Neural transducers offer an alignment-free framework for speech-to-text modeling, and hierarchical transducer architectures further improve multilingual joint automatic speech recognition (ASR) and speech translation (ST) by stacking a translation-focused encoder on top of an ASR encoder.However, extending hierarchical transducers to multilingual manyto-many settings remains challenging: fully shared models often suffer from negative transfer and unstable target-language generation, while training separate models for each direction is computationally prohibitive.We propose LCMA-SRT (Language-Conditional Mixtureof-Experts Adapters for Speech Recognition and Translation), which augments a hierarchical transducer with language-conditional Mixture-of-Experts (MoE) adapters.A sourceconditioned MoE adapter (SRC-MoE) uses source-language embeddings to reduce crosslanguage interference and improve multilingual ASR.A target-conditioned MoE adapter (TGT-MoE) uses the desired target language to reduce cross-target interference and stabilize targetlanguage generation in many-to-many ST.Experiments on Europarl-ST (9 languages, 72 directions) show that LCMA-SRT improves both ASR and ST within a single joint model, reducing average WER and improving BLEU and COMET over strong hierarchical transducer baselines.We release our code and models at https://github.com/linanjie0820/ LCMA-SRT.

This paper presents a computationally efficient approach for state-of-health (SOH) estimation. This approach has two novel components: large language model (LLM)-based prompt engineering for feature extraction and a edge-optimized extreme learning machine (TinyELM) for efficient SOH estimation. In our LLM-based feature engineering, we leverage GPT-5 to automatically generate potential health indicators from raw voltage and current data, then select two to four features that exhibit strong correlations with the target SOH. Building upon these features, we develop TinyELM as a streamlined and optimized variant of ELM designed for improved estimation accuracy and portability on resource-constrained edge devices. The performance of TinyELM is compared against several deep learning-based models as well as the conventional ELM model and linear-based models on three edge devices, namely, the STM32, Raspberry Pi 4, and Jetson Nano. Experimental results demonstrate that TinyELM achieves state-of-the-art performance in SOH estimation while delivering up to 28.4× faster inference than the conventional ELM and 3× faster inference than the other linear models. Furthermore, the results highlight the effectiveness of our LLM-based feature engineering, showing approximately 5× lower root mean square error on average compared to traditional and automated feature engineering approaches.

Empathy detection (ED) aims to understand the user’s empathy direction, while empathy intent classification (EIC) focuses on identifying the empathy intent behind the user’s utterance. Both tasks have garnered significant attention. Recent studies have shown that jointly training these tasks can improve model performance, as their correlation enhances the diversity of information. However, previous studies have relied solely on shallow information transfer between two task representations, failing to fully leverage the inter-task correlation, thus limiting performance. To this end, we propose a novel Label Co-occurrence Transformation Network (LCoT-Net), which models the correlation between the two tasks using the co-occurrence matrix of empathy and empathy intent labels as a medium. By performing category feature transformation at both the label and utterance levels, we achieve two-level mutual task guidance. Experimental results demonstrate that our model achieves competitive performance across various settings on two public datasets.

Traffic Engineering (TE), an important network application, manages and measures network traffic. It is responsible for monitoring network status, analyzing network traffic and realizing to improve overall network performance. However, when faced with different network applications and requirements, traditional network protocol is very onesided in considering link weight and the network equipment is difficult to obtain global information. Effective real-time and dynamic adjustion can not be guaranteed by traditional network. To solve the problem, an information entropy weight algorithm(IEW-algorithm) is proposed in this paper to calculate routing weights. This algorithm utilizes Software Defined Networking (SDN) with the character that centralizes all of network state information in SDN controller. It can be seen through the experiments that this algorithm can specify the optimal path for the application of different requirements based on link state.

