Multimodal Large Language Models (MLLMs) have significantly advanced visual tasks by integrating visual representations into large language models (LLMs). The textual modality, inherited from LLMs, equips MLLMs with abilities like instruction following and in-context learning. In contrast, the visual modality enhances performance in downstream tasks by leveraging rich semantic content, spatial information, and grounding capabilities. These intrinsic modalities work synergistically across various visual tasks. Our research initially reveals a persistent imbalance between these modalities, with text often dominating output generation during visual instruction tuning. This imbalance occurs when using both full fine-tuning and parameter-efficient fine-tuning (PEFT) methods. We then found that re-balancing these modalities can significantly reduce the number of trainable parameters required, inspiring a direction for further optimizing visual instruction tuning. We introduce Modality Linear Representation-Steering (MoReS) to achieve the goal. MoReS effectively re-balances the intrinsic modalities throughout the model, where the key idea is to steer visual representations through linear transformations in the visual subspace across each model layer. To validate our solution, we composed LLaVA Steering, a suite of models integrated with the proposed MoReS method. Evaluation results show that the composed LLaVA Steering models require, on average, 500 times fewer trainable parameters than LoRA needs while still achieving comparable performance across three visual benchmarks and eight visual question-answering tasks. Last, we present the LLaVA Steering Factory, an in-house developed platform that enables researchers to quickly customize various MLLMs with component-based architecture for seamlessly integrating state-of-the-art models, and evaluate their intrinsic modality imbalance.

Automated audio captioning (AAC) is an audio-to-text task to describe audio contents in natural language. Recently, the advancements in large language models (LLMs), with improvements in training approaches for audio encoders, have opened up possibilities for improving AAC. Thus, we explore enhancing AAC from three aspects: 1) a pre-trained audio encoder via consistent ensemble distillation (CED) is used to improve the effectivity of acoustic tokens, with a querying transformer (Q-Former) bridging the modality gap to LLM and compress acoustic tokens; 2) we investigate the advantages of using a Llama 2 with 7B parameters as the decoder; 3) another pre-trained LLM corrects text errors caused by insufficient training data and annotation ambiguities. Both the audio encoder and text decoder are optimized by low-rank adaptation (LoRA). Experiments show that each of these enhancements is effective. Our method obtains a 33.0 SPIDEr-FL score, outperforming the winner of DCASE 2023 Task 6A.

Audio-text retrieval is a challenging task, requiring the search for an audio clip or a text caption within a database. The predominant focus of existing research on English descriptions poses a limitation on the applicability of such models, given the abundance of non-English content in real-world data. To address these linguistic disparities, we propose a language enhancement (LE), using a multilingual text encoder (SONAR) to encode the text data with language-specific information. Additionally, we optimize the audio encoder through the application of consistent ensemble distillation (CED), enhancing support for variable-length audio-text retrieval. Our methodology excels in English audio-text retrieval, demonstrating state-of-the-art (SOTA) performance on commonly used datasets such as AudioCaps and Clotho. Simultaneously, the approach exhibits proficiency in retrieving content in seven other languages with only 10% of additional language-enhanced training data, yielding promising results. The source code is publicly available https://github.com/zyyan4/ml-clap.

Recently, deep learning-based beamforming algorithms have shown promising performance in target speech extraction tasks. However, most systems do not fully utilize spatial information. In this paper, we propose a target speech extraction network that utilizes spatial information to enhance the performance of neural beamformer. To achieve this, we first use the UNet-TCN structure to model input features and improve the estimation accuracy of the speech pre-separation module by avoiding information loss caused by direct dimensionality reduction in other models. Furthermore, we introduce a multi-head cross-attention mechanism that enhances the neural beamformer's perception of spatial information by making full use of the spatial information received by the array. Experimental results demonstrate that our approach, which incorporates a more reasonable target mask estimation network and a spatial information-based cross-attention mechanism into the neural beamformer, effectively improves speech separation performance.

