Xiaohai Tian

Siyin Wang, Wenyi Yu, Yudong Yang et al.

Speech quality assessment typically requires evaluating audio from multiple aspects, such as mean opinion score (MOS) and speaker similarity (SIM) \etc., which can be challenging to cover using one small model designed for a single task. In this paper, we propose leveraging recently introduced auditory large language models (LLMs) for automatic speech quality assessment. By employing task-specific prompts, auditory LLMs are finetuned to predict MOS, SIM and A/B testing results, which are commonly used for evaluating text-to-speech systems. Additionally, the finetuned auditory LLM is able to generate natural language descriptions assessing aspects like noisiness, distortion, discontinuity, and overall quality, providing more interpretable outputs. Extensive experiments have been performed on the NISQA, BVCC, SOMOS and VoxSim speech quality datasets, using open-source auditory LLMs such as SALMONN, Qwen-Audio, and Qwen2-Audio. For the natural language descriptions task, a commercial model Google Gemini 1.5 Pro is also evaluated. The results demonstrate that auditory LLMs achieve competitive performance compared to state-of-the-art task-specific small models in predicting MOS and SIM, while also delivering promising results in A/B testing and natural language descriptions. Our data processing scripts and finetuned model checkpoints can be found at https://github.com/bytedance/SALMONN.

Ming-Hao Hsu, Xiaohai Tian, Jun Zhang et al.

Speech Large Language Models (SLLMs) underperform their text counterparts on complex reasoning. We reveal that this modality gap is not a uniform cognitive deficit. Evaluating three diverse SLLMs, we show speech-to-text (S2T) matches or exceeds text-to-text (T2T) on spatial, syntactic, and factual tasks. However, on logical tasks requiring entity tracking, S2T accuracy collapses to chance. We diagnose this localized degradation as an entity binding failure: continuous speech features cause models to lose precise entity-property associations during implicit reasoning. To resolve this, we propose Entity-Aware Chain-of-Thought (EA-CoT), forcing SLLMs to explicitly enumerate entities and bind them to claims before reasoning. Strikingly, EA-CoT bridges the gap, even when spoken names are misrecognized, yielding up to a 24.4% absolute accuracy improvement. Ablations confirm these gains stem entirely from explicit semantic binding, reframing the gap as a resolvable bottleneck.

Kaiqi Fu, Shaojun Gao, Kai Wang et al.

Deep learning-based pronunciation scoring models highly rely on the availability of the annotated non-native data, which is costly and has scalability issues. To deal with the data scarcity problem, data augmentation is commonly used for model pretraining. In this paper, we propose a phone-level mixup, a simple yet effective data augmentation method, to improve the performance of word-level pronunciation scoring. Specifically, given a phoneme sequence from lexicon, the artificial augmented word sample can be generated by randomly sampling from the corresponding phone-level features in training data, while the word score is the average of their GOP scores. Benefit from the arbitrary phone-level combination, the mixup is able to generate any word with various pronunciation scores. Moreover, we utilize multi-source information (e.g., MFCC and deep features) to further improve the scoring system performance. The experiments conducted on the Speechocean762 show that the proposed system outperforms the baseline by adding the mixup data for pretraining, with Pearson correlation coefficients (PCC) increasing from 0.567 to 0.61. The results also indicate that proposed method achieves similar performance by using 1/10 unlabeled data of baseline. In addition, the experimental results also demonstrate the efficiency of our proposed multi-source approach.

Zhixian Zhao, Wenjie Tian, Xiaohai Tian et al.

Multimodal emotion analysis is shifting from static classification to generative reasoning. Beyond simple label prediction, robust affective reasoning must synthesize fine-grained signals such as facial micro-expressions and prosodic which shifts to decode the latent causality within complex social contexts. However, current Multimodal Large Language Models (MLLMs) face significant limitations in fine-grained perception, primarily due to data scarcity and insufficient cross-modal fusion. As a result, these models often exhibit unimodal dominance which leads to hallucinations in complex multimodal interactions, particularly when visual and acoustic cues are subtle, ambiguous, or even contradictory (e.g., in sarcastic scenery). To address this, we introduce SABER-LLM, a framework designed for robust multimodal reasoning. First, we construct SABER, a large-scale emotion reasoning dataset comprising 600K video clips, annotated with a novel six-dimensional schema that jointly captures audiovisual cues and causal logic. Second, we propose the structured evidence decomposition paradigm, which enforces a "perceive-then-reason" separation between evidence extraction and reasoning to alleviate unimodal dominance. The ability to perceive complex scenes is further reinforced by consistency-aware direct preference optimization, which explicitly encourages alignment among modalities under ambiguous or conflicting perceptual conditions. Experiments on EMER, EmoBench-M, and SABER-Test demonstrate that SABER-LLM significantly outperforms open-source baselines and achieves robustness competitive with closed-source models in decoding complex emotional dynamics. The dataset and model are available at https://github.com/zxzhao0/SABER-LLM.

