Vision-language models have revolutionized human-computer interaction and shown significant progress in multi-modal tasks. However, applying these models to complex visual tasks like medical image analysis remains challenging. In this study, we propose ViLaM, a unified Vision-Language transformer model that integrates instruction tuning predicated on a large language model. This approach enables us to optimally utilize the knowledge and reasoning capacities of large pre-trained language models for an array of tasks encompassing both language and vision. We employ frozen pre-trained encoders to encode and align both image and text features, enabling ViLaM to handle a variety of visual tasks following textual instructions. Besides, we've designed cycle training for referring expressions to address the need for high-quality, paired referring expression datasets for training large models in terms of both quantity and quality. We evaluated ViLaM's exceptional performance on public general datasets and further confirmed its generalizability on medical datasets. Importantly, we've observed the model's impressive zero-shot learning ability, indicating the potential future application of ViLaM in the medical field.
The Conformer has become the most popular encoder model for automatic speech recognition (ASR). It adds convolution modules to a transformer to learn both local and global dependencies. In this work we describe a faster, more memory-efficient, and better-performing transformer, called Zipformer. Modeling changes include: 1) a U-Net-like encoder structure where middle stacks operate at lower frame rates; 2) reorganized block structure with more modules, within which we re-use attention weights for efficiency; 3) a modified form of LayerNorm called BiasNorm allows us to retain some length information; 4) new activation functions SwooshR and SwooshL work better than Swish. We also propose a new optimizer, called ScaledAdam, which scales the update by each tensor's current scale to keep the relative change about the same, and also explictly learns the parameter scale. It achieves faster convergence and better performance than Adam. Extensive experiments on LibriSpeech, Aishell-1, and WenetSpeech datasets demonstrate the effectiveness of our proposed Zipformer over other state-of-the-art ASR models. Our code is publicly available at https://github.com/k2-fsa/icefall.
As the uplink sensing has the advantage of easy implementation, it attracts great attention in integrated sensing and communication (ISAC) system. This paper presents an uplink ISAC system based on multi-input multi-output orthogonal frequency division multiplexing (MIMO-OFDM) technology. The mutual information (MI) is introduced as a unified metric to evaluate the performance of communication and sensing. In this paper, firstly, the upper and lower bounds of communication and sensing MI are derived in details based on the interaction between communication and sensing. And the ISAC waveform is optimized by maximizing the weighted sum of sensing and communication MI. The Monte Carlo simulation results show that, compared with other waveform optimization schemes, the proposed ISAC scheme has the best overall performance.
Prompts are crucial to large language models as they provide context information such as topic or logical relationships. Inspired by this, we propose PromptASR, a framework that integrates prompts in end-to-end automatic speech recognition (E2E ASR) systems to achieve contextualized ASR with controllable style of transcriptions. Specifically, a dedicated text encoder encodes the text prompts and the encodings are injected into the speech encoder by cross-attending the features from two modalities. When using the ground truth text from preceding utterances as content prompt, the proposed system achieves 21.9% and 6.8% relative word error rate reductions on a book reading dataset and an in-house dataset compared to a baseline ASR system. The system can also take word-level biasing lists as prompt to improve recognition accuracy on rare words. An additional style prompt can be given to the text encoder and guide the ASR system to output different styles of transcriptions. The code is available at icefall.
In this paper, we introduce Libriheavy, a large-scale ASR corpus consisting of 50,000 hours of read English speech derived from LibriVox. To the best of our knowledge, Libriheavy is the largest freely-available corpus of speech with supervisions. Different from other open-sourced datasets that only provide normalized transcriptions, Libriheavy contains richer information such as punctuation, casing and text context, which brings more flexibility for system building. Specifically, we propose a general and efficient pipeline to locate, align and segment the audios in previously published Librilight to its corresponding texts. The same as Librilight, Libriheavy also has three training subsets small, medium, large of the sizes 500h, 5000h, 50000h respectively. We also extract the dev and test evaluation sets from the aligned audios and guarantee there is no overlapping speakers and books in training sets. Baseline systems are built on the popular CTC-Attention and transducer models. Additionally, we open-source our dataset creatation pipeline which can also be used to other audio alignment tasks.
