Guangzhi Sun

Changli Tang, Wenyi Yu, Guangzhi Sun et al.

Hearing is arguably an essential ability of artificial intelligence (AI) agents in the physical world, which refers to the perception and understanding of general auditory information consisting of at least three types of sounds: speech, audio events, and music. In this paper, we propose SALMONN, a speech audio language music open neural network, built by integrating a pre-trained text-based large language model (LLM) with speech and audio encoders into a single multimodal model. SALMONN enables the LLM to directly process and understand general audio inputs and achieve competitive performances on a number of speech and audio tasks used in training, such as automatic speech recognition and translation, auditory-information-based question answering, emotion recognition, speaker verification, and music and audio captioning etc. SALMONN also has a diverse set of emergent abilities unseen in the training, which includes but is not limited to speech translation to untrained languages, speech-based slot filling, spoken-query-based question answering, audio-based storytelling, and speech audio co-reasoning etc. The presence of cross-modal emergent abilities is studied, and a novel few-shot activation tuning approach is proposed to activate such abilities. To our knowledge, SALMONN is the first model of its type and can be regarded as a step towards AI with generic hearing abilities. The source code, model checkpoints and data are available at https://github.com/bytedance/SALMONN.

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.

Guangzhi Sun, Wenyi Yu, Changli Tang et al.

Audio-visual large language models (LLM) have drawn significant attention, yet the fine-grained combination of both input streams is rather under-explored, which is challenging but necessary for LLMs to understand general video inputs. To this end, a fine-grained audio-visual joint representation (FAVOR) learning framework for multimodal LLMs is proposed in this paper, which extends a text-based LLM to simultaneously perceive speech and audio events in the audio input stream and images or videos in the visual input stream, at the frame level. To fuse the audio and visual feature streams into joint representations and to align the joint space with the LLM input embedding space, we propose a causal Q-Former structure with a causal attention module to enhance the capture of causal relations of the audio-visual frames across time. An audio-visual evaluation benchmark (AVEB) is also proposed which comprises six representative single-modal tasks with five cross-modal tasks reflecting audio-visual co-reasoning abilities. While achieving competitive single-modal performance on audio, speech and image tasks in AVEB, FAVOR achieved over 20% accuracy improvements on the video question-answering task when fine-grained information or temporal causal reasoning is required. FAVOR, in addition, demonstrated remarkable video comprehension and reasoning abilities on tasks that are unprecedented by other multimodal LLMs. An interactive demo of FAVOR is available at https://github.com/BriansIDP/AudioVisualLLM.git, and the training code and model checkpoints will be released soon.

Guangzhi Sun, Xianrui Zheng, Chao Zhang et al.

End-to-end automatic speech recognition (ASR) and large language models, such as Whisper and GPT-2, have recently been scaled to use vast amounts of training data. Despite the large amount of training data, infrequent content words that occur in a particular task may still exhibit poor ASR performance, with contextual biasing a possible remedy. This paper investigates the effectiveness of neural contextual biasing for Whisper combined with GPT-2. Specifically, this paper proposes integrating an adapted tree-constrained pointer generator (TCPGen) component for Whisper and a dedicated training scheme to dynamically adjust the final output without modifying any Whisper model parameters. Experiments across three datasets show a considerable reduction in errors on biasing words with a biasing list of 1000 words. Contextual biasing was more effective when applied to domain-specific data and can boost the performance of Whisper and GPT-2 without losing their generality.

Potsawee Manakul, Guangzhi Sun, Warit Sirichotedumrong et al.

Audio language models process audio inputs using textual prompts for tasks like speech recognition and audio captioning. Although built on multilingual pre-trained components, most are trained primarily on English, limiting their usability for other languages. This paper evaluates audio language models on Thai, a low-resource language, and finds that they lack emergent cross-lingual abilities despite their multilingual foundations. To address this, we explore data mixtures that optimize audio language models for both a target language and English while integrating audio comprehension and speech instruction-following into a unified model. Our experiments provide insights into improving instruction-following in low-resource languages by balancing language-specific and multilingual training data. The proposed model, Typhoon-Audio, significantly outperforms existing open-source models and achieves performance comparable to state-of-the-art Gemini-1.5-Pro in both English and Thai.

Guangzhi Sun, Chao Zhang, Philip C. Woodland

Incorporating biasing words obtained as contextual knowledge is critical for many automatic speech recognition (ASR) applications. This paper proposes the use of graph neural network (GNN) encodings in a tree-constrained pointer generator (TCPGen) component for end-to-end contextual ASR. By encoding the biasing words in the prefix-tree with a tree-based GNN, lookahead for future wordpieces in end-to-end ASR decoding is achieved at each tree node by incorporating information about all wordpieces on the tree branches rooted from it, which allows a more accurate prediction of the generation probability of the biasing words. Systems were evaluated on the Librispeech corpus using simulated biasing tasks, and on the AMI corpus by proposing a novel visual-grounded contextual ASR pipeline that extracts biasing words from slides alongside each meeting. Results showed that TCPGen with GNN encodings achieved about a further 15% relative WER reduction on the biasing words compared to the original TCPGen, with a negligible increase in the computation cost for decoding.

Guangzhi Sun, Chao Zhang, Philip C Woodland

Contextual knowledge is essential for reducing speech recognition errors on high-valued long-tail words. This paper proposes a novel tree-constrained pointer generator (TCPGen) component that enables end-to-end ASR models to bias towards a list of long-tail words obtained using external contextual information. With only a small overhead in memory use and computation cost, TCPGen can structure thousands of biasing words efficiently into a symbolic prefix-tree and creates a neural shortcut between the tree and the final ASR output to facilitate the recognition of the biasing words. To enhance TCPGen, we further propose a novel minimum biasing word error (MBWE) loss that directly optimises biasing word errors during training, along with a biasing-word-driven language model discounting (BLMD) method during the test. All contextual ASR systems were evaluated on the public Librispeech audiobook corpus and the data from the dialogue state tracking challenges (DSTC) with the biasing lists extracted from the dialogue-system ontology. Consistent word error rate (WER) reductions were achieved with TCPGen, which were particularly significant on the biasing words with around 40\% relative reductions in the recognition error rates. MBWE and BLMD further improved the effectiveness of TCPGen and achieved more significant WER reductions on the biasing words. TCPGen also achieved zero-shot learning of words not in the audio training set with large WER reductions on the out-of-vocabulary words in the biasing list.

