arXiv subjects

Songjun Cao

Publications and source records attributed to Songjun Cao.

At least 19 recordsLinked to original sources

arXiv2026

STEB: A Speech-to-Speech Translation Expressiveness Benchmark for Evaluating Beyond Translation Fidelity

Speech-to-speech translation (S2ST) should preserve not only lexical meaning, but also expressive attributes: emotion, scenario style (e.g., news reporting vs. dramatic dialogue), and nonverbal vocalizations (NVs). Moreover, collecting cross-lingual target speech that is both translation-faithful and expressively aligned with the source is difficult at scale, making reference-based evaluation impractical. We introduce STEB (Speech-to-Speech Translation Expressiveness Benchmark), a 32.6-hour Chinese--English benchmark that evaluates both standard dimensions (translation fidelity, speaker similarity, duration alignment) and expressiveness dimensions (emotion, scenario style, NV preservation). For expressiveness evaluation, STEB uses a caption-then-summarize framework that converts speech into structured expressive attributes and compares source and hypothesis attributes with an LLM judge. Human validation shows statistically significant correlations with listener judgments across all expressive dimensions. We evaluate six S2ST systems covering cascaded systems, end-to-end models, and speech large language models. Many systems, especially cascaded ones, achieve strong translation fidelity, but they still struggle with emotion preservation (best: 3.82/5) and NV preservation (best: 2.31/5). These results reveal a gap between semantic transfer and expressive transfer, identifying expressiveness preservation as an open challenge for S2ST. Audio samples are available at https://cmots.github.io/steb.github.io/.

cs.SD↗

arXiv2026

MathVis-Fine: Aligning Visual Supervision with Necessity via Progressive Dependency-Guided Training for Multimodal Mathematical Reasoning

Chain-of-Thought (CoT) reasoning has extended from purely linguistic domains to multimodal scenarios; however, existing approaches often treat visual inputs as homogeneous or auxiliary signals, failing to capture the intricate and sample-specific dependencies between text and images in mathematical problem-solving. This gives rise to two core issues: first, the supervisory signals for visual content are generalized and coarse-grained, lacking adaptation to the actual necessity of visual information in each sample; second, training feedback becomes inaccurate when visual rewards are uniformly applied without distinguishing the complementary relationships among inputs. These limitations hinder models from achieving precise multimodal reasoning. In this work, we propose a framework for modeling fine-grained visual dependencies in mathematical reasoning. We first construct the MathVis-Fine dataset, augmenting fine-grained visual annotations with visual dependency ratings. Building upon this dataset, we introduce a two-stage progressive visual enhancement training paradigm that balances answer correctness rewards and visual grounding rewards according to the intrinsic visual dependency level of each sample, thereby mitigating reward bias and improving supervision accuracy. Extensive experiments demonstrate that the MathVis-Fine framework effectively enhances visual perception progressively based on visual dependency, offering a more precise training framework for multimodal mathematical reasoning. We will release the dataset upon acceptance.

cs.AI↗

arXiv2026

Diffusion Reconstruction towards Generalizable Audio Deepfake Detection

Achieving robust generalization against unseen attacks remains a challenge in Audio Deepfake Detection (ADD), driven by the rapid evolution of generative models. To address this, we propose a framework centered on hard sample classification. The core idea is that a model capable of distinguishing challenging hard samples is inherently equipped to handle simpler cases effectively. We investigate multiple reconstruction paradigms, identifying the diffusion-based method as optimal for generating hard samples. Furthermore, we leverage multi-layer feature aggregation and introduce a Regularization-Assisted Contrastive Learning (RACL) objective to enhance generalizability. Experiments demonstrate the superior generalization of our approach, with our best model achieving a significant reduction in the average Equal Error Rate (EER) compared to the baseline.

cs.SD↗

arXiv2026

Controllable Spoken Dialogue Generation: An LLM-Driven Grading System for K-12 Non-Native English Learners