This paper discusses compressive sensing with sparse measurement matrices. Sparse matrices have several attractive properties, like low computational complexity in both encoding and recovery, easy incremental updates to signals, and low storage requirement, etc. Typical examples of existing algorithms for sparse signal recovery with sparse measurement matrices include convex relaxation, matching pursuit, and Bayesian framework based approaches. In this paper, we propose an alternative technique to this problem. The proposed technique has a linear recovery complexity and relatively good empirical behavior. In this technique, we employ a permutation-based multi-dimensional measurement matrix, which is composed of several sub-matrices, each consisting of a block-diagonal matrix and a random permutation matrix. The measurement symbols generated from the same permutation matrix are referred to as a dimension. Such a measurement matrix brings some useful features to the measurement symbols. Fully exploiting these features enables us to design a simple and effective recovery algorithm. The proposed recovery algorithm employs an iterative process. In each iteration, the algorithm looks for measurement symbols with certain features, and uses such features to recover the source symbols related to these measurements. An iterative process is then applied, along with an interference cancellation operation, to reconstruct all source symbols gradually. The complexity of the proposed algorithm is relatively low, which grows linearly with the source signal length. Numerical results show that the proposed technique empirically offers a much lower sketch length than ℓ1-minimization-based convex relaxation and Bayesian framework based algorithms. It achieves the empirical lower bound of sketch length and linear recovery complexity at the same time.

This paper presents an efficient algorithm for the partial interference cancellation (PIC) group decoding. Partial interference cancellation is a linear operation that cancels interferences among symbol groups. The basic algorithm for the PIC operation requires a lot of matrix multiplications and inversions. Even though these matrix operations can be done in an efficient way by taking advantage of the Hermitian property, the computational burden is still high and may not always be acceptable. We propose a new algorithm for the PIC operation which requires much less and simpler matrix operations. The new algorithm significantly reduces the computational complexity of the PIC operation by one order of magnitude.

Uplink collaborative multiple-input-multiple-output (MIMO) processing among multiple base stations (BSs), also known as uplink multi-BS MIMO, is an effective way to solve the inter-cell interference in cellular systems. However, the significantly increased overhead on the backhaul network due to inter-BS information exchange has seriously restricted its application in practical systems with limited backhaul bandwidth. In this paper, we propose a three-stage technique for inter-BS information exchange in uplink multi-BS MIMO. A sub-optimal coordinated detection algorithm based on soft combining is considered to compromise between the performance and overhead. The major improvement over the existing solutions is that the proposed technique utilizes one property of soft combining to reduce the backhaul overhead: soft information combination is only needed for a very small part of the information bits that have mismatched local decoding results in different BSs. A three-stage information exchange technique is proposed to utilize this property, which allows the BSs to first locate the bits that need soft information combination and then exchange soft information for these bits only. By doing so, the backhaul overhead can be greatly reduced without performance degradation. Both simulation results and overhead analysis are used to demonstrate the benefit of the proposed technique. Comparisons with other coordinated detection algorithms like optimal joint detection and distributed interference cancellation show that it can provide a better tradeoff between the performance and overhead.

This paper discusses compressive sensing with digital sparse signals. The motivation is that most existing sparse signal recovery algorithms, like matching pursuit, convex relaxation and Bayesian framework, do not fully exploit the digital nature of signals when dealing with digital sparse signals, which result in certain performance losses. In this paper, we solve this problem via a permutation-based multi-dimensional sensing matrix and an iterative recovery algorithm with maximum likelihood (ML) local detectors. The sensing matrix considered consists of several sub-matrices, each composed of a random permutation matrix and a block-diagonal matrix. The measurements generated from the same permutation matrix are referred to as a dimension. The block-diagonal matrices allow the use of the low-complexity ML detector in each dimension, which best utilizes the digital nature of signals. The multi-dimensional structure of the sensing matrix enables information exchange between dimensions through an iterative process to achieve a near global-optimal estimation. Numerical results are used to show the rate-distortion performance of the proposed technique. It is shown that it can achieve much better rate-distortion than the existing approaches based on convex relaxation and Bayesian framework with digital source signals.