Yang Li, Cheng Yu, Guangzhi Sun et al.

Modelling prosody variation is critical for synthesizing natural and expressive speech in end-to-end text-to-speech (TTS) systems. In this paper, a cross-utterance conditional VAE (CUC-VAE) is proposed to estimate a posterior probability distribution of the latent prosody features for each phoneme by conditioning on acoustic features, speaker information, and text features obtained from both past and future sentences. At inference time, instead of the standard Gaussian distribution used by VAE, CUC-VAE allows sampling from an utterance-specific prior distribution conditioned on cross-utterance information, which allows the prosody features generated by the TTS system to be related to the context and is more similar to how humans naturally produce prosody. The performance of CUC-VAE is evaluated via a qualitative listening test for naturalness, intelligibility and quantitative measurements, including word error rates and the standard deviation of prosody attributes. Experimental results on LJ-Speech and LibriTTS data show that the proposed CUC-VAE TTS system improves naturalness and prosody diversity with clear margins.

Guangzhi Sun, Chao Zhang, Philip C. Woodland

End-to-end spoken language understanding (SLU) suffers from the long-tail word problem. This paper exploits contextual biasing, a technique to improve the speech recognition of rare words, in end-to-end SLU systems. Specifically, a tree-constrained pointer generator (TCPGen), a powerful and efficient biasing model component, is studied, which leverages a slot shortlist with corresponding entities to extract biasing lists. Meanwhile, to bias the SLU model output slot distribution, a slot probability biasing (SPB) mechanism is proposed to calculate a slot distribution from TCPGen. Experiments on the SLURP dataset showed consistent SLU-F1 improvements using TCPGen and SPB, especially on unseen entities. On a new split by holding out 5 slot types for the test, TCPGen with SPB achieved zero-shot learning with an SLU-F1 score over 50% compared to baselines which can not deal with it. In addition to slot filling, the intent classification accuracy was also improved.

Guangzhi Sun, Chao Zhang, Ivan Vulić et al.

Manually annotating fine-grained slot-value labels for task-oriented dialogue (ToD) systems is an expensive and time-consuming endeavour. This motivates research into slot-filling methods that operate with limited amounts of labelled data. Moreover, the majority of current work on ToD is based solely on text as the input modality, neglecting the additional challenges of imperfect automatic speech recognition (ASR) when working with spoken language. In this work, we propose a Knowledge-Aware Audio-Grounded generative slot-filling framework, termed KA2G, that focuses on few-shot and zero-shot slot filling for ToD with speech input. KA2G achieves robust and data-efficient slot filling for speech-based ToD by 1) framing it as a text generation task, 2) grounding text generation additionally in the audio modality, and 3) conditioning on available external knowledge (e.g. a predefined list of possible slot values). We show that combining both modalities within the KA2G framework improves the robustness against ASR errors. Further, the knowledge-aware slot-value generator in KA2G, implemented via a pointer generator mechanism, particularly benefits few-shot and zero-shot learning. Experiments, conducted on the standard speech-based single-turn SLURP dataset and a multi-turn dataset extracted from a commercial ToD system, display strong and consistent gains over prior work, especially in few-shot and zero-shot setups.

Wenyi Yu, Changli Tang, Guangzhi Sun et al.

The impressive capability and versatility of large language models (LLMs) have aroused increasing attention in automatic speech recognition (ASR), with several pioneering studies attempting to build integrated ASR models by connecting a speech encoder with an LLM. This paper presents a comparative study of three commonly used structures as connectors, including fully connected layers, multi-head cross-attention, and Q-Former. Speech encoders from the Whisper model series as well as LLMs from the Vicuna model series with different model sizes were studied. Experiments were performed on the commonly used LibriSpeech, Common Voice, and GigaSpeech datasets, where the LLMs with Q-Formers demonstrated consistent and considerable word error rate (WER) reductions over LLMs with other connector structures. Q-Former-based LLMs can generalise well to out-of-domain datasets, where 12% relative WER reductions over the Whisper baseline ASR model were achieved on the Eval2000 test set without using any in-domain training data from Switchboard. Moreover, a novel segment-level Q-Former is proposed to enable LLMs to recognise speech segments with a duration exceeding the limitation of the encoders, which results in 17% relative WER reductions over other connector structures on 90-second-long speech data.

Yang Li, Cheng Yu, Guangzhi Sun et al.

Speech synthesis systems powered by neural networks hold promise for multimedia production, but frequently face issues with producing expressive speech and seamless editing. In response, we present the Cross-Utterance Conditioned Variational Autoencoder speech synthesis (CUC-VAE S2) framework to enhance prosody and ensure natural speech generation. This framework leverages the powerful representational capabilities of pre-trained language models and the re-expression abilities of variational autoencoders (VAEs). The core component of the CUC-VAE S2 framework is the cross-utterance CVAE, which extracts acoustic, speaker, and textual features from surrounding sentences to generate context-sensitive prosodic features, more accurately emulating human prosody generation. We further propose two practical algorithms tailored for distinct speech synthesis applications: CUC-VAE TTS for text-to-speech and CUC-VAE SE for speech editing. The CUC-VAE TTS is a direct application of the framework, designed to generate audio with contextual prosody derived from surrounding texts. On the other hand, the CUC-VAE SE algorithm leverages real mel spectrogram sampling conditioned on contextual information, producing audio that closely mirrors real sound and thereby facilitating flexible speech editing based on text such as deletion, insertion, and replacement. Experimental results on the LibriTTS datasets demonstrate that our proposed models significantly enhance speech synthesis and editing, producing more natural and expressive speech.