Deep learning has revolutionized the process of new material discovery, with state-of-the-art models now able to predict material properties based solely on chemical compositions, thus eliminating the necessity for material structures. However, this cost-effective method has led to a trade-off in model accuracy. Specifically, the accuracy of Chemical Composition-based Property Prediction Models (CPMs) significantly lags behind that of Structure-based Property Prediction Models (SPMs). To tackle this challenge, we propose an innovative Teacher-Student (T-S) strategy, where a pre-trained SPM serves as the 'teacher' to enhance the accuracy of the CPM. Leveraging the T-S strategy, T-S CrabNet has risen to become the most accurate model among current CPMs. Initially, we demonstrated the universality of this strategy. On the Materials Project (MP) and Jarvis datasets, we validated the effectiveness of the T-S strategy in boosting the accuracy of CPMs with two distinct network structures, namely CrabNet and Roost. This led to CrabNet, under the guidance of the T-S strategy, emerging as the most accurate model among the current CPMs. Moreover, this strategy shows remarkable efficacy in small datasets. When predicting the formation energy on a small MP dataset comprising merely 5% of the samples, the T-S strategy boosted CrabNet's accuracy by 37.1%, exceeding the enhancement effect of the T-S strategy on the whole dataset.
With the support of integrated sensing and communication (ISAC) technology, mobile communication system will integrate the function of wireless sensing, thereby facilitating new intelligent applications such as smart city and intelligent transportation. Due to the limited sensing accuracy and sensing range of single base station (BS), multi-BS cooperative sensing can be applied to realize high-accurate, long-range and continuous sensing, exploiting the specific advantages of large-scale networked mobile communication system. This paper proposes a cooperative sensing method suitable to mobile communication systems, which applies symbol-level sensing information fusion to estimate the location and velocity of target. With the demodulation symbols obtained from the echo signals of multiple BSs, the phase features contained in the demodulation symbols are used in the fusion procedure, which realizes cooperative sensing with the synchronization level of mobile communication system. Compared with the signal-level fusion in the area of distributed aperture coherence-synthetic radars, the requirement of synchronization is much lower. When signal-to-noise ratio (SNR) is -5 dB, it is evaluated that symbol-level multi-BS cooperative sensing effectively improves the accuracy of distance and velocity estimation of target. Compared with single-BS sensing, the accuracy of distance and velocity estimation is improved by 40% and 72%, respectively. Compared with data-level multi-BS cooperative sensing based on maximum likelihood (ML) estimation, the accuracy of location and velocity estimation is improved by 12% and 63%, respectively. This work may provide a guideline for the design of multi-BS cooperative sensing system to exploit the widely deployed networked mobile communication system.
Neural Transducer and connectionist temporal classification (CTC) are popular end-to-end automatic speech recognition systems. Due to their frame-synchronous design, blank symbols are introduced to address the length mismatch between acoustic frames and output tokens, which might bring redundant computation. Previous studies managed to accelerate the training and inference of neural Transducers by discarding frames based on the blank symbols predicted by a co-trained CTC. However, there is no guarantee that the co-trained CTC can maximize the ratio of blank symbols. This paper proposes two novel regularization methods to explicitly encourage more blanks by constraining the self-loop of non-blank symbols in the CTC. It is interesting to find that the frame reduction ratio of the neural Transducer can approach the theoretical boundary. Experiments on LibriSpeech corpus show that our proposed method accelerates the inference of neural Transducer by 4 times without sacrificing performance. Our work is open-sourced and publicly available https://github.com/k2-fsa/icefall.
Connectionist Temporal Classification (CTC) suffers from the latency problem when applied to streaming models. We argue that in CTC lattice, the alignments that can access more future context are preferred during training, thereby leading to higher symbol delay. In this work we propose the delay-penalized CTC which is augmented with latency penalty regularization. We devise a flexible and efficient implementation based on the differentiable Finite State Transducer (FST). Specifically, by attaching a binary attribute to CTC topology, we can locate the frames that firstly emit non-blank tokens on the resulting CTC lattice, and add the frame offsets to the log-probabilities. Experimental results demonstrate the effectiveness of our proposed delay-penalized CTC, which is able to balance the delay-accuracy trade-off. Furthermore, combining the delay-penalized transducer enables the CTC model to achieve better performance and lower latency. Our work is open-sourced and publicly available https://github.com/k2-fsa/k2.