Dialogue serves as the most natural manner of human-computer interaction (HCI). Recent advancements in speech language models (SLM), have significantly enhanced speech-based conversational AI. However, these models are limited to turn-based conversation, lacking the ability to interact with humans in real-time spoken scenarios, for example, being interrupted when the generated content is not satisfactory. To address these limitations, we explore full duplex modeling (FDM) in interactive speech language models (iSLM), focusing on enhancing real-time interaction and, more explicitly, exploring the quintessential ability of interruption. We introduce a novel model design, namely listening-while-speaking language model (LSLM), an end-to-end system equipped with both listening and speaking channels. Our LSLM employs a token-based decoder-only TTS for speech generation and a streaming self-supervised learning (SSL) encoder for real-time audio input. LSLM fuses both channels for autoregressive generation and detects turn-taking in real time. Three fusion strategies—early fusion, middle fusion, and late fusion—are explored, with middle fusion achieving an optimal balance between speech generation and real-time interaction. Two experimental settings, command-based FDM and voice-based FDM, demonstrate LSLM’s robustness to noise and sensitivity to diverse instructions. Our results highlight LSLM’s capability to achieve duplex communication with minimal impact on existing systems. This study aims to advance the development of interactive speech dialogue systems, enhancing their applicability in real-world contexts.

Large Audio-Language Models (LALMs) have demonstrated remarkable performance in tasks involving audio perception and understanding, such as speech recognition and audio captioning. However, their reasoning capabilities-critical for solving complex real-world problems-remain underexplored. In this work, we conduct the first exploration into integrating Chain-of-Thought (CoT) reasoning into LALMs to enhance their reasoning ability across auditory modalities. We evaluate representative CoT methods, analyzing their performance in both information extraction and reasoning tasks across sound, music, and speech domains. Our findings reveal that CoT methods significantly improve performance on easy and medium tasks but encounter challenges with hard tasks, where reasoning chains can confuse the model rather than improve accuracy. Additionally, we identify a positive correlation between reasoning path length and accuracy, demonstrating the potential of scaling inference for advanced instruction-following and reasoning. This study not only highlights the promise of CoT in enhancing LALM reasoning capabilities but also identifies key limitations and provides actionable directions for future research. Index Terms-Chain-of-Thought (CoT), Reasoning, Large Audio Language Model (LALM)

Generative models have shown significant achievements in audio generation tasks. However, existing models struggle with complex and detailed prompts, leading to potential performance degradation. We hypothesize that this problem stems from the simplicity and scarcity of the training data. This work aims to create a large-scale audio dataset with rich captions for improving audio generation models. We first develop an automated pipeline to generate detailed captions by transforming predicted visual captions, audio captions, and tagging labels into comprehensive descriptions using a Large Language Model (LLM). The resulting dataset, Sound-VECaps, comprises 1.66M high-quality audio-caption pairs with enriched details including audio event orders, occurred places and environment information. We then demonstrate that training the text-to-audio generation models with Sound-VECaps significantly improves the performance on complex prompts. Furthermore, we conduct ablation studies of the models on several downstream audio-language tasks, showing the potential of Sound-VECaps in advancing audio-text representation learning.Dataset and demos are available at https://yyua8222.github.io/Sound-VECaps-demo/.

Several recent studies have attempted to autoregressively generate continuous speech representations without discrete speech tokens by combining diffusion and autoregressive models, yet they often face challenges with excessive computational loads or suboptimal outcomes. In this work, we propose Diffusion Transformer Autoregressive Modeling (DiTAR), a patch-based autoregressive framework combining a language model with a diffusion transformer. This approach significantly enhances the efficacy of autoregressive models for continuous tokens and reduces computational demands. DiTAR utilizes a divide-and-conquer strategy for patch generation, where the language model processes aggregated patch embeddings, and the diffusion transformer subsequently generates the next patch based on the output of the language model. For inference, we propose defining temperature as the time point of introducing noise during the reverse diffusion ODE to balance diversity and determinism. We also show in the extensive scaling analysis that DiTAR has superb scalability. In zero-shot speech generation, DiTAR achieves state-of-the-art performance in robustness, speaker similarity, and naturalness.