Yudong Yang, Xuezhen Zhang, Zhifeng Han et al.

Recent progress in large language models (LLMs) has enabled understanding of both speech and non-speech audio, but exposing new safety risks emerging from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech-Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech-audio composition mechanisms. SACRED-Bench adopts three mechanisms: (a) speech overlap and multi-speaker dialogue, which embeds harmful prompts beneath or alongside benign speech; (b) speech-audio mixture, which imply unsafe intent via non-speech audio alongside benign speech or audio; and (c) diverse spoken instruction formats (open-ended QA, yes/no) that evade text-only filters. Experiments show that, even Gemini 2.5 Pro, the state-of-the-art proprietary LLM, still exhibits 66% attack success rate in SACRED-Bench test set, exposing vulnerabilities under cross-modal, speech-audio composition attacks. To bridge this gap, we propose SALMONN-Guard, a safeguard LLM that jointly inspects speech, audio, and text for safety judgments, reducing attack success down to 20%. Our results highlight the need for audio-aware defenses for the safety of multimodal LLMs. The benchmark and SALMONN-Guard checkpoints can be found at https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench. Warning: this paper includes examples that may be offensive or harmful.

Xiao Zhan, Guangzhi Sun, Jose Such et al.

Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model's ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

Guangzhi Sun, Shutong Feng, Dongcheng Jiang et al.

Recently, advancements in large language models (LLMs) have shown an unprecedented ability across various language tasks. This paper investigates the potential application of LLMs to slot filling with noisy ASR transcriptions, via both in-context learning and task-specific fine-tuning. Dedicated prompt designs and fine-tuning approaches are proposed to improve the robustness of LLMs for slot filling with noisy ASR transcriptions. Moreover, a linearised knowledge injection (LKI) scheme is also proposed to integrate dynamic external knowledge into LLMs. Experiments were performed on SLURP to quantify the performance of LLMs, including GPT-3.5-turbo, GPT-4, LLaMA-13B and Vicuna-13B (v1.1 and v1.5) with different ASR error rates. The use of the proposed fine-tuning together with the LKI scheme for LLaMA-13B achieved an 8.3% absolute SLU-F1 improvement compared to the strong Flan-T5-base baseline system on a limited data setup.

Potsawee Manakul, Woody Haosheng Gan, Martijn Bartelds et al. · gatech

Current audio language models are predominantly text-first, either extending pre-trained text LLM backbones or relying on semantic-only audio tokens, limiting general audio modeling. This paper presents a systematic empirical study of native audio foundation models that apply next-token prediction to audio at scale, jointly modeling semantic content, acoustic details, and text to support both general audio generation and cross-modal capabilities. We provide comprehensive empirical insights for building such models: (1) We systematically investigate design choices -- data sources, text mixture ratios, and token composition -- establishing a validated training recipe. (2) We conduct the first scaling law study for discrete audio models via IsoFLOP analysis on 64 models spanning $3{\times}10^{18}$ to $3{\times}10^{20}$ FLOPs, finding that optimal data grows 1.6$\times$ faster than optimal model size. (3) We apply these lessons to train SODA (Scaling Open Discrete Audio), a suite of models from 135M to 4B parameters on 500B tokens, comparing against our scaling predictions and existing models. SODA serves as a flexible backbone for diverse audio/text tasks -- we demonstrate this by fine-tuning for voice-preserving speech-to-speech translation, using the same unified architecture.

Shutong Feng, Guangzhi Sun, Nurul Lubis et al.

Affect recognition, encompassing emotions, moods, and feelings, plays a pivotal role in human communication. In the realm of conversational artificial intelligence, the ability to discern and respond to human affective cues is a critical factor for creating engaging and empathetic interactions. This study investigates the capacity of large language models (LLMs) to recognise human affect in conversations, with a focus on both open-domain chit-chat dialogues and task-oriented dialogues. Leveraging three diverse datasets, namely IEMOCAP (Busso et al., 2008), EmoWOZ (Feng et al., 2022), and DAIC-WOZ (Gratch et al., 2014), covering a spectrum of dialogues from casual conversations to clinical interviews, we evaluate and compare LLMs' performance in affect recognition. Our investigation explores the zero-shot and few-shot capabilities of LLMs through in-context learning as well as their model capacities through task-specific fine-tuning. Additionally, this study takes into account the potential impact of automatic speech recognition errors on LLM predictions. With this work, we aim to shed light on the extent to which LLMs can replicate human-like affect recognition capabilities in conversations.

Adel Moumen, Guangzhi Sun, Philip C. Woodland

Spoken Language Models (SLMs) aim to learn linguistic competence directly from speech using discrete units, widening access to Natural Language Processing (NLP) technologies for languages with limited written resources. However, progress has been largely English-centric due to scarce spoken evaluation benchmarks and training data, making cross-lingual learning difficult. We present a cross-lingual interleaving method that mixes speech tokens across languages without textual supervision. We also release an EN-FR training dataset, TinyStories (~42k hours), together with EN-FR spoken StoryCloze and TopicCloze benchmarks for cross-lingual semantic evaluation, both synthetically generated using GPT-4. On 360M and 1B SLMs under matched training-token budgets, interleaving improves monolingual semantic accuracy, enables robust cross-lingual continuation, and strengthens cross-lingual hidden-state alignment. Taken together, these results indicate that cross-lingual interleaving is a simple, scalable route to building multilingual SLMs that understand and converse across languages. All resources will be made open-source to support reproducibility.

Theodor Nguyen, Guangzhi Sun, Xianrui Zheng et al.