Ming-Hao Hsu, Xueyao Zhang, Xiaohai Tian et al.

Recent advancements in Large Speech-Language Models have significantly bridged the gap between acoustic signals and linguistic understanding. However, a persistent performance disparity remains in speech-based input tasks compared to direct text inference. In this paper, we investigate the dynamic roots of this modality gap beyond static geometric alignment, analyzing how speech and text representations evolve layer-by-layer. We evaluate four open-weight end-to-end models on SpeechMMLU and VoiceBench BBH. Using cross-layer CKA analysis with speech-text token alignment, we find that speech representations exhibit a broad cross-layer alignment band, attributable to the redundant nature of speech where semantic content spans multiple frames. We show that these alignment patterns are structurally stable across different analysis configurations. Crucially, simple statistical calibration is insufficient and can be detrimental when applied at the input layer, indicating that the modality gap is not a mere distribution shift. Overall, our results suggest that the bottleneck lies in condensing redundant speech into stable late-layer decisions, motivating future solutions that operate at the token or temporal granularity instead of feature-level matching.

Siyin Wang, Wenyi Yu, Xianzhao Chen et al.

This paper explores a novel perspective to speech quality assessment by leveraging natural language descriptions, offering richer, more nuanced insights than traditional numerical scoring methods. Natural language feedback provides instructive recommendations and detailed evaluations, yet existing datasets lack the comprehensive annotations needed for this approach. To bridge this gap, we introduce QualiSpeech, a comprehensive low-level speech quality assessment dataset encompassing 11 key aspects and detailed natural language comments that include reasoning and contextual insights. Additionally, we propose the QualiSpeech Benchmark to evaluate the low-level speech understanding capabilities of auditory large language models (LLMs). Experimental results demonstrate that finetuned auditory LLMs can reliably generate detailed descriptions of noise and distortion, effectively identifying their types and temporal characteristics. The results further highlight the potential for incorporating reasoning to enhance the accuracy and reliability of quality assessments. The dataset will be released at https://huggingface.co/datasets/tsinghua-ee/QualiSpeech.

Wenyi Yu, Siyin Wang, Xiaoyu Yang et al.

In order to enable fluid and natural human-machine speech interaction, existing full-duplex conversational systems often adopt modular architectures with auxiliary components such as voice activity detectors, interrupters, conversation state predictors, or multiple LLMs. These systems, however, suffer from error accumulation across modules and struggle with key challenges such as context-dependent barge-in and echo cancellation. Recent approaches, most notably Moshi, simplify the pipeline by injecting audio codecs into the token space of a single LLM. However, such methods still incur significant performance degradation when operating on the speech rather than text modality. In this paper, we introduce SALMONN-omni, the first single, standalone full-duplex speech LLM that operates without audio codecs in its token space. It features a novel dynamic thinking mechanism within the LLM backbone, enabling the model to learn when to transition between speaking and listening states. Experiments on widely used benchmarks for spoken question answering and open-domain dialogue show that SALMONN-omni achieves at least 30\% relative performance improvement over existing open-source full-duplex models and performs highly competitively to half-duplex and turn-based systems, despite using substantially less training data. Moreover, SALMONN-omni demonstrates strong performance in complex conversational scenarios, including turn-taking, backchanneling, echo cancellation and context-dependent barge-in, with further improvements achieved through reinforcement learning. Some demo conversations between user and SALMONN-omni are provided in the following repository https://github.com/bytedance/SALMONN.

Junyi Ao, Yuancheng Wang, Xiaohai Tian et al.

