UniTalker: Conversational Speech-Visual Synthesis

Conversational Speech Synthesis (CSS) is a key task in the user-agent interaction area, aiming to generate more expressive and empathetic speech for users. However, it is well-known that "listening" and "eye contact" play crucial roles in conveying emotions during real-world interpersonal communication. Existing CSS research is limited to perceiving only text and speech within the dialogue context, which restricts its effectiveness. Moreover, speech-only responses further constrain the interactive experience. To address these limitations, we introduce a Conversational Speech-Visual Synthesis (CSVS) task as an extension of traditional CSS. By leveraging multimodal dialogue context, it provides users with coherent audiovisual responses. To this end, we develop a CSVS system named UniTalker, which is a unified model that seamlessly integrates multimodal perception and multimodal rendering capabilities. Specifically, it leverages a large-scale language model to comprehensively understand multimodal cues in the dialogue context, including speaker, text, speech, and the talking-face animations. After that, it employs multi-task sequence prediction to first infer the target utterance's emotion and then generate empathetic speech and natural talking-face animations. To ensure that the generated speech-visual content remains consistent in terms of emotion, content, and duration, we introduce three key optimizations: 1) Designing a specialized neural landmark codec to tokenize and reconstruct facial expression sequences. 2) Proposing a bimodal speech-visual hard alignment decoding strategy. 3) Applying emotion-guided rendering during the generation stage. Comprehensive objective and subjective experiments demonstrate that our model synthesizes more empathetic speech and provides users with more natural and emotionally consistent talking-face animations.

Accent Normalization Using Self-Supervised Discrete Tokens with Non-Parallel Data

Accent normalization converts foreign-accented speech into native-like speech while preserving speaker identity. We propose a novel pipeline using self-supervised discrete tokens and non-parallel training data. The system extracts tokens from source speech, converts them through a dedicated model, and synthesizes the output using flow matching. Our method demonstrates superior performance over a frame-to-frame baseline in naturalness, accentedness reduction, and timbre preservation across multiple English accents. Through token-level phonetic analysis, we validate the effectiveness of our token-based approach. We also develop two duration preservation methods, suitable for applications such as dubbing.

IML-Spikeformer: Input-aware Multi-Level Spiking Transformer for Speech Processing

Spiking Neural Networks (SNNs), inspired by biological neural mechanisms, represent a promising neuromorphic computing paradigm that offers energy-efficient alternatives to traditional Artificial Neural Networks (ANNs). Despite proven effectiveness, SNN architectures have struggled to achieve competitive performance on large-scale speech processing task. Two key challenges hinder progress: (1) the high computational overhead during training caused by multi-timestep spike firing, and (2) the absence of large-scale SNN architectures tailored to speech processing tasks. To overcome the issues, we introduce Input-aware Multi-Level Spikeformer, i.e. IML-Spikeformer, a spiking Transformer architecture specifically designed for large-scale speech processing. Central to our design is the Input-aware Multi-Level Spike (IMLS) mechanism, which simulate multi-timestep spike firing within a single timestep using an adaptive, input-aware thresholding scheme. IML-Spikeformer further integrates a Reparameterized Spiking Self-Attention (RepSSA) module with a Hierarchical Decay Mask (HDM), forming the HD-RepSSA module. This module enhances the precision of attention maps and enables modeling of multi-scale temporal dependencies in speech signals. Experiments demonstrate that IML-Spikeformer achieves word error rates of 6.0\% on AiShell-1 and 3.4\% on Librispeech-960, comparable to conventional ANN transformers while reducing theoretical inference energy consumption by 4.64$\times$ and 4.32$\times$ respectively. IML-Spikeformer marks an advance of scalable SNN architectures for large-scale speech processing in both task performance and energy efficiency.

VP-SelDoA: Visual-prompted Selective DoA Estimation of Target Sound via Semantic-Spatial Matching

Audio-visual sound source localization (AV-SSL) identifies the position of a sound source by exploiting the complementary strengths of auditory and visual signals. However, existing AV-SSL methods encounter three major challenges: 1) inability to selectively isolate the target sound source in multi-source scenarios, 2) misalignment between semantic visual features and spatial acoustic features, and 3) overreliance on paired audio-visual data. To overcome these limitations, we introduce Cross-Instance Audio-Visual Localization (CI-AVL), a novel task that leverages images from different instances of the same sound event category to localize target sound sources, thereby reducing dependence on paired data while enhancing generalization capabilities. Our proposed VP-SelDoA tackles this challenging task through a semantic-level modality fusion and employs a Frequency-Temporal ConMamba architecture to generate target-selective masks for sound isolation. We further develop a Semantic-Spatial Matching mechanism that aligns the heterogeneous semantic and spatial features via integrated cross- and self-attention mechanisms. To facilitate the CI-AVL research, we construct a large-scale dataset named VGG-SSL, comprising 13,981 spatial audio clips across 296 sound event categories. Extensive experiments show that our proposed method outperforms state-of-the-art audio-visual localization methods, achieving a mean absolute error (MAE) of 12.04 and an accuracy (ACC) of 78.23%.