We propose DiffSpEx, a generative target speaker extraction method based on score-based generative modelling through stochastic differential equations. DiffSpEx deploys a continuous-time stochastic diffusion process in the complex short-time Fourier transform domain, starting from the target speaker source and converging to a Gaussian distribution centred on the mixture of sources. For the reverse-time process, a parametrised score function is conditioned on a target speaker embedding to extract the target speaker from the mixture of sources. We utilise ECAPA-TDNN target speaker embeddings and condition the score function alternately on the SDE time embedding and the target speaker embedding. The potential of DiffSpEx is demonstrated with the WSJ0-2mix dataset, achieving an SI-SDR of 12.9 dB and a NISQA score of 3.56. Moreover, we show that fine-tuning a pre-trained DiffSpEx model to a specific speaker further improves performance, enabling personalisation in target speaker extraction.

Qiuming Zhao, Guangzhi Sun, Chao Zhang et al.

Recent end-to-end automatic speech recognition (ASR) models have become increasingly larger, making them particularly challenging to be deployed on resource-constrained devices. Model quantisation is an effective solution that sometimes causes the word error rate (WER) to increase. In this paper, a novel strategy of personalisation for a quantised model (PQM) is proposed, which combines speaker adaptive training (SAT) with model quantisation to improve the performance of heavily compressed models. Specifically, PQM uses a 4-bit NormalFloat Quantisation (NF4) approach for model quantisation and low-rank adaptation (LoRA) for SAT. Experiments have been performed on the LibriSpeech and the TED-LIUM 3 corpora. Remarkably, with a 7x reduction in model size and 1% additional speaker-specific parameters, 15.1% and 23.3% relative WER reductions were achieved on quantised Whisper and Conformer-based attention-based encoder-decoder ASR models respectively, comparing to the original full precision models.

Yudong Yang, Zhan Liu, Wenyi Yu et al.

Diffusion-based generative models have recently achieved remarkable results in speech and vocal enhancement due to their ability to model complex speech data distributions. While these models generalize well to unseen acoustic environments, they may not achieve the same level of fidelity as the discriminative models specifically trained to enhance particular acoustic conditions. In this paper, we propose Ex-Diff, a novel score-based diffusion model that integrates the latent representations produced by a discriminative model to improve speech and vocal enhancement, which combines the strengths of both generative and discriminative models. Experimental results on the widely used MUSDB dataset show relative improvements of 3.7% in SI-SDR and 10.0% in SI-SIR compared to the baseline diffusion model for speech and vocal enhancement tasks, respectively. Additionally, case studies are provided to further illustrate and analyze the complementary nature of generative and discriminative models in this context.

Junli Chen, Changli Tang, Yixuan Li et al.

Visual information, such as subtitles in a movie, often helps automatic speech recognition. In this paper, we propose Donut-Whisper, an audio-visual ASR model with dual encoder to leverage visual information to improve speech recognition performance in both English and Chinese. Donut-Whisper combines the advantage of the linear and the Q-Former-based modality alignment structures via a cross-attention module, generating more powerful audio-visual features. Meanwhile, we propose a lightweight knowledge distillation scheme showcasing the potential of using audio-visual models to teach audio-only models to achieve better performance. Moreover, we propose a new multilingual audio-visual speech recognition dataset based on movie clips containing both Chinese and English partitions. As a result, Donut-Whisper achieved significantly better performance on both English and Chinese partition of the dataset compared to both Donut and Whisper large V3 baselines. In particular, an absolute 5.75% WER reduction and a 16.5% absolute CER reduction were achieved on the English and Chinese sets respectively compared to the Whisper ASR baseline.

Guangzhi Sun, Yudong Yang, Jimin Zhuang et al.

While recent advancements in reasoning optimization have significantly enhanced the capabilities of large language models (LLMs), existing efforts to improve reasoning have been limited to solving mathematical problems and focusing on visual graphical inputs, neglecting broader applications in general video understanding.This paper proposes video-SALMONN-o1, the first open-source reasoning-enhanced audio-visual LLM designed for general video understanding tasks. To enhance its reasoning abilities, we develop a reasoning-intensive dataset featuring challenging audio-visual questions with step-by-step solutions. We also propose process direct preference optimization (pDPO), which leverages contrastive step selection to achieve efficient step-level reward modelling tailored for multimodal inputs. Additionally, we introduce RivaBench, the first reasoning-intensive video understanding benchmark, featuring over 4,000 high-quality, expert-curated question-answer pairs across scenarios such as standup comedy, academic presentations, and synthetic video detection. video-SALMONN-o1 achieves 3-8% accuracy improvements over the LLaVA-OneVision baseline across different video reasoning benchmarks. Besides, pDPO achieves 6-8% improvements compared to the supervised fine-tuning model on RivaBench. Enhanced reasoning enables video-SALMONN-o1 zero-shot synthetic video detection capabilities.

Yixuan Li, Changli Tang, Jimin Zhuang et al.

Human vision is dynamic and continuous. However, in video understanding with multimodal large language models (LLMs), existing methods primarily rely on static features extracted from images sampled at a fixed low frame rate of frame-per-second (FPS) $\leqslant$2, leading to critical visual information loss. In this paper, we introduce F-16, the first multimodal LLM designed for high-frame-rate video understanding. By increasing the frame rate to 16 FPS and compressing visual tokens within each 1-second clip, F-16 efficiently captures dynamic visual features while preserving key semantic information. Experimental results demonstrate that higher frame rates considerably enhance video understanding across multiple benchmarks, providing a new approach to improving video LLMs beyond scaling model size or training data. F-16 achieves state-of-the-art performance among 7-billion-parameter video LLMs on both general and fine-grained video understanding benchmarks, such as Video-MME and TemporalBench. Furthermore, F-16 excels in complex spatiotemporal tasks, including high-speed sports analysis (\textit{e.g.}, basketball, football, gymnastics, and diving), outperforming SOTA proprietary visual models like GPT-4o and Gemini-1.5-pro. Additionally, we introduce a novel decoding method for F-16 that enables highly efficient low-frame-rate inference without requiring model retraining. We will release the source code, model checkpoints, and data at \href{https://github.com/bytedance/F-16}{https://github.com/bytedance/F-16}.

