Existing large-scale zero-shot text-to-speech (TTS) models deliver high speech quality but suffer from slow inference speeds due to massive parameters. To address this issue, this paper introduces ZipVoice, a high-quality flow-matching-based zero-shot TTS model with a compact model size and fast inference speed. Key designs include: 1) a Zipformer-based flow-matching decoder to maintain adequate modeling capabilities under constrained size; 2) Average upsampling-based initial speech-text alignment and Zipformer-based text encoder to improve speech intelligibility; 3) A flow distillation method to reduce sampling steps and eliminate the inference overhead associated with classifier-free guidance. Experiments on 100k hours multilingual datasets show that ZipVoice matches state-of-the-art models in speech quality, while being 3 times smaller and up to 30 times faster than a DiT-based flow-matching baseline. Codes, model checkpoints and demo samples are publicly available.
Multilingual speech-to-speech translation (S2ST) aims to directly convert spoken utterances from multiple source languages into fluent and intelligible speech in a target language. Despite recent progress, several critical challenges persist: 1) achieving high-quality S2ST remains a significant obstacle; 2) most existing S2ST methods rely heavily on large-scale parallel speech corpora, which are difficult and resource-intensive to obtain. To tackle these challenges, we introduce S2ST-Omni, a novel, efficient, and scalable framework tailored for multilingual speech-to-speech translation. Specifically, we decompose S2ST into speech-to-text translation (S2TT) and text-to-speech synthesis (TTS). To enable high-quality S2TT while mitigating reliance on large-scale parallel speech corpora, we leverage powerful pretrained models: Whisper for robust audio understanding and Qwen 3.0 for advanced text comprehension. A lightweight speech adapter is introduced to bridge the modality gap between speech and text representations, facilitating effective utilization of pretrained multimodal knowledge. To ensure both translation accuracy and real-time responsiveness, we adopt a streaming speech generation model in the TTS stage, which generates the target speech in an autoregressive manner. Extensive experiments conducted on the CVSS benchmark demonstrate that S2ST-Omni consistently surpasses several state-of-the-art S2ST baselines in translation quality, highlighting its effectiveness and superiority.
Automated speaking assessment (ASA) on opinion expressions is often hampered by the scarcity of labeled recordings, which restricts prompt diversity and undermines scoring reliability. To address this challenge, we propose a novel training paradigm that leverages a large language models (LLM) to generate diverse responses of a given proficiency level, converts responses into synthesized speech via speaker-aware text-to-speech synthesis, and employs a dynamic importance loss to adaptively reweight training instances based on feature distribution differences between synthesized and real speech. Subsequently, a multimodal large language model integrates aligned textual features with speech signals to predict proficiency scores directly. Experiments conducted on the LTTC dataset show that our approach outperforms methods relying on real data or conventional augmentation, effectively mitigating low-resource constraints and enabling ASA on opinion expressions with cross-modal information.
Recent advancements in generative artificial intelligence have significantly transformed the field of style-captioned text-to-speech synthesis (CapTTS). However, adapting CapTTS to real-world applications remains challenging due to the lack of standardized, comprehensive datasets and limited research on downstream tasks built upon CapTTS. To address these gaps, we introduce CapSpeech, a new benchmark designed for a series of CapTTS-related tasks, including style-captioned text-to-speech synthesis with sound events (CapTTS-SE), accent-captioned TTS (AccCapTTS), emotion-captioned TTS (EmoCapTTS), and text-to-speech synthesis for chat agent (AgentTTS). CapSpeech comprises over 10 million machine-annotated audio-caption pairs and nearly 0.36 million human-annotated audio-caption pairs. In addition, we introduce two new datasets collected and recorded by a professional voice actor and experienced audio engineers, specifically for the AgentTTS and CapTTS-SE tasks. Alongside the datasets, we conduct comprehensive experiments using both autoregressive and non-autoregressive models on CapSpeech. Our results demonstrate high-fidelity and highly intelligible speech synthesis across a diverse range of speaking styles. To the best of our knowledge, CapSpeech is the largest available dataset offering comprehensive annotations for CapTTS-related tasks. The experiments and findings further provide valuable insights into the challenges of developing CapTTS systems.