Large language models (LLMs) often fail to meet the pedagogical needs of K-12 English learners in non-native contexts due to a proficiency mismatch. To address this widespread challenge, we introduce a proficiency-aligned framework that adapts LLM outputs to learner abilities, using China's national curriculum (CSE) as a representative case. Our framework enables precise control over lexical complexity through a four-tier grading system, supported by a comprehensive suite of new resources: graded vocabulary lists and a multi-turn dialogue corpus. Our core technical contribution is the \textbf{DDPO} algorithm,Diversity Driven Policy Optimization, a multi-turn GRPO-based approach designed to preserve dialogue diversity while holistically optimizing dialogue quality. This method significantly outperforms conventional approaches, achieving low out-of-vocabulary rates and high diversity while enhancing conversational naturalness and pedagogical value. While grounded in the CSE, our framework is designed for flexibility and can be readily adapted to other educational standards. Our models, data, and code will all be open-sourced, providing a scalable platform for personalized English speaking practice that effectively addresses the unique challenges faced by K-12 learners in non-immersive environments.

cs.CL↗

arXiv2026

Thinking with Constructions: A Benchmark and Policy Optimization for Visual-Text Interleaved Geometric Reasoning

Geometric reasoning inherently requires "thinking with constructions" -- the dynamic manipulation of visual aids to bridge the gap between problem conditions and solutions. However, existing Multimodal Large Language Models (MLLMs) are largely confined to passive inference with static diagrams, lacking the strategic knowledge of when and how to construct effective visual aids. To address this, we present a framework for Visual-Text Interleaved Chain-of-Thought. We first introduce GeoAux-Bench, the first benchmark comprising 4,334 geometry problems that aligns textual construction steps with ground-truth visual updates. Our pilot study reveals two critical insights: (1) interleaved visual-textual aids outperform single-modality counterparts, which cannot losslessly capture geometric synergy; and (2) valid constructions act as entropy reducers, strongly correlating with reduced reasoning perplexity. Building on these findings, we propose Action Applicability Policy Optimization (A2PO), a reinforcement learning paradigm for mastering strategic construction. A2PO employs Adaptive Reward Shaping to regulate the timing and quality of visual aids via counterfactual sampling to distinguish necessary from redundant constructions. Experiments demonstrate our approach enables MLLMs to leverage selective auxiliary constructions, yielding a 3.51% gain over strong baselines. Code and data are available on GitHub.

cs.AI↗

arXiv2026

Leveraging large multimodal models for audio-video deepfake detection: a pilot study

Audio-visual deepfake detection (AVD) is increasingly important as modern generators can fabricate convincing speech and video. Most current multimodal detectors are small, task-specific models: they work well on curated tests but scale poorly and generalize weakly across domains. We introduce AV-LMMDetect, a supervised fine-tuned (SFT) large multimodal model that casts AVD as a prompted yes/no classification - "Is this video real or fake?". Built on Qwen 2.5 Omni, it jointly analyzes audio and visual streams for deepfake detection and is trained in two stages: lightweight LoRA alignment followed by audio-visual encoder full fine-tuning. On FakeAVCeleb and Mavos-DD, AV-LMMDetect matches or surpasses prior methods and sets a new state of the art on Mavos-DD datasets.

cs.SD↗

arXiv2025

PseudoVC: Improving One-shot Voice Conversion with Pseudo Paired Data

As parallel training data is scarce for one-shot voice conversion (VC) tasks, waveform reconstruction is typically performed by various VC systems. A typical one-shot VC system comprises a content encoder and a speaker encoder. However, two types of mismatches arise: one for the inputs to the content encoder during training and inference, and another for the inputs to the speaker encoder. To address these mismatches, we propose a novel VC training method called \textit{PseudoVC} in this paper. First, we introduce an innovative information perturbation approach named \textit{Pseudo Conversion} to tackle the first mismatch problem. This approach leverages pretrained VC models to convert the source utterance into a perturbed utterance, which is fed into the content encoder during training. Second, we propose an approach termed \textit{Speaker Sampling} to resolve the second mismatch problem, which will substitute the input to the speaker encoder by another utterance from the same speaker during training. Experimental results demonstrate that our proposed \textit{Pseudo Conversion} outperforms previous information perturbation methods, and the overall \textit{PseudoVC} method surpasses publicly available VC models. Audio examples are available.