Speech encompasses a wealth of information, including but not limited to content, paralinguistic, and environmental information. This comprehensive nature of speech significantly impacts communication and is crucial for human-computer interaction. Chat-Oriented Large Language Models (LLMs), known for their general-purpose assistance capabilities, have evolved to handle multi-modal inputs, including speech. Although these models can be adept at recognizing and analyzing speech, they often fall short of generating appropriate responses. We argue that this is due to the lack of principles on task definition and model development, which requires open-source datasets and metrics suitable for model evaluation. To bridge the gap, we present SD-Eval, a benchmark dataset aimed at multidimensional evaluation of spoken dialogue understanding and generation. SD-Eval focuses on paralinguistic and environmental information and includes 7,303 utterances, amounting to 8.76 hours of speech data. The data is aggregated from eight public datasets, representing four perspectives: emotion, accent, age, and background sound. To assess the SD-Eval benchmark dataset, we implement three different models and construct a training set following a process similar to that of SD-Eval. The training set contains 1,052.72 hours of speech data and 724.4k utterances. We also conduct a comprehensive evaluation using objective evaluation methods (e.g. BLEU and ROUGE), subjective evaluations and LLM-based metrics for the generated responses. Models conditioned with paralinguistic and environmental information outperform their counterparts in both objective and subjective measures. Moreover, experiments demonstrate that LLM-based metrics show a higher correlation with human evaluation compared to traditional metrics. We open-source SD-Eval at https://github.com/amphionspace/SD-Eval.

Wenyi Yu, Siyin Wang, Xiaoyu Yang et al.

Full-duplex multimodal large language models (LLMs) provide a unified framework for addressing diverse speech understanding and generation tasks, enabling more natural and seamless human-machine conversations. Unlike traditional modularised conversational AI systems, which separate speech recognition, understanding, and text-to-speech generation into distinct components, multimodal LLMs operate as single end-to-end models. This streamlined design eliminates error propagation across components and fully leverages the rich non-verbal information embedded in input speech signals. We introduce SALMONN-omni, a codec-free, full-duplex speech understanding and generation model capable of simultaneously listening to its own generated speech and background sounds while speaking. To support this capability, we propose a novel duplex spoken dialogue framework incorporating a ``thinking'' mechanism that facilitates asynchronous text and speech generation relying on embeddings instead of codecs (quantized speech and audio tokens). Experimental results demonstrate SALMONN-omni's versatility across a broad range of streaming speech tasks, including speech recognition, speech enhancement, and spoken question answering. Additionally, SALMONN-omni excels at managing turn-taking, barge-in, and echo cancellation scenarios, establishing its potential as a robust prototype for full-duplex conversational AI systems. To the best of our knowledge, SALMONN-omni is the first codec-free model of its kind. A full technical report along with model checkpoints will be released soon.

Junyi Ao, Dekun Chen, Xiaohai Tian et al.

Large Language Models (LLMs) have recently shown remarkable ability to process not only text but also multimodal inputs such as speech and audio. However, most existing models primarily focus on analyzing input signals using text instructions, overlooking scenarios in which speech instructions and audio are mixed and serve as inputs to the model. To address these challenges, we introduce Solla, a novel framework designed to understand speech-based questions and hear the acoustic context concurrently. Solla incorporates an audio tagging module to effectively identify and represent audio events, as well as an ASR-assisted prediction method to improve comprehension of spoken content. To rigorously evaluate Solla and other publicly available models, we propose a new benchmark dataset called SA-Eval, which includes three tasks: audio event classification, audio captioning, and audio question answering. SA-Eval has diverse speech instruction with various speaking styles, encompassing two difficulty levels, easy and hard, to capture the range of real-world acoustic conditions. Experimental results show that Solla performs on par with or outperforms baseline models on both the easy and hard test sets, underscoring its effectiveness in jointly understanding speech and audio.

Siyin Wang, Wenyi Yu, Xianzhao Chen et al.

Human interaction is inherently multimodal and full-duplex: we listen while watching, speak while acting, and fluidly adapt to turn-taking and interruptions. Realizing these capabilities is essential for building models simulating humans. We present ELLSA (End-to-end Listen, Look, Speak and Act), which, to our knowledge, is the first full-duplex, end-to-end model that simultaneously perceives and generates across vision, text, speech, and action within a single architecture, enabling interaction patterns previously out of reach, yielding more natural, human-like behaviors. At its core is a novel SA-MoE architecture (Self-Attention Mixture-of-Experts) that routes each modality to specialized experts and fuses them through a unified attention backbone. This provides a generalizable solution for joint multimodal perception and concurrent generation, leveraging strong pre-trained components while enabling efficient modality integration and mitigating modality interference. On speech-interaction and robot-manipulation benchmarks, ELLSA matches modality-specific baselines, while uniquely supporting advanced multimodal and full-duplex behaviors such as dialogue and action turn-taking, defective instruction rejection, speaking-while-acting, context-grounded visual question answering, and action barge-ins. We contend that ELLSA represents a step toward more natural and general interactive intelligence, contributing to the broader pursuit of artificial general intelligence. All data, code and model checkpoints will be released upon acceptance.