Speech quality assessment is a critical process in selecting text-to-speech synthesis (TTS) or voice conversion models. Evaluation of voice synthesis can be done using objective metrics or subjective metrics. Although there are many objective metrics like the Perceptual Evaluation of Speech Quality (PESQ), Perceptual Objective Listening Quality Assessment (POLQA) or Short-Time Objective Intelligibility (STOI) but none of them is feasible in selecting the best model. On the other hand subjective metric like Mean Opinion Score is highly reliable but it requires a lot of manual efforts and are time-consuming. To counter the issues in MOS Evaluation, we have developed a novel model, Speaker Agnostic Latent Features (SALF)-Mean Opinion Score (MOS) which is a small-sized, end-to-end, highly generalized and scalable model for predicting MOS score on a scale of 5. We use the sequences of convolutions and stack them to get the latent features of the audio samples to get the best state-of-the-art results based on mean squared error (MSE), Linear Concordance Correlation coefficient (LCC), Spearman Rank Correlation Coefficient (SRCC) and Kendall Rank Correlation Coefficient (KTAU).
Unlike traditional cascaded pipelines, end-to-end (E2E) spoken dialogue systems preserve full differentiability and capture non-phonemic information, making them well-suited for modeling spoken interactions. However, existing E2E approaches often require large-scale training data and generates responses lacking semantic coherence. We propose a simple yet effective strategy leveraging a chain-of-thought (CoT) formulation, ensuring that training on conversational data remains closely aligned with the multimodal language model (LM)'s pre-training on speech recognition~(ASR), text-to-speech synthesis (TTS), and text LM tasks. Our method achieves over 1.5 ROUGE-1 improvement over the baseline, successfully training spoken dialogue systems on publicly available human-human conversation datasets, while being compute-efficient enough to train on just 300 hours of public human-human conversation data, such as the Switchboard. We will publicly release our models and training code.
This paper explores multi-modal controllable Text-to-Speech Synthesis (TTS) where the voice can be generated from face image, and the characteristics of output speech (e.g., pace, noise level, distance, tone, place) can be controllable with natural text description. Specifically, we aim to mitigate the following three challenges in face-driven TTS systems. 1) To overcome the limited audio quality of audio-visual speech corpora, we propose a training method that additionally utilizes high-quality audio-only speech corpora. 2) To generate voices not only from real human faces but also from artistic portraits, we propose augmenting the input face image with stylization. 3) To consider one-to-many possibilities in face-to-voice mapping and ensure consistent voice generation at the same time, we propose to first employ sampling-based decoding and then use prompting with generated speech samples. Experimental results validate the proposed model's effectiveness in face-driven voice synthesis.
Zero-shot streaming text-to-speech is an important research topic in human-computer interaction. Existing methods primarily use a lookahead mechanism, relying on future text to achieve natural streaming speech synthesis, which introduces high processing latency. To address this issue, we propose SMLLE, a streaming framework for generating high-quality speech frame-by-frame. SMLLE employs a Transducer to convert text into semantic tokens in real time while simultaneously obtaining duration alignment information. The combined outputs are then fed into a fully autoregressive (AR) streaming model to reconstruct mel-spectrograms. To further stabilize the generation process, we design a Delete < Bos > Mechanism that allows the AR model to access future text introducing as minimal delay as possible. Experimental results suggest that the SMLLE outperforms current streaming TTS methods and achieves comparable performance over sentence-level TTS systems. Samples are available on https://anonymous.4open.science/w/demo_page-48B7/.
The rise of Large Audio Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing AJailBench-Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories, converted from textual jailbreak attacks using realistic text to speech synthesis. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms.
Tibetan is a low-resource language with minimal parallel speech corpora spanning its three major dialects-\"U-Tsang, Amdo, and Kham-limiting progress in speech modeling. To address this issue, we propose FMSD-TTS, a few-shot, multi-speaker, multi-dialect text-to-speech framework that synthesizes parallel dialectal speech from limited reference audio and explicit dialect labels. Our method features a novel speaker-dialect fusion module and a Dialect-Specialized Dynamic Routing Network (DSDR-Net) to capture fine-grained acoustic and linguistic variations across dialects while preserving speaker identity. Extensive objective and subjective evaluations demonstrate that FMSD-TTS significantly outperforms baselines in both dialectal expressiveness and speaker similarity. We further validate the quality and utility of the synthesized speech through a challenging speech-to-speech dialect conversion task. Our contributions include: (1) a novel few-shot TTS system tailored for Tibetan multi-dialect speech synthesis, (2) the public release of a large-scale synthetic Tibetan speech corpus generated by FMSD-TTS, and (3) an open-source evaluation toolkit for standardized assessment of speaker similarity, dialect consistency, and audio quality.