eess.AS↗

arXiv2025

MPE-TTS: Customized Emotion Zero-Shot Text-To-Speech Using Multi-Modal Prompt

Most existing Zero-Shot Text-To-Speech(ZS-TTS) systems generate the unseen speech based on single prompt, such as reference speech or text descriptions, which limits their flexibility. We propose a customized emotion ZS-TTS system based on multi-modal prompt. The system disentangles speech into the content, timbre, emotion and prosody, allowing emotion prompts to be provided as text, image or speech. To extract emotion information from different prompts, we propose a multi-modal prompt emotion encoder. Additionally, we introduce an prosody predictor to fit the distribution of prosody and propose an emotion consistency loss to preserve emotion information in the predicted prosody. A diffusion-based acoustic model is employed to generate the target mel-spectrogram. Both objective and subjective experiments demonstrate that our system outperforms existing systems in terms of naturalness and similarity. The samples are available at https://mpetts-demo.github.io/mpetts_demo/.

cs.SD↗

arXiv2025

Detect All-Type Deepfake Audio: Wavelet Prompt Tuning for Enhanced Auditory Perception

The rapid advancement of audio generation technologies has escalated the risks of malicious deepfake audio across speech, sound, singing voice, and music, threatening multimedia security and trust. While existing countermeasures (CMs) perform well in single-type audio deepfake detection (ADD), their performance declines in cross-type scenarios. This paper is dedicated to studying the all-type ADD task. We are the first to comprehensively establish an all-type ADD benchmark to evaluate current CMs, incorporating cross-type deepfake detection across speech, sound, singing voice, and music. Then, we introduce the prompt tuning self-supervised learning (PT-SSL) training paradigm, which optimizes SSL front-end by learning specialized prompt tokens for ADD, requiring 458x fewer trainable parameters than fine-tuning (FT). Considering the auditory perception of different audio types, we propose the wavelet prompt tuning (WPT)-SSL method to capture type-invariant auditory deepfake information from the frequency domain without requiring additional training parameters, thereby enhancing performance over FT in the all-type ADD task. To achieve an universally CM, we utilize all types of deepfake audio for co-training. Experimental results demonstrate that WPT-XLSR-AASIST achieved the best performance, with an average EER of 3.58% across all evaluation sets.

cs.SD↗

arXiv2025

DiffCSS: Diverse and Expressive Conversational Speech Synthesis with Diffusion Models

Conversational speech synthesis (CSS) aims to synthesize both contextually appropriate and expressive speech, and considerable efforts have been made to enhance the understanding of conversational context. However, existing CSS systems are limited to deterministic prediction, overlooking the diversity of potential responses. Moreover, they rarely employ language model (LM)-based TTS backbones, limiting the naturalness and quality of synthesized speech. To address these issues, in this paper, we propose DiffCSS, an innovative CSS framework that leverages diffusion models and an LM-based TTS backbone to generate diverse, expressive, and contextually coherent speech. A diffusion-based context-aware prosody predictor is proposed to sample diverse prosody embeddings conditioned on multimodal conversational context. Then a prosody-controllable LM-based TTS backbone is developed to synthesize high-quality speech with sampled prosody embeddings. Experimental results demonstrate that the synthesized speech from DiffCSS is more diverse, contextually coherent, and expressive than existing CSS systems

cs.SD↗

arXiv2025

Neural Codec Source Tracing: Toward Comprehensive Attribution in Open-Set Condition