SpeechRefiner: Towards Perceptual Quality Refinement for Front-End Algorithms

Speech pre-processing techniques such as denoising, de-reverberation, and separation, are commonly employed as front-ends for various downstream speech processing tasks. However, these methods can sometimes be inadequate, resulting in residual noise or the introduction of new artifacts. Such deficiencies are typically not captured by metrics like SI-SNR but are noticeable to human listeners. To address this, we introduce SpeechRefiner, a post-processing tool that utilizes Conditional Flow Matching (CFM) to improve the perceptual quality of speech. In this study, we benchmark SpeechRefiner against recent task-specific refinement methods and evaluate its performance within our internal processing pipeline, which integrates multiple front-end algorithms. Experiments show that SpeechRefiner exhibits strong generalization across diverse impairment sources, significantly enhancing speech perceptual quality. Audio demos can be found at https://speechrefiner.github.io/SpeechRefiner/.

Incorporating Linguistic Constraints from External Knowledge Source for Audio-Visual Target Speech Extraction

Audio-visual target speaker extraction (AV-TSE) models primarily rely on target visual cues to isolate the target speaker's voice from others. We know that humans leverage linguistic knowledge, such as syntax and semantics, to support speech perception. Inspired by this, we explore the potential of pre-trained speech-language models (PSLMs) and pre-trained language models (PLMs) as auxiliary knowledge sources for AV-TSE. In this study, we propose incorporating the linguistic constraints from PSLMs or PLMs for the AV-TSE model as additional supervision signals. Without introducing any extra computational cost during inference, the proposed approach consistently improves speech quality and intelligibility. Furthermore, we evaluate our method in multi-language settings and visual cue-impaired scenarios and show robust performance gains.

SongBloom: Coherent Song Generation via Interleaved Autoregressive Sketching and Diffusion Refinement

Generating music with coherent structure, harmonious instrumental and vocal elements remains a significant challenge in song generation. Existing language models and diffusion-based methods often struggle to balance global coherence with local fidelity, resulting in outputs that lack musicality or suffer from incoherent progression and mismatched lyrics. This paper introduces $\textbf{SongBloom}$, a novel framework for full-length song generation that leverages an interleaved paradigm of autoregressive sketching and diffusion-based refinement. SongBloom employs an autoregressive diffusion model that combines the high fidelity of diffusion models with the scalability of language models. Specifically, it gradually extends a musical sketch from short to long and refines the details from coarse to fine-grained. The interleaved generation paradigm effectively integrates prior semantic and acoustic context to guide the generation process. Experimental results demonstrate that SongBloom outperforms existing methods across both subjective and objective metrics and achieves performance comparable to the state-of-the-art commercial music generation platforms. Audio samples are available on our demo page: https://cypress-yang.github.io/SongBloom\_demo.

Exploring Length Generalization For Transformer-based Speech Enhancement

Transformer network architecture has proven effective in speech enhancement. However, as its core module, self-attention suffers from quadratic complexity, making it infeasible for training on long speech utterances. In practical scenarios, speech enhancement models are often required to perform on noisy speech at run-time that is substantially longer than the training utterances. It remains a challenge how a Transformer-based speech enhancement model can generalize to long speech utterances. In this paper, extensive empirical studies are conducted to explore the model's length generalization ability. In particular, we conduct speech enhancement experiments on four training objectives and evaluate with five metrics. Our studies establish that positional encoding is an effective instrument to dampen the effect of utterance length on speech enhancement. We first explore several existing positional encoding methods, and the results show that relative positional encoding methods exhibit a better length generalization property than absolute positional encoding methods. Additionally, we also explore a simpler and more effective positional encoding scheme, i.e. LearnLin, that uses only one trainable parameter for each attention head to scale the real relative position between time frames, which learns the different preferences on short- or long-term dependencies of these heads. The results demonstrate that our proposal exhibits excellent length generalization ability with comparable or superior performance than other state-of-the-art positional encoding strategies.

From Word to World: Evaluate and Mitigate Culture Bias via Word Association Test

The human-centered word association test (WAT) serves as a cognitive proxy, revealing sociocultural variations through lexical-semantic patterns. We extend this test into an LLM-adaptive, free-relation task to assess the alignment of large language models (LLMs) with cross-cultural cognition. To mitigate the culture preference, we propose CultureSteer, an innovative approach that integrates a culture-aware steering mechanism to guide semantic representations toward culturally specific spaces. Experiments show that current LLMs exhibit significant bias toward Western cultural (notably in American) schemas at the word association level. In contrast, our model substantially improves cross-cultural alignment, surpassing prompt-based methods in capturing diverse semantic associations. Further validation on culture-sensitive downstream tasks confirms its efficacy in fostering cognitive alignment across cultures. This work contributes a novel methodological paradigm for enhancing cultural awareness in LLMs, advancing the development of more inclusive language technologies.

Towards Emotionally Consistent Text-Based Speech Editing: Introducing EmoCorrector and The ECD-TSE Dataset

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.