Generating accurate sounds for complex audio-visual scenes is challenging, especially in the presence of multiple objects and sound sources. In this paper, we propose an interactive object-aware audio generation model that grounds sound generation in user-selected visual objects within images. Our method integrates object-centric learning into a conditional latent diffusion model, which learns to associate image regions with their corresponding sounds through multi-modal attention. At test time, our model employs image segmentation to allow users to interactively generate sounds at the object level. We theoretically validate that our attention mechanism functionally approximates test-time segmentation masks, ensuring the generated audio aligns with selected objects. Quantitative and qualitative evaluations show that our model outperforms baselines, achieving better alignment between objects and their associated sounds.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. These findings underscore the urgent need for greater research attention in audio-language reasoning, including both data and algorithm innovation. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.

Recent language model (LM) advancements have showcased impressive zero-shot voice conversion (VC) performance.However, existing LM-based VC models usually apply offline conversion from source semantics to acoustic features, demanding the complete source speech and limiting their deployment to realtime applications.In this paper, we introduce StreamVoice, a novel streaming LM-based model for zero-shot VC, facilitating real-time conversion given arbitrary speaker prompts and source speech.Specifically, to enable streaming capability, StreamVoice employs a fully causal context-aware LM with a temporalindependent acoustic predictor, while alternately processing semantic and acoustic features at each time step of autoregression which eliminates the dependence on complete source speech.To address the potential performance degradation from the incomplete context in streaming processing, we enhance the contextawareness of the LM through two strategies: 1) teacher-guided context foresight, using a teacher model to summarize the present and future semantic context during training to guide the model's forecasting for missing context; 2) semantic masking strategy, promoting acoustic prediction from preceding corrupted semantic and acoustic input, enhancing context-learning ability.Notably, StreamVoice is the first LMbased streaming zero-shot VC model without any future look-ahead.Experiments demonstrate StreamVoice's streaming conversion capability while achieving zero-shot performance comparable to non-streaming VC systems.

With the huge success of GPT models in natural language processing, there is a growing interest in applying language modeling approaches to speech tasks. Currently, the dominant architecture in speech-to-speech translation (S2ST) remains the encoder-decoder paradigm, creating a need to investigate the impact of language modeling approaches in this area. In this study, we introduce PolyVoice, a language model-based framework designed for S2ST systems. Our framework comprises three decoder-only language models: a translation language model, a duration language model, and a speech synthesis language model. These language models employ different types of prompts to extract learned information effectively. By utilizing unsupervised semantic units, our framework can transfer semantic information across these models, making it applicable even to unwritten languages. We evaluate our system on Chinese $\rightarrow$ English and English $\rightarrow$ Spanish language pairs. Experimental results demonstrate that \method outperforms the state-of-the-art encoder-decoder model, producing voice-cloned speech with high translation and audio quality. Speech samples are available at https://polyvoice.github.io.

StreamVoice has recently pushed the boundaries of zero-shot voice conversion (VC) in the streaming domain. It uses a streamable language model (LM) with a context-aware approach to convert semantic features from automatic speech recognition (ASR) into acoustic features with the desired speaker timbre. Despite its innovations, StreamVoice faces challenges due to its dependency on a streaming ASR within a cascaded framework, which complicates system deployment and optimization, affects VC system's design and performance based on the choice of ASR, and struggles with conversion stability when faced with low-quality semantic inputs. To overcome these limitations, we introduce StreamVoice+, an enhanced LM-based end-to-end streaming framework that operates independently of streaming ASR. StreamVoice+ integrates a semantic encoder and a connector with the original StreamVoice framework, now trained using a non-streaming ASR. This model undergoes a two-stage training process: initially, the StreamVoice backbone is pre-trained for voice conversion and the semantic encoder for robust semantic extraction. Subsequently, the system is fine-tuned end-to-end, incorporating a LoRA matrix to activate comprehensive streaming functionality. Furthermore, StreamVoice+ mainly introduces two strategic enhancements to boost conversion quality: a residual compensation mechanism in the connector to ensure effective semantic transmission and a self-refinement strategy that leverages pseudo-parallel speech pairs generated by the conversion backbone to improve speech decoupling. Experiments demonstrate that StreamVoice+ not only achieves higher naturalness and speaker similarity in voice conversion than its predecessor but also provides versatile support for both streaming and non-streaming conversion scenarios.