Current research in audio deepfake detection is gradually transitioning from binary classification to multi-class tasks, referred as audio deepfake source tracing task. However, existing studies on source tracing consider only closed-set scenarios and have not considered the challenges posed by open-set conditions. In this paper, we define the Neural Codec Source Tracing (NCST) task, which is capable of performing open-set neural codec classification and interpretable ALM detection. Specifically, we constructed the ST-Codecfake dataset for the NCST task, which includes bilingual audio samples generated by 11 state-of-the-art neural codec methods and ALM-based out-ofdistribution (OOD) test samples. Furthermore, we establish a comprehensive source tracing benchmark to assess NCST models in open-set conditions. The experimental results reveal that although the NCST models perform well in in-distribution (ID) classification and OOD detection, they lack robustness in classifying unseen real audio. The ST-codecfake dataset and code are available.

cs.SD↗

arXiv2024

A Transcription Prompt-based Efficient Audio Large Language Model for Robust Speech Recognition

Audio-LLM introduces audio modality into a large language model (LLM) to enable a powerful LLM to recognize, understand, and generate audio. However, during speech recognition in noisy environments, we observed the presence of illusions and repetition issues in audio-LLM, leading to substitution and insertion errors. This paper proposes a transcription prompt-based audio-LLM by introducing an ASR expert as a transcription tokenizer and a hybrid Autoregressive (AR) Non-autoregressive (NAR) decoding approach to solve the above problems. Experiments on 10k-hour WenetSpeech Mandarin corpus show that our approach decreases 12.2% and 9.6% CER relatively on Test_Net and Test_Meeting evaluation sets compared with baseline. Notably, we reduce the decoding repetition rate on the evaluation set to zero, showing that the decoding repetition problem has been solved fundamentally.

cs.SD↗

arXiv2023

DistillW2V2: A Small and Streaming Wav2vec 2.0 Based ASR Model

Wav2vec 2.0 (W2V2) has shown impressive performance in automatic speech recognition (ASR). However, the large model size and the non-streaming architecture make it hard to be used under low-resource or streaming scenarios. In this work, we propose a two-stage knowledge distillation method to solve these two problems: the first step is to make the big and non-streaming teacher model smaller, and the second step is to make it streaming. Specially, we adopt the MSE loss for the distillation of hidden layers and the modified LF-MMI loss for the distillation of the prediction layer. Experiments are conducted on Gigaspeech, Librispeech, and an in-house dataset. The results show that the distilled student model (DistillW2V2) we finally get is 8x faster and 12x smaller than the original teacher model. For the 480ms latency setup, the DistillW2V2's relative word error rate (WER) degradation varies from 9% to 23.4% on test sets, which reveals a promising way to extend the W2V2's application scope.

eess.AS↗

arXiv2022

Censer: Curriculum Semi-supervised Learning for Speech Recognition Based on Self-supervised Pre-training

Recent studies have shown that the benefits provided by self-supervised pre-training and self-training (pseudo-labeling) are complementary. Semi-supervised fine-tuning strategies under the pre-training framework, however, remain insufficiently studied. Besides, modern semi-supervised speech recognition algorithms either treat unlabeled data indiscriminately or filter out noisy samples with a confidence threshold. The dissimilarities among different unlabeled data are often ignored. In this paper, we propose Censer, a semi-supervised speech recognition algorithm based on self-supervised pre-training to maximize the utilization of unlabeled data. The pre-training stage of Censer adopts wav2vec2.0 and the fine-tuning stage employs an improved semi-supervised learning algorithm from slimIPL, which leverages unlabeled data progressively according to their pseudo labels' qualities. We also incorporate a temporal pseudo label pool and an exponential moving average to control the pseudo labels' update frequency and to avoid model divergence. Experimental results on Libri-Light and LibriSpeech datasets manifest our proposed method achieves better performance compared to existing approaches while being more unified.

cs.SD↗

arXiv2022

A practical framework for multi-domain speech recognition and an instance sampling method to neural language modeling

