronization and outperforms all baseline systems.

Zongyang Du, Berrak Sisman, Kun Zhou et al.

Traditional voice conversion(VC) has been focused on speaker identity conversion for speech with a neutral expression. We note that emotional expression plays an essential role in daily communication, and the emotional style of speech can be speaker-dependent. In this paper, we study the technique to jointly convert the speaker identity and speaker-dependent emotional style, that is called expressive voice conversion. We propose a StarGAN-based framework to learn a many-to-many mapping across different speakers, that takes into account speaker-dependent emotional style without the need for parallel data. To achieve this, we condition the generator on emotional style encoding derived from a pre-trained speech emotion recognition(SER) model. The experiments validate the effectiveness of our proposed framework in both objective and subjective evaluations. To our best knowledge, this is the first study on expressive voice conversion.

Kun Zhou, Berrak Sisman, Rui Liu et al.

In this paper, we first provide a review of the state-of-the-art emotional voice conversion research, and the existing emotional speech databases. We then motivate the development of a novel emotional speech database (ESD) that addresses the increasing research need. With this paper, the ESD database is now made available to the research community. The ESD database consists of 350 parallel utterances spoken by 10 native English and 10 native Chinese speakers and covers 5 emotion categories (neutral, happy, angry, sad and surprise). More than 29 hours of speech data were recorded in a controlled acoustic environment. The database is suitable for multi-speaker and cross-lingual emotional voice conversion studies. As case studies, we implement several state-of-the-art emotional voice conversion systems on the ESD database. This paper provides a reference study on ESD in conjunction with its release.

Rui Liu, Berrak Sisman, Haizhou Li

Emotional text-to-speech synthesis (ETTS) has seen much progress in recent years. However, the generated voice is often not perceptually identifiable by its intended emotion category. To address this problem, we propose a new interactive training paradigm for ETTS, denoted as i-ETTS, which seeks to directly improve the emotion discriminability by interacting with a speech emotion recognition (SER) model. Moreover, we formulate an iterative training strategy with reinforcement learning to ensure the quality of i-ETTS optimization. Experimental results demonstrate that the proposed i-ETTS outperforms the state-of-the-art baselines by rendering speech with more accurate emotion style. To our best knowledge, this is the first study of reinforcement learning in emotional text-to-speech synthesis.

Kun Zhou, Berrak Sisman, Haizhou Li

Emotional voice conversion (EVC) aims to change the emotional state of an utterance while preserving the linguistic content and speaker identity. In this paper, we propose a novel 2-stage training strategy for sequence-to-sequence emotional voice conversion with a limited amount of emotional speech data. We note that the proposed EVC framework leverages text-to-speech (TTS) as they share a common goal that is to generate high-quality expressive voice. In stage 1, we perform style initialization with a multi-speaker TTS corpus, to disentangle speaking style and linguistic content. In stage 2, we perform emotion training with a limited amount of emotional speech data, to learn how to disentangle emotional style and linguistic information from the speech. The proposed framework can perform both spectrum and prosody conversion and achieves significant improvement over the state-of-the-art baselines in both objective and subjective evaluation.

Kun Zhou, Berrak Sisman, Haizhou Li

Emotional voice conversion (EVC) aims to convert the emotion of speech from one state to another while preserving the linguistic content and speaker identity. In this paper, we study the disentanglement and recomposition of emotional elements in speech through variational autoencoding Wasserstein generative adversarial network (VAW-GAN). We propose a speaker-dependent EVC framework based on VAW-GAN, that includes two VAW-GAN pipelines, one for spectrum conversion, and another for prosody conversion. We train a spectral encoder that disentangles emotion and prosody (F0) information from spectral features; we also train a prosodic encoder that disentangles emotion modulation of prosody (affective prosody) from linguistic prosody. At run-time, the decoder of spectral VAW-GAN is conditioned on the output of prosodic VAW-GAN. The vocoder takes the converted spectral and prosodic features to generate the target emotional speech. Experiments validate the effectiveness of our proposed method in both objective and subjective evaluations.

Kun Zhou, Berrak Sisman, Rui Liu et al.

Emotional voice conversion aims to transform emotional prosody in speech while preserving the linguistic content and speaker identity. Prior studies show that it is possible to disentangle emotional prosody using an encoder-decoder network conditioned on discrete representation, such as one-hot emotion labels. Such networks learn to remember a fixed set of emotional styles. In this paper, we propose a novel framework based on variational auto-encoding Wasserstein generative adversarial network (VAW-GAN), which makes use of a pre-trained speech emotion recognition (SER) model to transfer emotional style during training and at run-time inference. In this way, the network is able to transfer both seen and unseen emotional style to a new utterance. We show that the proposed framework achieves remarkable performance by consistently outperforming the baseline framework. This paper also marks the release of an emotional speech dataset (ESD) for voice conversion, which has multiple speakers and languages.