Automatic dubbing, which generates a corresponding version of the input speech in another language, can be widely utilized in many real-world scenarios, such as video and game localization. In addition to synthesizing the translated scripts, automatic dubbing further transfers the speaking style in the original language to the dubbed speeches to give audiences the impression that the characters are speaking in their native tongue. However, state-of-the-art automatic dubbing systems only model the transfer on the duration and speaking rate, disregarding the other aspects of speaking style, such as emotion, intonation and emphasis, which are also crucial to fully understand the characters and speech. In this paper, we propose a joint multiscale cross-lingual speaking style transfer framework to simultaneously model the bidirectional speaking style transfer between two languages at both the global scale (i.e., utterance level) and local scale (i.e., word level). The global and local speaking styles in each language are extracted and utilized to predict the global and local speaking styles in the other language with an encoder-decoder framework for each direction and a shared bidirectional attention mechanism for both directions. A multiscale speaking style-enhanced FastSpeech 2 is then utilized to synthesize the desired speech with the predicted global and local speaking styles for each language. The experimental results demonstrate the effectiveness of our proposed framework, which outperforms a baseline with only duration transfer in objective and subjective evaluations.

Zero-shot speaker cloning aims to synthesize speech for any target speaker unseen during TTS system building, given only a single speech reference of the speaker at hand. Although more practical in real applications, the current zero-shot methods still produce speech with undesirable naturalness and speaker similarity. Moreover, endowing the target speaker with arbitrary speaking styles in the zero-shot setup has not been considered. This is because the unique challenge ofzero-shot speaker and style cloningis to learn the disentangled speaker and style representations from only short references representing an arbitrary speaker and an arbitrary style. To address this challenge, we proposeU-Style, which employs Grad-TTS as the backbone, particularly cascading aspeaker-specific encoderand astyle-specific encoderbetween the text encoder and the diffusion decoder. Thus, leveraging signal perturbation, U-Style is explicitly decomposed into speaker- and style-specific modeling parts, achieving better speaker and style disentanglement. To improve unseen speaker and style modeling ability, these two encoders conduct multi-level speaker and style modeling by skip-connected U-nets, incorporating the representation extraction and information reconstruction process. Besides, to improve the naturalness of synthetic speech, we adopt mean-based instance normalization and style adaptive layer normalization in these encoders to perform representation extraction and condition adaptation, respectively. Experiments show that U-Style significantly surpasses the state-of-the-art methods in unseen speaker cloning regarding naturalness and speaker similarity. Notably, U-Style can transfer the style from an unseen source speaker to another unseen target speaker, achieving flexible combinations of desired speaker timbre and style in zero-shot voice cloning.

Although audio generation shares commonalities across different types of audio, such as speech, music, and sound effects, designing models for each type requires careful consideration of specific objectives and biases that can significantly differ from those of other types. To bring us closer to a unified perspective of audio generation, this paper proposes a holistic framework that utilizes the same learning method for speech, music, and sound effect generation. Our framework utilizes a general representation of audio, called “language of audio” (LOA). Any audio can be translated into LOA based on AudioMAE, a self-supervised pre-trained representation learning model. In the generation process, we translate other modalities into LOA by using a GPT-2 model, and we perform self-supervised audio generation learning with a latent diffusion model conditioned on the LOA of audio in our training set. The proposed framework naturally brings advantages such as reusable self-supervised pretrained latent diffusion models. Experiments on the major benchmarks of text-to-audio, text-to-music, and text-to-speech with three AudioLDM 2 variants demonstrate competitive performance of the AudioLDM 2 variants framework against previous approaches. Our code, pretrained model, and demo are available athttps://audioldm.github.io/audioldm2.