Automatic speech recognition (ASR) systems used on smart phones or vehicles are usually required to process speech queries from very different domains. In such situations, a vanilla ASR system usually fails to perform well on every domain. This paper proposes a multi-domain ASR framework for Tencent Map, a navigation app used on smart phones and in-vehicle infotainment systems. The proposed framework consists of three core parts: a basic ASR module to generate n-best lists of a speech query, a text classification module to determine which domain the speech query belongs to, and a reranking module to rescore n-best lists using domain-specific language models. In addition, an instance sampling based method to training neural network language models (NNLMs) is proposed to address the data imbalance problem in multi-domain ASR. In experiments, the proposed framework was evaluated on navigation domain and music domain, since navigating and playing music are two main features of Tencent Map. Compared to a general ASR system, the proposed framework achieves a relative 13% $\sim$ 22% character error rate reduction on several test sets collected from Tencent Map and our in-car voice assistant.

eess.AS↗

arXiv2022

Improving CTC-based speech recognition via knowledge transferring from pre-trained language models

Recently, end-to-end automatic speech recognition models based on connectionist temporal classification (CTC) have achieved impressive results, especially when fine-tuned from wav2vec2.0 models. Due to the conditional independence assumption, CTC-based models are always weaker than attention-based encoder-decoder models and require the assistance of external language models (LMs). To solve this issue, we propose two knowledge transferring methods that leverage pre-trained LMs, such as BERT and GPT2, to improve CTC-based models. The first method is based on representation learning, in which the CTC-based models use the representation produced by BERT as an auxiliary learning target. The second method is based on joint classification learning, which combines GPT2 for text modeling with a hybrid CTC/attention architecture. Experiment on AISHELL-1 corpus yields a character error rate (CER) of 4.2% on the test set. When compared to the vanilla CTC-based models fine-tuned from the wav2vec2.0 models, our knowledge transferring method reduces CER by 16.1% relatively without external LMs.

cs.CL↗

arXiv2021

Improving Hybrid CTC/Attention End-to-end Speech Recognition with Pretrained Acoustic and Language Model

Recently, self-supervised pretraining has achieved impressive results in end-to-end (E2E) automatic speech recognition (ASR). However, the dominant sequence-to-sequence (S2S) E2E model is still hard to fully utilize the self-supervised pre-training methods because its decoder is conditioned on acoustic representation thus cannot be pretrained separately. In this paper, we propose a pretrained Transformer (Preformer) S2S ASR architecture based on hybrid CTC/attention E2E models to fully utilize the pretrained acoustic models (AMs) and language models (LMs). In our framework, the encoder is initialized with a pretrained AM (wav2vec2.0). The Preformer leverages CTC as an auxiliary task during training and inference. Furthermore, we design a one-cross decoder (OCD), which relaxes the dependence on acoustic representations so that it can be initialized with pretrained LM (DistilGPT2). Experiments are conducted on the AISHELL-1 corpus and achieve a $4.6\%$ character error rate (CER) on the test set. Compared with our vanilla hybrid CTC/attention Transformer baseline, our proposed CTC/attention-based Preformer yields $27\%$ relative CER reduction. To the best of our knowledge, this is the first work to utilize both pretrained AM and LM in a S2S ASR system.

eess.AS↗

arXiv2021

Improving Streaming Transformer Based ASR Under a Framework of Self-supervised Learning

Recently self-supervised learning has emerged as an effective approach to improve the performance of automatic speech recognition (ASR). Under such a framework, the neural network is usually pre-trained with massive unlabeled data and then fine-tuned with limited labeled data. However, the non-streaming architecture like bidirectional transformer is usually adopted by the neural network to achieve competitive results, which can not be used in streaming scenarios. In this paper, we mainly focus on improving the performance of streaming transformer under the self-supervised learning framework. Specifically, we propose a novel two-stage training method during fine-tuning, which combines knowledge distilling and self-training. The proposed training method achieves 16.3% relative word error rate (WER) reduction on Librispeech noisy test set. Finally, by only using the 100h clean subset of Librispeech as the labeled data and the rest (860h) as the unlabeled data, our streaming transformer based model obtains competitive WERs 3.5/8.7 on Librispeech clean/noisy test sets.

eess.AS↗