In this paper, we present new constructions of quasi-orthogonal space-time block codes for 4 and 6 transmit antennas. The proposed codes possess the following properties: full-rate, full transmit diversity, even power distribution and low decoding complexity, which make these two codes very practical spacetime coding schemes. For the case of 4 transmit antennas, the code can be decoded pair-by-pair, with each pair containing only two real symbols. It can be proved that for QAM signals, our proposed code has the same coding gain as the existing optimal quasi-orthogonal space-time block codes with minimum decoding complexity. For the case of 6 transmit antennas, the code can be decoded triplet-by-triplet, with each triplet containing three real symbols. We also present simple concatenation schemes of our proposed space-time block (STB) codes and multi-dimensional trellis coded modulation (MTCM) codes. Simulation results show that the performance of our concatenation scheme is comparable to the previously proposed STB-MTCM concatenation scheme while having a much lower decoding complexity.

In this paper, a correction is made to the criterion for a linear dispersion space-time block code to achieve full diversity with the partial interference cancellation (PIC) group decoding recently published in the above titled paper (ibid., vol. 55, no. 10, pp. 4366-4385).

Image segmentation is always very important for computer vision and pattern recognition. Moreover, how to fast extract objects from a given image is still a problem for real time image processing. Most of the traditional region-based models depend on global information to converge to minimum error segmentation, but they are always time-consuming, and result in no effective segmentation. In this paper, we propose a region-based model with weight matrix to detect objects fast based on least square method. The basic ideal of our model is to build up a minimum error functional by approximating objects and background of original image with two constants respectively. At the same time, we introduce a weight matrix into the region-based model, which can enhance the weight of objects while reducing the influence from background. Our method can fast converge through alternating iterations under least square method. We also compare it with other region-based methods to show the improvements that can be achieved. Experimental results show the advantages of our method in terms of efficiency in image segmentation without losing accuracy.

Cyclic division algebra (CDA) has recently become a major technique to construct space–time block codes with nonvanishing determinant (NVD). One of the key steps in this technique is the determination of non-norm elements and a sufficient condition for the determination has been given by Kiran and Rajan lately based on algebraic number theory. In this paper, based on Kiran and Rajan's condition, we present a more elementary condition for non-norm elements when signals are QAM or HEX, which is easier to check. With this elementary condition, non-norm elements with smaller absolute values than the existing ones can be found.

In this paper, we propose a partial interference cancellation (PIC) group decoding strategy/scheme for linear dispersive space-time block codes (STBC) and a design criterion for the codes to achieve full diversity when the PIC group decoding is used at the receiver. A PIC group decoding decodes the symbols embedded in an STBC by dividing them into several groups and decoding each group separately after a linear PIC operation is implemented. It can be viewed as an intermediate decoding between the maximum likelihood (ML) receiver that decodes all the embedded symbols together, i.e., all the embedded symbols are in a single group, and the zero-forcing (ZF) receiver that decodes all the embedded symbols separately and independently, i.e., each group has and only has one embedded symbol, after the ZF operation is implemented. The PIC group decoding provides a framework to adjust the complexity-performance tradeoff by choosing the sizes of the information symbol groups. Our proposed design criterion (group independence) for the PIC group decoding to achieve full diversity is an intermediate condition between the loosest ML full rank criterion of codewords and the strongest ZF linear independence condition of the column vectors in the equivalent channel matrix. We also propose asymptotic optimal (AO) group decoding algorithm which is an intermediate decoding between the MMSE decoding algorithm and the ML decoding algorithm. The design criterion for the PIC group decoding can be applied to the AO group decoding algorithm because of its asymptotic optimality. It is well-known that the symbol rate for a full rank linear STBC can be full, i.e.,nt, fornttransmit antennas. It has been recently shown that its rate is upper bounded by 1 if a code achieves full diversity with a linear receiver. The intermediate criterion proposed in this paper provides the possibility for codes of rates betweenntand 1 that achieve full diversity with the PIC group decoding. This therefore provides a complexity-performance-rate tradeoff. Some design examples are given.