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.

Guangzhi Sun, Xiao Zhan, Shutong Feng et al.

Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware SafEty Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments (p<0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts.

Yudong Yang, Jimin Zhuang, Guangzhi Sun et al.

Audio often serves as an auxiliary modality in video understanding tasks of audio-visual large language models (LLMs), merely assisting in the comprehension of visual information. However, a thorough understanding of videos significantly depends on auditory information, as audio offers critical context, emotional cues, and semantic meaning that visual data alone often lacks. This paper proposes an audio-centric video understanding benchmark (AVUT) to evaluate the video comprehension capabilities of multimodal LLMs with a particular focus on auditory information. AVUT introduces a suite of carefully designed audio-centric tasks, holistically testing the understanding of both audio content and audio-visual interactions in videos. Moreover, this work points out the text shortcut problem that largely exists in other benchmarks where the correct answer can be found from question text alone without needing videos. AVUT addresses this problem by proposing a answer permutation-based filtering mechanism. A thorough evaluation across a diverse range of open-source and proprietary multimodal LLMs is performed, followed by the analyses of deficiencies in audio-visual LLMs. Demos and data are available at https://github.com/lark-png/AVUT.

Zhe Chen, Heyang Liu, Wenyi Yu et al.

Publishing open-source academic video recordings is an emergent and prevalent approach to sharing knowledge online. Such videos carry rich multimodal information including speech, the facial and body movements of the speakers, as well as the texts and pictures in the slides and possibly even the papers. Although multiple academic video datasets have been constructed and released, few of them support both multimodal content recognition and understanding tasks, which is partially due to the lack of high-quality human annotations. In this paper, we propose a novel multimodal, multigenre, and multipurpose audio-visual academic lecture dataset (M$^3$AV), which has almost 367 hours of videos from five sources covering computer science, mathematics, and medical and biology topics. With high-quality human annotations of the slide text and spoken words, in particular high-valued name entities, the dataset can be used for multiple audio-visual recognition and understanding tasks. Evaluations performed on contextual speech recognition, speech synthesis, and slide and script generation tasks demonstrate that the diversity of M$^3$AV makes it a challenging dataset.

Mengjie Qian, Guangzhi Sun, Mark J. F. Gales et al.

Large language models (LLMs) are increasingly applied as automatic evaluators for natural language generation assessment often using pairwise comparative judgements. Existing approaches typically rely on single judges or aggregate multiple judges assuming equal reliability. In practice, LLM judges vary substantially in performance across tasks and aspects, and their judgment probabilities may be biased and inconsistent. Furthermore, human-labelled supervision for judge calibration may be unavailable. We first empirically demonstrate that inconsistencies in LLM comparison probabilities exist and show that it limits the effectiveness of direct probability-based ranking. To address this, we study the LLM-as-a-jury setting and propose BT-sigma, a judge-aware extension of the Bradley-Terry model that introduces a discriminator parameter for each judge to jointly infer item rankings and judge reliability from pairwise comparisons alone. Experiments on benchmark NLG evaluation datasets show that BT-sigma consistently outperforms averaging-based aggregation methods, and that the learned discriminator strongly correlates with independent measures of the cycle consistency of LLM judgments. Further analysis reveals that BT-sigma can be interpreted as an unsupervised calibration mechanism that improves aggregation by modelling judge reliability.

Xiaoliang Luo, Akilles Rechardt, Guangzhi Sun et al.

Scientific discoveries often hinge on synthesizing decades of research, a task that potentially outstrips human information processing capacities. Large language models (LLMs) offer a solution. LLMs trained on the vast scientific literature could potentially integrate noisy yet interrelated findings to forecast novel results better than human experts. To evaluate this possibility, we created BrainBench, a forward-looking benchmark for predicting neuroscience results. We find that LLMs surpass experts in predicting experimental outcomes. BrainGPT, an LLM we tuned on the neuroscience literature, performed better yet. Like human experts, when LLMs were confident in their predictions, they were more likely to be correct, which presages a future where humans and LLMs team together to make discoveries. Our approach is not neuroscience-specific and is transferable to other knowledge-intensive endeavors.

Changli Tang, Yixuan Li, Yudong Yang et al.

We present video-SALMONN 2, a family of audio-visual large language models that set new state-of-the-art (SOTA) results in video description and question answering (QA). Our core contribution is multi-round direct preference optimisation (MrDPO), paired with a caption-quality objective that jointly rewards completeness and factual accuracy. Unlike standard DPO with a fixed reference policy, MrDPO periodically refreshes the reference by bootstrapping from a newly re-initialised lightweight adapter trained on the latest preferences, avoiding reference staleness and enabling continual improvement. This strategy produces captions that are consistently more detailed and accurate than those from proprietary systems such as GPT-4o and Gemini-1.5 Pro. We further distil these gains by using our model to generate a high-quality video-caption corpus for supervised fine-tuning of new models, transferring benefits beyond captioning to strong performance on complex video-QA tasks. Across widely used audio-visual and visual-only understanding benchmarks (including Video-MME, WorldSense, AVUT, Video-Holmes, DailyOmni, MLVU, and LVBench), our 3B and 7B models achieve SOTA results at comparable scales, while the 72B model surpasses all other open-source systems. Our source code, models, and data are released at \href{https://github.com/bytedance/video-SALMONN-2}{https://github.com/bytedance/video-SALMONN-2}.

Guangzhi Sun, Wenyi Yu, Changli Tang et al.