Electroencephalography (EEG) plays a vital role in detecting how brain responses to different stimulus. In this paper, we propose a novel Shallow-Deep Attention-based Network (SDANet) to classify the correct auditory stimulus evoking the EEG signal. It adopts the Attention-based Correlation Module (ACM) to discover the connection between auditory speech and EEG from global aspect, and the Shallow-Deep Similarity Classification Module (SDSCM) to decide the classification result via the embeddings learned from the shallow and deep layers. Moreover, various training strategies and data augmentation are used to boost the model robustness. Experiments are conducted on the dataset provided by Auditory EEG challenge (ICASSP Signal Processing Grand Challenge 2023). Results show that the proposed model has a significant gain over the baseline on the match-mismatch track.

In this paper, we investigate representation learning for low-resource keyword spotting (KWS). The main challenges of KWS are limited labeled data and limited available device resources. To address those challenges, we explore representation learning for KWS by self-supervised contrastive learning and self-training with pretrained model. First, local-global contrastive siamese networks (LGCSiam) are designed to learn similar utterance-level representations for similar audio samplers by proposed local-global contrastive loss without requiring ground-truth. Second, a self-supervised pretrained Wav2Vec 2.0 model is applied as a constraint module (WVC) to force the KWS model to learn frame-level acoustic representations. By the LGCSiam and WVC modules, the proposed small-footprint KWS model can be pretrained with unlabeled data. Experiments on speech commands dataset show that the self-training WVC module and the self-supervised LGCSiam module significantly improve accuracy, especially in the case of training on a small labeled dataset.

Existing weakly supervised sound event detection (WSSED) work has not explored both types of co-occurrences simultaneously, i.e., some sound events often co-occur, and their occurrences are usually accompanied by specific background sounds, so they would be inevitably entangled, causing misclassification and biased localization results with only clip-level supervision. To tackle this issue, we first establish a structural causal model (SCM) to reveal that the context is the main cause of co-occurrence confounders that mislead the model to learn spurious correlations between frames and clip-level labels. Based on the causal analysis, we propose a causal intervention (CI) method for WSSED to remove the negative impact of co-occurrence confounders by iteratively accumulating every possible context of each class and then re-projecting the contexts to the frame-level features for making the event boundary clearer. Experiments show that our method effectively improves the performance on multiple datasets and can generalize to various baseline models.

In most cases, bilingual TTS needs to handle three types of input scripts: first language only, second language only, and second language embedded in the first language. In the latter two situations, the pronunciation and intonation of the second language are usually quite different due to the influence of the first language. Therefore, it is a big challenge to accurately model the pronunciation and intonation of the second language in different contexts without mutual interference. This paper builds a Mandarin-English TTS system to acquire more standard spoken English speech from a monolingual Chinese speaker. We introduce phonology embedding to capture the English differences between different phonology. Embedding mask is applied to language embedding for distinguishing information between different languages and to phonology embedding for focusing on English expression. We specially design an embedding strength modulator to capture the dynamic strength of language and phonology. Experiments show that our approach can produce significantly more natural and standard spoken English speech of the monolingual Chinese speaker. From analysis, we find that suitable phonology control contributes to better performance in different scenarios.

Sequence expansion between encoder and decoder is a critical challenge in sequence-to-sequence TTS. Attention-based methods achieve great naturalness but suffer from unstable issues like missing and repeating phonemes, not to mention accurate duration control. Duration-informed methods, on the contrary, seem to easily adjust phoneme duration but show obvious degradation in speech naturalness. This paper proposes PAMA-TTS to address the problem. It takes the advantage of both flexible attention and explicit duration models. Based on the monotonic attention mechanism, PAMA-TTS also leverages token duration and relative position of a frame, especially countdown information, i.e. in how many future frames the present phoneme will end. They help the attention to move forward along the token sequence in a soft but reliable control. Experimental results prove that PAMA-TTS achieves the highest naturalness, while has on-par or even better duration controllability than the duration-informed model.

Neural networks based vocoders have recently demonstrated the powerful ability to synthesize high quality speech. These models usually generate samples by conditioning on some spectrum features, such as Mel-spectrum. However, these features are extracted by using speech analysis module including some processing based on the human knowledge. In this work, we proposed RawNet, a truly end-to-end neural vocoder, which use a coder network to learn the higher representation of signal, and an autoregressive voder network to generate speech sample by sample. The coder and voder together act like an auto-encoder network, and could be jointly trained directly on raw waveform without any human-designed features. The experiments on the Copy-Synthesis tasks show that RawNet can achieve the comparative synthesized speech quality with LPCNet, with a smaller model architecture and faster speech generation at the inference step.