In user cooperative communications, time offsets among the relay nodes may distort the structure of a distributed space-time code, and therefore impair the diversity gain. In this paper, a distributed linear convolutive space-time coding scheme is proposed for asynchronous scenario. We present a systematic construction method of distributed linear convolutive space-time codes that can achieve the full diversity under any delay profile. Furthermore, we show that with the MMSE decision feedback equalizer (MMSE-DFE) receiver, our proposed codes also achieve the full diversity, even when time offsets occur.

In this paper, we propose a partial interference cancellation (PIC) group decoding for linear dispersive space-time block codes (STBC) and a design criterion for the codes to achieve full diversity when the PIC group decoding is used at the receiver. A PIC group decoding decodes the symbols embedded in an STBC by dividing them into several groups and decoding each group separately after a linear PIC operation is implemented. It can be viewed as an intermediate decoding between the maximum likelihood (ML) receiver that decodes all the embedded symbols together, i.e., all embedded symbols are in a single group, and a zero-forcing (ZF) receiver that decodes all the embedded symbols separately and independently, i.e., each group has and only has one embedded symbol, after the ZF operation is implemented. The PIC group decoding provides a framework to adjust the complexity-performance tradeoff by choosing the sizes of the symbol groups. Our proposed design criterion (group independence) for the PIC group decoding to achieve full diversity is an intermediate condition between the loosest ML full rank criterion of codewords and the strongest ZF linear independence condition of the column vectors in the equivalent channel matrix. A design example is also given.

In user cooperative communications, time offsets among the relay nodes may distort the structure of a distributed space-time code, and therefore impair the diversity gain. In this paper, we first recall some of the existing schemes for dealing with this problem and then introduce a distributed linear space- time convolutional coding scheme for asynchronous scenario. We present a systematic construction of distributed linear space- time convolutional codes that can achieve the full diversity under any delay profile. Furthermore, we show that with the MMSE decision feedback equalizer (MMSE-DFE) receiver, our proposed codes also achieve the asynchronous full diversity.

Jing and Hassibi recently presented a detailed analysis on the problem of distributed space-time coding for a synchronized wireless relay network, where only linear transforms are implemented at relay nodes, and proposed a linear dispersion coding scheme. In this letter, we extend it to asynchronous wireless networks by employing OFDM to combat timing errors from relay nodes. Based on a layered structure, we present a distributed space-time code design achieving full spatial diversity for an asynchronous wireless relay network.

In user cooperative communications, time offsets among the relay nodes may distort the structure of a distributed space-time code, and therefore impair the diversity gain. In this paper, a distributed linear convolutive space-time coding scheme is proposed for asynchronous cooperative communication networks. We propose systematic construction methods of linear convolutive space-time codes (DLC-STC) that are .insensitive to the timing errors, i.e., they achieve the full spatial diversity under any delay profile. We then study the diversity property of the delay-tolerant DLC-STC with suboptimal receivers that have low decoding complexity. We show that with the ZF, MMSE and MMSE-DFE receivers, our proposed codes also achieve the full spatial diversity in the asynchronous scenario.

Cyclic division algebra (CDA) has recently become a major technique to construct space-time block codes with nonvanishing determinant (NVD). One of the key steps in this technique is the determination of non-norm elements and a sufficient condition for the determination has been given by Kiran and Rajan lately based on algebraic number theory. In this paper, based on Kiran and Rajan's condition, we present a more elementary condition for non-norm elements when signals are QAM, i.e., in Z[i], or HEX, i.e., Z[j], which is easier to check. With this elementary condition, non-norm elements with smaller absolute values than the existing ones can be found.