Speech understanding as an element of the more generic video understanding using audio-visual large language models (av-LLMs) is a crucial yet understudied aspect. This paper proposes video-SALMONN, a single end-to-end av-LLM for video processing, which can understand not only visual frame sequences, audio events and music, but speech as well. To obtain fine-grained temporal information required by speech understanding, while keeping efficient for other video elements, this paper proposes a novel multi-resolution causal Q-Former (MRC Q-Former) structure to connect pre-trained audio-visual encoders and the backbone large language model. Moreover, dedicated training approaches including the diversity loss and the unpaired audio-visual mixed training scheme are proposed to avoid frames or modality dominance. On the introduced speech-audio-visual evaluation benchmark, video-SALMONN achieves more than 25\% absolute accuracy improvements on the video-QA task and over 30\% absolute accuracy improvements on audio-visual QA tasks with human speech. In addition, video-SALMONN demonstrates remarkable video comprehension and reasoning abilities on tasks that are unprecedented by other av-LLMs. Our training code and model checkpoints are available at \texttt{\url{https://github.com/bytedance/SALMONN/}}.

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.

Guangzhi Sun, Potsawee Manakul, Adian Liusie et al.

Multimodal foundation models are prone to hallucination, generating outputs that either contradict the input or are not grounded by factual information. Given the diversity in architectures, training data and instruction tuning techniques, there can be large variations in systems' susceptibility to hallucinations. To assess system hallucination robustness, hallucination ranking approaches have been developed for specific tasks such as image captioning, question answering, summarization, or biography generation. However, these approaches typically compare model outputs to gold-standard references or labels, limiting hallucination benchmarking for new domains. This work proposes "CrossCheckGPT", a reference-free universal hallucination ranking for multimodal foundation models. The core idea of CrossCheckGPT is that the same hallucinated content is unlikely to be generated by different independent systems, hence cross-system consistency can provide meaningful and accurate hallucination assessment scores. CrossCheckGPT can be applied to any model or task, provided that the information consistency between outputs can be measured through an appropriate distance metric. Focusing on multimodal large language models that generate text, we explore two information consistency measures: CrossCheck-explicit and CrossCheck-implicit. We showcase the applicability of our method for hallucination ranking across various modalities, namely the text, image, and audio-visual domains. Further, we propose the first audio-visual hallucination benchmark, "AVHalluBench", and illustrate the effectiveness of CrossCheckGPT, achieving correlations of 98% and 89% with human judgements on MHaluBench and AVHalluBench, respectively.

Nineli Lashkarashvili, Wen Wu, Guangzhi Sun et al.

Foundation models have shown superior performance for speech emotion recognition (SER). However, given the limited data in emotion corpora, finetuning all parameters of large pre-trained models for SER can be both resource-intensive and susceptible to overfitting. This paper investigates parameter-efficient finetuning (PEFT) for SER. Various PEFT adaptors are systematically studied for both classification of discrete emotion categories and prediction of dimensional emotional attributes. The results demonstrate that the combination of PEFT methods surpasses full finetuning with a significant reduction in the number of trainable parameters. Furthermore, a two-stage adaptation strategy is proposed to adapt models trained on acted emotion data, which is more readily available, to make the model more adept at capturing natural emotional expressions. Both intra- and cross-corpus experiments validate the efficacy of the proposed approach in enhancing the performance on both the source and target domains.

Ziyun Cui, Ziyang Zhang, Guangzhi Sun et al.

Advances in large language models raise the question of how alignment techniques will adapt as models become increasingly complex and humans will only be able to supervise them weakly. Weak-to-Strong mimics such a scenario where weak model supervision attempts to harness the full capabilities of a much stronger model. This work extends Weak-to-Strong to WeakS-to-Strong by exploring an ensemble of weak models which simulate the variability in human opinions. Confidence scores are estimated using a Bayesian approach to guide the WeakS-to-Strong generalization. Furthermore, we extend the application of WeakS-to-Strong from text classification tasks to text generation tasks where more advanced strategies are investigated for supervision. Moreover, direct preference optimization is applied to advance the student model's preference learning, beyond the basic learning framework of teacher forcing. Results demonstrate the effectiveness of the proposed approach for the reliability of a strong student model, showing potential for superalignment.

Guangzhi Sun, Potsawee Manakul, Xiao Zhan et al.

Unlearning has emerged as a critical capability for large language models (LLMs) to support data privacy, regulatory compliance, and ethical AI deployment. Recent techniques often rely on obfuscation by injecting incorrect or irrelevant information to suppress knowledge. Such methods effectively constitute knowledge addition rather than true removal, often leaving models vulnerable to probing. In this paper, we formally distinguish unlearning from obfuscation and introduce a probing-based evaluation framework to assess whether existing approaches genuinely remove targeted information. Moreover, we propose DF-MCQ, a novel unlearning method that flattens the model predictive distribution over automatically generated multiple-choice questions using KL-divergence, effectively removing knowledge about target individuals and triggering appropriate refusal behaviour. Experimental results demonstrate that DF-MCQ achieves unlearning with over 90% refusal rate and a random choice-level uncertainty that is much higher than obfuscation on probing questions.

Qiuming Zhao, Guangzhi Sun, Chao Zhang

Language diversity presents a significant challenge in speech-to-text (S2T) tasks, such as automatic speech recognition and translation. Traditional multi-lingual multi-task training approaches aim to address this by jointly optimising multiple speech recognition and translation tasks across various languages. While models like Whisper, built on these strategies, demonstrate strong performance, they still face issues of high computational cost, language interference, suboptimal training configurations, and limited extensibility. To overcome these challenges, we introduce LoRS-Merging (low-rank and sparse model merging), a novel technique designed to efficiently integrate models trained on different languages or tasks while preserving performance and reducing computational overhead. LoRS-Merging combines low-rank and sparse pruning to retain essential structures while eliminating redundant parameters, mitigating language interference, and enhancing extensibility. Experimental results across 10 languages demonstrate that LoRS-Merging significantly outperforms multi-lingual multi-task training, sequential training, and other merging methods, achieving over 20% improvement in normalised performance. Our findings suggest that model merging, particularly LoRS-Merging, is a scalable and effective complement to traditional multi-lingual training strategies for S2T applications.