Kaiqi Fu, Shaojun Gao, Shuju Shi et al.

Speech fluency/disfluency can be evaluated by analyzing a range of phonetic and prosodic features. Deep neural networks are commonly trained to map fluency-related features into the human scores. However, the effectiveness of deep learning-based models is constrained by the limited amount of labeled training samples. To address this, we introduce a self-supervised learning (SSL) approach that takes into account phonetic and prosody awareness for fluency scoring. Specifically, we first pre-train the model using a reconstruction loss function, by masking phones and their durations jointly on a large amount of unlabeled speech and text prompts. We then fine-tune the pre-trained model using human-annotated scoring data. Our experimental results, conducted on datasets such as Speechocean762 and our non-native datasets, show that our proposed method outperforms the baseline systems in terms of Pearson correlation coefficients (PCC). Moreover, we also conduct an ablation study to better understand the contribution of phonetic and prosody factors during the pre-training stage.

Bidisha Sharma, Xiaoxue Gao, Karthika Vijayan et al.

We present a database of parallel recordings of speech and singing, collected and released by the Human Language Technology (HLT) laboratory at the National University of Singapore (NUS), that is called NUS-HLT Speak-Sing (NHSS) database. We release this database to the public to support research activities, that include, but not limited to comparative studies of acoustic attributes of speech and singing signals, cooperative synthesis of speech and singing voices, and speech-to-singing conversion. This database consists of recordings of sung vocals of English pop songs, the spoken counterpart of lyrics of the songs read by the singers in their natural reading manner, and manually prepared utterance-level and word-level annotations. The audio recordings in the NHSS database correspond to 100 songs sung and spoken by 10 singers, resulting in a total of 7 hours of audio data. There are 5 male and 5 female singers, singing and reading the lyrics of 10 songs each. In this paper, we discuss the design methodology of the database, analyse the similarities and dissimilarities in characteristics of speech and singing voices, and provide some strategies to address relationships between these characteristics for converting one to another. We develop benchmark systems, which can be used as reference for speech-to-singing alignment, spectral mapping, and conversion using the NHSS database.

Hongqiang Du, Xiaohai Tian, Lei Xie et al.

We propose a novel training scheme to optimize voice conversion network with a speaker identity loss function. The training scheme not only minimizes frame-level spectral loss, but also speaker identity loss. We introduce a cycle consistency loss that constrains the converted speech to maintain the same speaker identity as reference speech at utterance level. While the proposed training scheme is applicable to any voice conversion networks, we formulate the study under the average model voice conversion framework in this paper. Experiments conducted on CMU-ARCTIC and CSTR-VCTK corpus confirm that the proposed method outperforms baseline methods in terms of speaker similarity.

Rohan Kumar Das, Tomi Kinnunen, Wen-Chin Huang et al.

The Voice Conversion Challenge 2020 is the third edition under its flagship that promotes intra-lingual semiparallel and cross-lingual voice conversion (VC). While the primary evaluation of the challenge submissions was done through crowd-sourced listening tests, we also performed an objective assessment of the submitted systems. The aim of the objective assessment is to provide complementary performance analysis that may be more beneficial than the time-consuming listening tests. In this study, we examined five types of objective assessments using automatic speaker verification (ASV), neural speaker embeddings, spoofing countermeasures, predicted mean opinion scores (MOS), and automatic speech recognition (ASR). Each of these objective measures assesses the VC output along different aspects. We observed that the correlations of these objective assessments with the subjective results were high for ASV, neural speaker embedding, and ASR, which makes them more influential for predicting subjective test results. In addition, we performed spoofing assessments on the submitted systems and identified some of the VC methods showing a potentially high security risk.

Yi Zhao, Wen-Chin Huang, Xiaohai Tian et al.

The voice conversion challenge is a bi-annual scientific event held to compare and understand different voice conversion (VC) systems built on a common dataset. In 2020, we organized the third edition of the challenge and constructed and distributed a new database for two tasks, intra-lingual semi-parallel and cross-lingual VC. After a two-month challenge period, we received 33 submissions, including 3 baselines built on the database. From the results of crowd-sourced listening tests, we observed that VC methods have progressed rapidly thanks to advanced deep learning methods. In particular, speaker similarity scores of several systems turned out to be as high as target speakers in the intra-lingual semi-parallel VC task. However, we confirmed that none of them have achieved human-level naturalness yet for the same task. The cross-lingual conversion task is, as expected, a more difficult task, and the overall naturalness and similarity scores were lower than those for the intra-lingual conversion task. However, we observed encouraging results, and the MOS scores of the best systems were higher than 4.0. We also show a few additional analysis results to aid in understanding cross-lingual VC better.