We propose a shallow flow matching (SFM) mechanism to enhance flow matching (FM)-based text-to-speech (TTS) models within a coarse-to-fine generation paradigm. SFM constructs intermediate states along the FM paths using coarse output representations. During training, we introduce an orthogonal projection method to adaptively determine the temporal position of these states, and apply a principled construction strategy based on a single-segment piecewise flow. The SFM inference starts from the intermediate state rather than pure noise and focuses computation on the latter stages of the FM paths. We integrate SFM into multiple TTS models with a lightweight SFM head. Experiments show that SFM consistently improves the naturalness of synthesized speech in both objective and subjective evaluations, while significantly reducing inference when using adaptive-step ODE solvers. Demo and codes are available at https://ydqmkkx.github.io/SFMDemo/.
Recent works have shown that modelling raw waveform directly from text in an end-to-end (E2E) fashion produces more natural-sounding speech than traditional neural text-to-speech (TTS) systems based on a cascade or two-stage approach. However, current E2E state-of-the-art models are computationally complex and memory-consuming, making them unsuitable for real-time offline on-device applications in low-resource scenarios. To address this issue, we propose a Lightweight E2E-TTS (LE2E) model that generates high-quality speech requiring minimal computational resources. We evaluate the proposed model on the LJSpeech dataset and show that it achieves state-of-the-art performance while being up to $90\%$ smaller in terms of model parameters and $10\times$ faster in real-time-factor. Furthermore, we demonstrate that the proposed E2E training paradigm achieves better quality compared to an equivalent architecture trained in a two-stage approach. Our results suggest that LE2E is a promising approach for developing real-time, high quality, low-resource TTS applications for on-device applications.
This paper introduces a novel multi-Agent framework that automates the end to end production of Qinqiang opera by integrating Large Language Models , visual generation, and Text to Speech synthesis. Three specialized agents collaborate in sequence: Agent1 uses an LLM to craft coherent, culturally grounded scripts;Agent2 employs visual generation models to render contextually accurate stage scenes; and Agent3 leverages TTS to produce synchronized, emotionally expressive vocal performances. In a case study on Dou E Yuan, the system achieved expert ratings of 3.8 for script fidelity, 3.5 for visual coherence, and 3.8 for speech accuracy-culminating in an overall score of 3.6, a 0.3 point improvement over a Single Agent baseline. Ablation experiments demonstrate that removing Agent2 or Agent3 leads to drops of 0.4 and 0.5 points, respectively, underscoring the value of modular collaboration. This work showcases how AI driven pipelines can streamline and scale the preservation of traditional performing arts, and points toward future enhancements in cross modal alignment, richer emotional nuance, and support for additional opera genres.
Recent zero-shot text-to-speech (TTS) systems face a common dilemma: autoregressive (AR) models suffer from slow generation and lack duration controllability, while non-autoregressive (NAR) models lack temporal modeling and typically require complex designs. In this paper, we introduce a novel pseudo-autoregressive (PAR) codec language modeling approach that unifies AR and NAR modeling. Combining explicit temporal modeling from AR with parallel generation from NAR, PAR generates dynamic-length spans at fixed time steps. Building on PAR, we propose PALLE, a two-stage TTS system that leverages PAR for initial generation followed by NAR refinement. In the first stage, PAR progressively generates speech tokens along the time dimension, with each step predicting all positions in parallel but only retaining the left-most span. In the second stage, low-confidence tokens are iteratively refined in parallel, leveraging the global contextual information. Experiments demonstrate that PALLE, trained on LibriTTS, outperforms state-of-the-art systems trained on large-scale data, including F5-TTS, E2-TTS, and MaskGCT, on the LibriSpeech test-clean set in terms of speech quality, speaker similarity, and intelligibility, while achieving up to ten times faster inference speed. Audio samples are available at https://anonymous-palle.github.io.