Xiaoliang Luo, Guangzhi Sun, Bradley C. Love

Recently, large language models (LLMs) have outperformed human experts in predicting the results of neuroscience experiments (Luo et al., 2024). What is the basis for this performance? One possibility is that statistical patterns in that specific scientific literature, as opposed to emergent reasoning abilities arising from broader training, underlie LLMs' performance. To evaluate this possibility, we trained (next word prediction) a relatively small 124M-parameter GPT-2 model on 1.3 billion tokens of domain-specific knowledge. Despite being orders of magnitude smaller than larger LLMs trained on trillions of tokens, small models achieved expert-level performance in predicting neuroscience results. Small models trained on the neuroscience literature succeeded when they were trained from scratch using a tokenizer specifically trained on neuroscience text or when the neuroscience literature was used to finetune a pretrained GPT-2. Our results indicate that expert-level performance may be attained by even small LLMs through domain-specific, auto-regressive training approaches.

Guangzhi Sun, Anmol Kagrecha, Potsawee Manakul et al.

Large Language Models (LLMs) are increasingly used to assess NLP tasks due to their ability to generate human-like judgments. Single LLMs were used initially, however, recent work suggests using multiple LLMs as judges yields improved performance. An important step in exploiting multiple judgements is the combination stage, aggregation. Existing methods in NLP either assign equal weight to all LLM judgments or are designed for specific tasks such as hallucination detection. This work focuses on aggregating predictions from multiple systems where no reference labels are available. A new method called SkillAggregation is proposed, which learns to combine estimates from LLM judges without needing additional data or ground truth. It extends the Crowdlayer aggregation method, developed for image classification, to exploit the judge estimates during inference. The approach is compared to a range of standard aggregation methods on HaluEval-Dialogue, TruthfulQA and Chatbot Arena tasks. SkillAggregation outperforms Crowdlayer on all tasks, and yields the best performance over all approaches on the majority of tasks.

Guangzhi Sun, Yixuan Li, Xiaodong Wu et al.

Continuous, high-frame-rate, high-resolution processing of long video streams is critical for future AI agents, yet current video-understanding LLMs struggle to scale. Offline, fixed-frame-number methods require the stream length to adapt frame rates; streaming methods constrain memory by merging or discarding tokens, losing information. We propose video-SALMONN S, a streaming audio-visual LLM that, to our knowledge, is the first to process 3-hour videos at 1 FPS and 360p resolution under a fixed memory budget. Our model introduces (i) a test-time-training (TTT) memory module that continually updates token representations to capture long-range dependencies by replacing token merging, and (ii) a prompt-dependent memory reader that selectively retrieves context-relevant content from fixed-size memory. The TTT module is optimised with a Hessian-free conjugate-gradient procedure (TTT_HF) for efficient adaptation. On long-video benchmarks (Video-MME, LVBench, VideoEvalPro), video-SALMONN S sustains high-quality understanding on multi-hour videos with 10k frames and 1M tokens. Our 8B-parameter model achieves 74.2% overall and 67.8% on the Video-MME long split, outperforming both offline and streaming baselines.

Adel Moumen, Guangzhi Sun, Philip C Woodland

Speech Large Language Models route speech encoder representations into an LLM decoder that typically accounts for over 90% of total parameters. We study how much of this decoder capacity is actually needed for speech tasks. Across two LLM families and three scales (1-8B), we show that decoder redundancy is largely inherited from the pretrained LLM: text and speech inputs yield similar redundant blocks. We then measure excess capacity by pruning decoder layers and analysing post-pruning healing to increase robustness. Our findings show that 7-8B models retain good ASR performance with only 60% of decoder layers, and the same trend extends to smaller scales with reduced pruning tolerance. We then generalise to speech translation, and show that the same blocks of layers are redundant across speech encoders, tasks and languages, indicating that a more global redundancy structure exists, enabling a single pruned and multi-tasks SpeechLLM backbone to be deployed.

Mengqi Wang, Zhan Liu, Zengrui Jin et al.

Diffusion-based large language models (DLLMs) have recently attracted growing interest as an alternative to autoregressive decoders. In this work, we present an empirical study on using the diffusion-based large language model LLaDA for automatic speech recognition (ASR). We first investigate its use as an external deliberation-based processing module for Whisper-LLaMA transcripts. By leveraging the bidirectional attention and denoising capabilities of LLaDA, we explore random masking, low-confidence masking, and semi-autoregressive strategies, showing that Whisper-LLaDA substantially reduces WER compared with the baseline. On LibriSpeech, the best cascade system achieves 2.25%/4.94% WER on test-clean/test-other, representing a 12.3% relative improvement over the Whisper-LLaMA baseline on the test-other split. In contrast, a plain-text LLaDA without acoustic features fails to improve accuracy, highlighting the importance of audio-conditioned embeddings. We further evaluate Whisper-LLaDA as a standalone decoder for ASR with diffusion-based and semi-autoregressive decoding. Most experimental configurations achieve faster inference than the Whisper-LLaMA baseline, although recognition accuracy is slightly lower. These findings offer an empirical view of diffusion-based LLMs for ASR and point to promising directions for improvements.