Rui Liu, Berrak Sisman, Haizhou Li

Attention-based end-to-end text-to-speech synthesis (TTS) is superior to conventional statistical methods in many ways. Transformer-based TTS is one of such successful implementations. While Transformer TTS models the speech frame sequence well with a self-attention mechanism, it does not associate input text with output utterances from a syntactic point of view at sentence level. We propose a novel neural TTS model, denoted as GraphSpeech, that is formulated under graph neural network framework. GraphSpeech encodes explicitly the syntactic relation of input lexical tokens in a sentence, and incorporates such information to derive syntactically motivated character embeddings for TTS attention mechanism. Experiments show that GraphSpeech consistently outperforms the Transformer TTS baseline in terms of spectrum and prosody rendering of utterances.

Rui Liu, Berrak Sisman, Feilong Bao et al.

Tacotron-based end-to-end speech synthesis has shown remarkable voice quality. However, the rendering of prosody in the synthesized speech remains to be improved, especially for long sentences, where prosodic phrasing errors can occur frequently. In this paper, we extend the Tacotron-based speech synthesis framework to explicitly model the prosodic phrase breaks. We propose a multi-task learning scheme for Tacotron training, that optimizes the system to predict both Mel spectrum and phrase breaks. To our best knowledge, this is the first implementation of multi-task learning for Tacotron based TTS with a prosodic phrasing model. Experiments show that our proposed training scheme consistently improves the voice quality for both Chinese and Mongolian systems.

Zongyang Du, Kun Zhou, Berrak Sisman et al.

Cross-lingual voice conversion aims to change source speaker's voice to sound like that of target speaker, when source and target speakers speak different languages. It relies on non-parallel training data from two different languages, hence, is more challenging than mono-lingual voice conversion. Previous studies on cross-lingual voice conversion mainly focus on spectral conversion with a linear transformation for F0 transfer. However, as an important prosodic factor, F0 is inherently hierarchical, thus it is insufficient to just use a linear method for conversion. We propose the use of continuous wavelet transform (CWT) decomposition for F0 modeling. CWT provides a way to decompose a signal into different temporal scales that explain prosody in different time resolutions. We also propose to train two CycleGAN pipelines for spectrum and prosody mapping respectively. In this way, we eliminate the need for parallel data of any two languages and any alignment techniques. Experimental results show that our proposed Spectrum-Prosody-CycleGAN framework outperforms the Spectrum-CycleGAN baseline in subjective evaluation. To our best knowledge, this is the first study of prosody in cross-lingual voice conversion.

Junchen Lu, Kun Zhou, Berrak Sisman et al.

Singing voice conversion aims to convert singer's voice from source to target without changing singing content. Parallel training data is typically required for the training of singing voice conversion system, that is however not practical in real-life applications. Recent encoder-decoder structures, such as variational autoencoding Wasserstein generative adversarial network (VAW-GAN), provide an effective way to learn a mapping through non-parallel training data. In this paper, we propose a singing voice conversion framework that is based on VAW-GAN. We train an encoder to disentangle singer identity and singing prosody (F0 contour) from phonetic content. By conditioning on singer identity and F0, the decoder generates output spectral features with unseen target singer identity, and improves the F0 rendering. Experimental results show that the proposed framework achieves better performance than the baseline frameworks.

Berrak Sisman, Junichi Yamagishi, Simon King et al.

Speaker identity is one of the important characteristics of human speech. In voice conversion, we change the speaker identity from one to another, while keeping the linguistic content unchanged. Voice conversion involves multiple speech processing techniques, such as speech analysis, spectral conversion, prosody conversion, speaker characterization, and vocoding. With the recent advances in theory and practice, we are now able to produce human-like voice quality with high speaker similarity. In this paper, we provide a comprehensive overview of the state-of-the-art of voice conversion techniques and their performance evaluation methods from the statistical approaches to deep learning, and discuss their promise and limitations. We will also report the recent Voice Conversion Challenges (VCC), the performance of the current state of technology, and provide a summary of the available resources for voice conversion research.

Rui Liu, Berrak Sisman, Guanglai Gao et al.