With the advancement of speech synthesis technology, users have higher expectations for the naturalness and expressiveness of synthesized speech. But previous research ignores the importance of prompt selection. This study proposes a text-to-speech (TTS) framework based on Retrieval-Augmented Generation (RAG) technology, which can dynamically adjust the speech style according to the text content to achieve more natural and vivid communication effects. We have constructed a speech style knowledge database containing high-quality speech samples in various contexts and developed a style matching scheme. This scheme uses embeddings, extracted by Llama, PER-LLM-Embedder,and Moka, to match with samples in the knowledge database, selecting the most appropriate speech style for synthesis. Furthermore, our empirical research validates the effectiveness of the proposed method. Our demo can be viewed at: https://thuhcsi.github.io/icme2025-AutoStyle-TTS
Recent advances in text-to-speech synthesis have achieved notable success in generating high-quality short utterances for individual speakers. However, these systems still face challenges when extending their capabilities to long, multi-speaker, and spontaneous dialogues, typical of real-world scenarios such as podcasts. These limitations arise from two primary challenges: 1) long speech: podcasts typically span several minutes, exceeding the upper limit of most existing work; 2) spontaneity: podcasts are marked by their spontaneous, oral nature, which sharply contrasts with formal, written contexts; existing works often fall short in capturing this spontaneity. In this paper, we propose MoonCast, a solution for high-quality zero-shot podcast generation, aiming to synthesize natural podcast-style speech from text-only sources (e.g., stories, technical reports, news in TXT, PDF, or Web URL formats) using the voices of unseen speakers. To generate long audio, we adopt a long-context language model-based audio modeling approach utilizing large-scale long-context speech data. To enhance spontaneity, we utilize a podcast generation module to generate scripts with spontaneous details, which have been empirically shown to be as crucial as the text-to-speech modeling itself. Experiments demonstrate that MoonCast outperforms baselines, with particularly notable improvements in spontaneity and coherence.
ASVspoof 5 is the fifth edition in a series of challenges which promote the study of speech spoofing and deepfake attacks as well as the design of detection solutions. We introduce the ASVspoof 5 database which is generated in a crowdsourced fashion from data collected in diverse acoustic conditions (cf. studio-quality data for earlier ASVspoof databases) and from ~2,000 speakers (cf. ~100 earlier). The database contains attacks generated with 32 different algorithms, also crowdsourced, and optimised to varying degrees using new surrogate detection models. Among them are attacks generated with a mix of legacy and contemporary text-to-speech synthesis and voice conversion models, in addition to adversarial attacks which are incorporated for the first time. ASVspoof 5 protocols comprise seven speaker-disjoint partitions. They include two distinct partitions for the training of different sets of attack models, two more for the development and evaluation of surrogate detection models, and then three additional partitions which comprise the ASVspoof 5 training, development and evaluation sets. An auxiliary set of data collected from an additional 30k speakers can also be used to train speaker encoders for the implementation of attack algorithms. Also described herein is an experimental validation of the new ASVspoof 5 database using a set of automatic speaker verification and spoof/deepfake baseline detectors. With the exception of protocols and tools for the generation of spoofed/deepfake speech, the resources described in this paper, already used by participants of the ASVspoof 5 challenge in 2024, are now all freely available to the community.
Speech synthesis has significantly advanced from statistical methods to deep neural network architectures, leading to various text-to-speech (TTS) models that closely mimic human speech patterns. However, capturing nuances such as emotion and style in speech synthesis is challenging. To address this challenge, we introduce an approach centered on prompt-based emotion control. The proposed architecture incorporates emotion and intensity control across multi-speakers. Furthermore, we leverage large language models (LLMs) to manipulate speech prosody while preserving linguistic content. Using embedding emotional cues, regulating intensity levels, and guiding prosodic variations with prompts, our approach infuses synthesized speech with human-like expressiveness and variability. Lastly, we demonstrate the effectiveness of our approach through a systematic exploration of the control mechanisms mentioned above.
In recent years, several text-to-speech systems have been proposed to synthesize natural speech in zero-shot, few-shot, and low-resource scenarios. However, these methods typically require training with data from many different speakers. The speech quality across the speaker set typically is diverse and imposes an upper limit on the quality achievable for the low-resource speaker. In the current work, we achieve high-quality speech synthesis using as little as five minutes of speech from the desired speaker by augmenting the low-resource speaker data with noise and employing multiple sampling techniques during training. Our method requires only four high-quality, high-resource speakers, which are easy to obtain and use in practice. Our low-complexity method achieves improved speaker similarity compared to the state-of-the-art zero-shot method HierSpeech++ and the recent low-resource method AdapterMix while maintaining comparable naturalness. Our proposed approach can also reduce the data requirements for speech synthesis for new speakers and languages.
This study explores speaker-specific features encoded in speaker embeddings and intermediate layers of speech self-supervised learning (SSL) models. By utilising a probing method, we analyse features such as pitch, tempo, and energy across prominent speaker embedding models and speech SSL models, including HuBERT, WavLM, and Wav2vec 2.0. The results reveal that speaker embeddings like CAM++ excel in energy classification, while speech SSL models demonstrate superior performance across multiple features due to their hierarchical feature encoding. Intermediate layers effectively capture a mix of acoustic and para-linguistic information, with deeper layers refining these representations. This investigation provides insights into model design and highlights the potential of these representations for downstream applications, such as speaker verification and text-to-speech synthesis, while laying the groundwork for exploring additional features and advanced probing methods.