Guangzhi Sun, Chao Zhang, Phil Woodland

The incorporation of biasing words obtained through contextual knowledge is of paramount importance in automatic speech recognition (ASR) applications. This paper proposes an innovative method for achieving end-to-end contextual ASR using graph neural network (GNN) encodings based on the tree-constrained pointer generator method. GNN node encodings facilitate lookahead for future word pieces in the process of ASR decoding at each tree node by incorporating information about all word pieces on the tree branches rooted from it. This results in a more precise prediction of the generation probability of the biasing words. The study explores three GNN encoding techniques, namely tree recursive neural networks, graph convolutional network (GCN), and GraphSAGE, along with different combinations of the complementary GCN and GraphSAGE structures. The performance of the systems was evaluated using the Librispeech and AMI corpus, following the visual-grounded contextual ASR pipeline. The findings indicate that using GNN encodings achieved consistent and significant reductions in word error rate (WER), particularly for words that are rare or have not been seen during the training process. Notably, the most effective combination of GNN encodings obtained more than 60% WER reduction for rare and unseen words compared to standard end-to-end systems.

Guangzhi Sun, Chao Zhang, Philip C. Woodland

Contextual knowledge is important for real-world automatic speech recognition (ASR) applications. In this paper, a novel tree-constrained pointer generator (TCPGen) component is proposed that incorporates such knowledge as a list of biasing words into both attention-based encoder-decoder and transducer end-to-end ASR models in a neural-symbolic way. TCPGen structures the biasing words into an efficient prefix tree to serve as its symbolic input and creates a neural shortcut between the tree and the final ASR output distribution to facilitate recognising biasing words during decoding. Systems were trained and evaluated on the Librispeech corpus where biasing words were extracted at the scales of an utterance, a chapter, or a book to simulate different application scenarios. Experimental results showed that TCPGen consistently improved word error rates (WERs) compared to the baselines, and in particular, achieved significant WER reductions on the biasing words. TCPGen is highly efficient: it can handle 5,000 biasing words and distractors and only add a small overhead to memory use and computation cost.

Guangzhi Sun, Chao Zhang, Phil Woodland

Significant progress has recently been made in speaker diarisation after the introduction of d-vectors as speaker embeddings extracted from neural network (NN) speaker classifiers for clustering speech segments. To extract better-performing and more robust speaker embeddings, this paper proposes a c-vector method by combining multiple sets of complementary d-vectors derived from systems with different NN components. Three structures are used to implement the c-vectors, namely 2D self-attentive, gated additive, and bilinear pooling structures, relying on attention mechanisms, a gating mechanism, and a low-rank bilinear pooling mechanism respectively. Furthermore, a neural-based single-pass speaker diarisation pipeline is also proposed in this paper, which uses NNs to achieve voice activity detection, speaker change point detection, and speaker embedding extraction. Experiments and detailed analyses are conducted on the challenging AMI and NIST RT05 datasets which consist of real meetings with 4--10 speakers and a wide range of acoustic conditions. For systems trained on the AMI training set, relative speaker error rate (SER) reductions of 13% and 29% are obtained by using c-vectors instead of d-vectors on the AMI dev and eval sets respectively, and a relative reduction of 15% in SER is observed on RT05, which shows the robustness of the proposed methods. By incorporating VoxCeleb data into the training set, the best c-vector system achieved 7%, 17% and16% relative SER reduction compared to the d-vector on the AMI dev, eval, and RT05 sets respectively

Guangzhi Sun, Yu Zhang, Ron J. Weiss et al.

This paper proposes a hierarchical, fine-grained and interpretable latent variable model for prosody based on the Tacotron 2 text-to-speech model. It achieves multi-resolution modeling of prosody by conditioning finer level representations on coarser level ones. Additionally, it imposes hierarchical conditioning across all latent dimensions using a conditional variational auto-encoder (VAE) with an auto-regressive structure. Evaluation of reconstruction performance illustrates that the new structure does not degrade the model while allowing better interpretability. Interpretations of prosody attributes are provided together with the comparison between word-level and phone-level prosody representations. Moreover, both qualitative and quantitative evaluations are used to demonstrate the improvement in the disentanglement of the latent dimensions.

Guangzhi Sun, Yu Zhang, Ron J. Weiss et al.

Recent neural text-to-speech (TTS) models with fine-grained latent features enable precise control of the prosody of synthesized speech. Such models typically incorporate a fine-grained variational autoencoder (VAE) structure, extracting latent features at each input token (e.g., phonemes). However, generating samples with the standard VAE prior often results in unnatural and discontinuous speech, with dramatic prosodic variation between tokens. This paper proposes a sequential prior in a discrete latent space which can generate more naturally sounding samples. This is accomplished by discretizing the latent features using vector quantization (VQ), and separately training an autoregressive (AR) prior model over the result. We evaluate the approach using listening tests, objective metrics of automatic speech recognition (ASR) performance, and measurements of prosody attributes. Experimental results show that the proposed model significantly improves the naturalness in random sample generation. Furthermore, initial experiments demonstrate that randomly sampling from the proposed model can be used as data augmentation to improve the ASR performance.

Guangzhi Sun, Chao Zhang, Phil Woodland

Speaker diarisation systems often cluster audio segments using speaker embeddings such as i-vectors and d-vectors. Since different types of embeddings are often complementary, this paper proposes a generic framework to improve performance by combining them into a single embedding, referred to as a c-vector. This combination uses a 2-dimensional (2D) self-attentive structure, which extends the standard self-attentive layer by averaging not only across time but also across different types of embeddings. Two types of 2D self-attentive structure in this paper are the simultaneous combination and the consecutive combination, adopting a single and multiple self-attentive layers respectively. The penalty term in the original self-attentive layer which is jointly minimised with the objective function to encourage diversity of annotation vectors is also modified to obtain not only different local peaks but also the overall trends in the multiple annotation vectors. Experiments on the AMI meeting corpus show that our modified penalty term improves the d- vector relative speaker error rate (SER) by 6% and 21% for d-vector systems, and a 10% further relative SER reduction can be obtained using the c-vector from our best 2D self-attentive structure.