Zero-shot voice conversion (VC) converts source speech into the voice of any desired speaker using only one utterance of the speaker without requiring additional model updates. Typical methods use a speaker representation from a pre-trained speaker verification (SV) model or learn speaker representation during VC training to achieve zero-shot VC. However, existing speaker modeling methods overlook the variation of speaker information richness in temporal and frequency channel dimensions of speech. This insufficient speaker modeling hampers the ability of the VC model to accurately represent unseen speakers who are not in the training dataset. In this study, we present a robust zero-shot VC model withmulti-leveltemporal-channelretrieval, referred to as MTCR-VC. Specifically, to flexibly adapt to the dynamic-variant speaker characteristic in the temporal and channel axis of the speech, we propose a novel fine-grained speaker modeling method, calledtemporal-channelretrieval (TCR), to find outwhenandwherespeaker information appears in speech. It retrieves variable-length speaker representation from both temporal and channel dimensions under the guidance of a pre-trained SV model. Besides, inspired by the hierarchical process of human speech production, the MTCR speaker module stacks several TCR blocks to extract speaker representations from multi-granularity levels. Furthermore, we introduce a cycle-based training strategy to simulate zero-shot inference recurrently to achieve better speech disentanglement and reconstruction. To drive this process, we adopt perceptual constraints on three aspects: content, style, and speaker. Experiments demonstrate that MTCR-VC is superior to the previous zero-shot VC methods in modeling speaker timbre while maintaining good speech naturalness.

Streaming voice conversion (VC) is the task of converting the voice of one person to another in real-time. Previous streaming VC methods use phonetic posteriorgrams (PPGs) extracted from automatic speech recognition (ASR) systems to represent speaker-independent information. However, PPGs lack the prosody and phonation information of the source speaker, and streaming PPGs contain undesired leaked timbre of the source speaker. In this paper, we propose to use intermediate bottleneck features (IBFs) to replace PPGs. VC systems trained with IBFs retain more prosody and phonation information of the source speaker. Furthermore, we propose a non-streaming teacher guidance (TG) framework that addresses the timbre leakage problem. Experiments show that our proposed IBFs and the TG framework achieve a state-of-the-art streaming VC naturalness of 3.85, a content consistency of 3.77, and a timbre similarity of 3.77 under a future receptive field of 160 ms which significantly outperform previous streaming VC systems.

Conveying the linguistic content and maintaining the source speech’s speaking style, such as intonation and emotion, is essential in voice conversion (VC). However, in a low-resource situation, where only limited utterances from the target speaker are accessible, existing VC methods are hard to meet this requirement and capture the target speaker’s timber. In this work, a novel VC model, referred to as MFC-StyleVC, is proposed for the low-resource VC task. Specifically, speaker timbre constraint generated by clustering method is newly proposed to guide target speaker timbre learning in different stages. Meanwhile, to prevent over-fitting to the target speaker’s limited data, perceptual regularization constraints explicitly maintain model performance on specific aspects, including speaking style, linguistic content, and speech quality. Besides, a simulation mode is introduced to simulate the inference process to alleviate the mis-match between training and inference. Extensive experiments performed on highly expressive speech demonstrate the superiority of the proposed method in low-resource VC.

Recent progress in music generation has been remarkably advanced by the state-of-the-art MusicLM, which comprises a hierarchy of three LMs, respectively, for semantic, coarse acoustic, and fine acoustic modelings. Yet, sampling with the MusicLM requires processing through these LMs one by one to obtain the fine-grained acoustic tokens, making it computationally expensive and prohibitive for a real-time generation. Efficient music generation with a quality on par with MusicLM remains a significant challenge. In this paper, we present **M**e**L**o**D**y (**M** for music; **L** for LM; **D** for diffusion), an LM-guided diffusion model that generates music audios of state-of-the-art quality meanwhile reducing 95.7\% to 99.6\% forward passes in MusicLM, respectively, for sampling 10s to 30s music. MeLoDy inherits the highest-level LM from MusicLM for semantic modeling, and applies a novel dual-path diffusion (DPD) model and an audio VAE-GAN to efficiently decode the conditioning semantic tokens into waveform. DPD is proposed to simultaneously model the coarse and fine acoustics by incorporating the semantic information into segments of latents effectively via cross-attention at each denoising step. Our experimental results suggest the superiority of MeLoDy, not only in its practical advantages on sampling speed and infinitely continuable generation, but also in its state-of-the-art musicality, audio quality, and text correlation. Our samples are available at https://Efficient-MeLoDy.github.io/.