Rohan Kumar Das, Xiaohai Tian, Tomi Kinnunen et al.

Security of automatic speaker verification (ASV) systems is compromised by various spoofing attacks. While many types of non-proactive attacks (and their defenses) have been studied in the past, attacker's perspective on ASV, represents a far less explored direction. It can potentially help to identify the weakest parts of ASV systems and be used to develop attacker-aware systems. We present an overview on this emerging research area by focusing on potential threats of adversarial attacks on ASV, spoofing countermeasures, or both. We conclude the study with discussion on selected attacks and leveraging from such knowledge to improve defense mechanisms against adversarial attacks.

Xiaohai Tian, Eng Siong Chng, Haizhou Li

In a typical voice conversion system, vocoder is commonly used for speech-to-features analysis and features-to-speech synthesis. However, vocoder can be a source of speech quality degradation. This paper presents a vocoder-free voice conversion approach using WaveNet for non-parallel training data. Instead of dealing with the intermediate features, the proposed approach utilizes the WaveNet to map the Phonetic PosteriorGrams (PPGs) to the waveform samples directly. In this way, we avoid the estimation errors caused by vocoder and feature conversion. Additionally, as PPG is assumed to be speaker independent, the proposed method also reduces the feature mismatch problem in WaveNet vocoder based approaches. Experimental results conducted on the CMU-ARCTIC database show that the proposed approach significantly outperforms the baseline approaches in terms of speech quality.

Xiaohai Tian, Zhizheng Wu, Xiong Xiao et al.

Spoofing detection for automatic speaker verification (ASV), which is to discriminate between live speech and attacks, has received increasing attentions recently. However, all the previous studies have been done on the clean data without significant additive noise. To simulate the real-life scenarios, we perform a preliminary investigation of spoofing detection under additive noisy conditions, and also describe an initial database for this task. The noisy database is based on the ASVspoof challenge 2015 database and generated by artificially adding background noises at different signal-to-noise ratios (SNRs). Five different additive noises are included. Our preliminary results show that using the model trained from clean data, the system performance degrades significantly in noisy conditions. Phase-based feature is more noise robust than magnitude-based features. And the systems perform significantly differ under different noise scenarios.

Hy Quy Nguyen, Siu Wa Lee, Xiaohai Tian et al.

Voice conversion methods have advanced rapidly over the last decade. Studies have shown that speaker characteristics are captured by spectral feature as well as various prosodic features. Most existing conversion methods focus on the spectral feature as it directly represents the timbre characteristics, while some conversion methods have focused only on the prosodic feature represented by the fundamental frequency. In this paper, a comprehensive framework using deep neural networks to convert both timbre and prosodic features is proposed. The timbre feature is represented by a high-resolution spectral feature. The prosodic features include F0, intensity and duration. It is well known that DNN is useful as a tool to model high-dimensional features. In this work, we show that DNN initialized by our proposed autoencoder pretraining yields good quality DNN conversion models. This pretraining is tailor-made for voice conversion and leverages on autoencoder to capture the generic spectral shape of source speech. Additionally, our framework uses segmental DNN models to capture the evolution of the prosodic features over time. To reconstruct the converted speech, the spectral feature produced by the DNN model is combined with the three prosodic features produced by the DNN segmental models. Our experimental results show that the application of both prosodic and high-resolution spectral features leads to quality converted speech as measured by objective evaluation and subjective listening tests.

Bo Fan, Siu Wa Lee, Xiaohai Tian et al.

State-of-the-art statistical parametric speech synthesis (SPSS) generally uses a vocoder to represent speech signals and parameterize them into features for subsequent modeling. Magnitude spectrum has been a dominant feature over the years. Although perceptual studies have shown that phase spectrum is essential to the quality of synthesized speech, it is often ignored by using a minimum phase filter during synthesis and the speech quality suffers. To bypass this bottleneck in vocoded speech, this paper proposes a phase-embedded waveform representation framework and establishes a magnitude-phase joint modeling platform for high-quality SPSS. Our experiments on waveform reconstruction show that the performance is better than that of the widely-used STRAIGHT. Furthermore, the proposed modeling and synthesis platform outperforms a leading-edge, vocoded, deep bidirectional long short-term memory recurrent neural network (DBLSTM-RNN)-based baseline system in various objective evaluation metrics conducted.