We propose a novel training strategy for Tacotron-based text-to-speech (TTS) system to improve the expressiveness of speech. One of the key challenges in prosody modeling is the lack of reference that makes explicit modeling difficult. The proposed technique doesn't require prosody annotations from training data. It doesn't attempt to model prosody explicitly either, but rather encodes the association between input text and its prosody styles using a Tacotron-based TTS framework. Our proposed idea marks a departure from the style token paradigm where prosody is explicitly modeled by a bank of prosody embeddings. The proposed training strategy adopts a combination of two objective functions: 1) frame level reconstruction loss, that is calculated between the synthesized and target spectral features; 2) utterance level style reconstruction loss, that is calculated between the deep style features of synthesized and target speech. The proposed style reconstruction loss is formulated as a perceptual loss to ensure that utterance level speech style is taken into consideration during training. Experiments show that the proposed training strategy achieves remarkable performance and outperforms a state-of-the-art baseline in both naturalness and expressiveness. To our best knowledge, this is the first study to incorporate utterance level perceptual quality as a loss function into Tacotron training for improved expressiveness.

Kun Zhou, Berrak Sisman, Mingyang Zhang et al.

Emotional voice conversion aims to convert the emotion of speech from one state to another while preserving the linguistic content and speaker identity. The prior studies on emotional voice conversion are mostly carried out under the assumption that emotion is speaker-dependent. We consider that there is a common code between speakers for emotional expression in a spoken language, therefore, a speaker-independent mapping between emotional states is possible. In this paper, we propose a speaker-independent emotional voice conversion framework, that can convert anyone's emotion without the need for parallel data. We propose a VAW-GAN based encoder-decoder structure to learn the spectrum and prosody mapping. We perform prosody conversion by using continuous wavelet transform (CWT) to model the temporal dependencies. We also investigate the use of F0 as an additional input to the decoder to improve emotion conversion performance. Experiments show that the proposed speaker-independent framework achieves competitive results for both seen and unseen speakers.

Rui Liu, Berrak Sisman, Feilong Bao et al.

Tacotron-based text-to-speech (TTS) systems directly synthesize speech from text input. Such frameworks typically consist of a feature prediction network that maps character sequences to frequency-domain acoustic features, followed by a waveform reconstruction algorithm or a neural vocoder that generates the time-domain waveform from acoustic features. As the loss function is usually calculated only for frequency-domain acoustic features, that doesn't directly control the quality of the generated time-domain waveform. To address this problem, we propose a new training scheme for Tacotron-based TTS, referred to as WaveTTS, that has 2 loss functions: 1) time-domain loss, denoted as the waveform loss, that measures the distortion between the natural and generated waveform; and 2) frequency-domain loss, that measures the Mel-scale acoustic feature loss between the natural and generated acoustic features. WaveTTS ensures both the quality of the acoustic features and the resulting speech waveform. To our best knowledge, this is the first implementation of Tacotron with joint time-frequency domain loss. Experimental results show that the proposed framework outperforms the baselines and achieves high-quality synthesized speech.

Kun Zhou, Berrak Sisman, Haizhou Li

Emotional voice conversion aims to convert the spectrum and prosody to change the emotional patterns of speech, while preserving the speaker identity and linguistic content. Many studies require parallel speech data between different emotional patterns, which is not practical in real life. Moreover, they often model the conversion of fundamental frequency (F0) with a simple linear transform. As F0 is a key aspect of intonation that is hierarchical in nature, we believe that it is more adequate to model F0 in different temporal scales by using wavelet transform. We propose a CycleGAN network to find an optimal pseudo pair from non-parallel training data by learning forward and inverse mappings simultaneously using adversarial and cycle-consistency losses. We also study the use of continuous wavelet transform (CWT) to decompose F0 into ten temporal scales, that describes speech prosody at different time resolution, for effective F0 conversion. Experimental results show that our proposed framework outperforms the baselines both in objective and subjective evaluations.

Rui Liu, Berrak Sisman, Jingdong Li et al.

While neural end-to-end text-to-speech (TTS) is superior to conventional statistical methods in many ways, the exposure bias problem in the autoregressive models remains an issue to be resolved. The exposure bias problem arises from the mismatch between the training and inference process, that results in unpredictable performance for out-of-domain test data at run-time. To overcome this, we propose a teacher-student training scheme for Tacotron-based TTS by introducing a distillation loss function in addition to the feature loss function. We first train a Tacotron2-based TTS model by always providing natural speech frames to the decoder, that serves as a teacher model. We then train another Tacotron2-based model as a student model, of which the decoder takes the predicted speech frames as input, similar to how the decoder works during run-time inference. With the distillation loss, the student model learns the output probabilities from the teacher model, that is called knowledge distillation. Experiments show that our proposed training scheme consistently improves the voice quality for out-of-domain test data both in Chinese and English systems.