Language model (LM) based audio generation frameworks, e.g., AudioLM, have recently achieved new state-of-the-art performance in zero-shot audio generation. In this paper, we explore the feasibility of LMs forzero-shot voice conversion. An intuitive approach is to follow AudioLM – Tokenizing speech into semantic and acoustic tokens respectively by HuBERT and SoundStream, and converting source semantic tokens to target acoustic tokens conditioned on acoustic tokens of the target speaker. However, such an approach encounters several issues: 1) the linguistic content contained in semantic tokens may get dispersed during multi-layer modeling while the lengthy speech input in the voice conversion task makes contextual learning even harder; 2) the semantic tokens still contain speaker-related information, which may be leaked to the target speech, lowering the target speaker similarity; 3) the generation diversity in the sampling of the LM can lead to unexpected outcomes during inference, leading to unnatural pronunciation and speech quality degradation. To mitigate these problems, we proposeLM-VC, a two-stage language modeling approach that generates coarse acoustic tokens for recovering the source linguistic content and target speaker's timbre, and then reconstructs the fine for acoustic details as converted speech. Specifically, to enhance content preservation and facilitates better disentanglement, a masked prefix LM with a mask prediction strategy is used for coarse acoustic modeling. This model is encouraged to recover the masked content from the surrounding context and generate target speech based on the target speaker's utterance and corrupted semantic tokens. Besides, to further alleviate the sampling error in the generation, an external LM, which employs window attention to capture the local acoustic relations, is introduced to participate in the coarse acoustic modeling through shallow fusion. Finally, a prefix LM reconstructs fine acoustic tokens from the coarse and results in the converted speech. Experiments demonstrate that LM-VC outperforms competitive systems in speech naturalness and speaker similarity.

While the performance of cross-lingual TTS based on monolingual corpora has been significantly improved recently, generating cross-lingual speech still suffers from the foreign accent problem, leading to limited naturalness. Besides, current cross-lingual methods ignore modeling emotion, which is indispensable paralinguistic information in speech delivery. In this paper, we propose DiCLET-TTS, a Diffusion model based Cross-Lingual Emotion Transfer method that can transfer emotion from a source speaker to the intra- and cross-lingual target speakers. Specifically, to relieve the foreign accent problem while improving the emotion expressiveness, the terminal distribution of the forward diffusion process is parameterized into a speaker-irrelevant but emotion-related linguistic prior by a prior text encoder with the emotion embedding as a condition. To address the weaker emotional expressiveness problem caused by speaker disentanglement in emotion embedding, a novel orthogonal projection based emotion disentangling module (OP-EDM) is proposed to learn the speaker-irrelevant but emotion-discriminative embedding. Moreover, a condition-enhanced DPM decoder is introduced to strengthen the modeling ability of the speaker and the emotion in the reverse diffusion process to further improve emotion expressiveness in speech delivery. Cross-lingual emotion transfer experiments show the superiority of DiCLET-TTS over various competitive models and the good design of OP-EDM in learning speaker-irrelevant but emotion-discriminative embedding.

In addition to conveying the linguistic content from source speech to converted speech, maintaining the speaking style of source speech also plays an important role in the voice conversion (VC) task, which is essential in many scenarios with highly expressive source speech, such as dubbing and data augmentation. Previous work generally took explicit prosodic features or fixed-length style embedding extracted from source speech to model the speaking style of source speech, which is insufficient to achieve comprehensive style modeling and target speaker timbre preservation. Inspired by the style's multi-scale nature of human speech, a multi-scale style modeling method for the VC task, referred to as MSM-VC, is proposed in this article. MSM-VC models the speaking style of source speech from different levels, i.e., global, local, and frame levels. To effectively convey the speaking style and meanwhile prevent timbre leakage from source speech to converted speech, each level's style is modeled by specific representation. Specifically, prosodic features, pre-trained ASR model's bottleneck features, and features extracted by a model trained with a self-supervised strategy are adopted to model the frame, local, and global-level styles, respectively. Besides, to balance the performance of source style modeling and target speaker timbre preservation, an explicit constraint module consisting of a pre-trained speech emotion recognition model and a speaker classifier is introduced to MSM-VC. This explicit constraint module also makes it possible to simulate the style transfer inference process during the training to improve the disentanglement ability and alleviate the mismatch between training and inference. Experiments performed on the highly expressive speech corpus demonstrate that MSM-VC is superior to the state-of-the-art VC methods for modeling source speech style while maintaining good speech quality and speaker similarity. Furthermore, ablation analysis indicates the indispensable of every style level's modeling and the effectiveness of each module.

With the increasing popularity of speech synthesis products, the industry has put forward more requirements for personalized speech synthesis: (1) How to use low-resource, easily accessible data to clone a person’s voice. (2) How to clone a person’s voice while controlling the style and prosody. To solve the above two problems, we proposed the Hieratron model framework in which the prosody and timbre are modeled separately using two modules, therefore, the independent control of timbre and the other characteristics of audio can be achieved while generating speech. The practice shows that, for very limited target speaker data in the wild, Hieratron has obvious advantages over the traditional method, in addition to controlling the style and language of the generated speech, the mean opinion score on speech quality of the generated speech has also been improved by more than 0.2 points.

Although deep learning and end-to-end models have been widely used and shown their superiority in automatic speech recognition (ASR) and text-to-speech (TTS) synthesis, state-of-the-art forced alignment (FA) models are still based on hidden Markov model (HMM). HMM has limited view of contextual information and is developed with long pipelines, leading to error accumulation and unsatisfactory performance. Inspired by the capability of attention mechanism in capturing long term contextual information and learning alignments in ASR and TTS, we propose a neural network based end-to-end forced aligner called NeuFA, in which a novel bidirectional attention mechanism plays an essential role. NeuFA integrates the alignment learning of both ASR and TTS tasks in a unified framework by learning bidirectional alignment information from a shared attention matrix in the proposed bidirectional attention mechanism. Alignments are extracted from the learnt attention weights and optimized by the ASR, TTS and FA tasks in a multi-task learning manner. Experimental results demonstrate the effectiveness of our proposed model, with mean absolute error (MAE) on test set drops from 25.8 ms to 23.7 ms at word level, and from 18.0 ms to 15.7 ms at phoneme level compared with state-of-the-art HMM based model.

To support applications of speech-driven interactive systems in various conversational scenarios, text-to-speech (TTS) synthesis needs to understand the conversational context and determine appropriate speaking styles in its synthesized speeches. These speaking styles are influenced by the dependencies between the multi-modal information in the context at both global scale (i.e. utterance level) and local scale (i.e. word level). However, the dependency modeling and speaking style inference at the local scale are largely missing in state-of-the-art TTS systems, resulting in the synthesis of incorrect or improper speaking styles. In this paper, to learn the dependencies in conversations at both global and local scales and to improve the synthesis of speaking styles, we propose a context modeling method which models the dependencies among the multi-modal information in context with multi-scale relational graph convolutional network (MSRGCN). The learnt multi-modal context information at multiple scales is then utilized to infer the global and local speaking styles of the current utterance for speech synthesis. Experiments demonstrate the effectiveness of the proposed approach, and ablation studies reflect the contributions from modeling multi-modal information and multi-scale dependencies.

Dubbing is a post-production process of re-recording actors’ dialogues, which is extensively used in filmmaking and video production. It is usually performed manually by professional voice actors who read lines with proper prosody, and in synchronization with the pre-recorded videos. In this work, we propose Neural Dubber, the first neural network model to solve a novel automatic video dubbing (AVD) task: synthesizing human speech synchronized with the given video from the text. Neural Dubber is a multi-modal text-to-speech (TTS) model that utilizes the lip movement in the video to control the prosody of the generated speech. Furthermore, an image-based speaker embedding (ISE) module is developed for the multi-speaker setting, which enables Neural Dubber to generate speech with a reasonable timbre according to the speaker’s face. Experiments on the chemistry lecture single-speaker dataset and LRS2 multi-speaker dataset show that Neural Dubber can generate speech audios on par with state-of-the-art TTS models in terms of speech quality. Most importantly, both qualitative and quantitative evaluations show that Neural Dubber can control the prosody of synthesized speech by the video, and generate high-fidelity speech temporally synchronized with the video.