Empirical Results for Adjusting Truncated Backpropagation Through Time while Training Neural Audio Effects

Dynamic range compression (DRC) by hexapeptide libraries increases MS/MS-based identification of lower-abundance proteins in complex mixtures. However, two unanswered questions impede fully realizing DRC's potential in shotgun proteomics. First, does DRC enhance identification of post-translationally modified proteins? Second, can DRC be incorporated into a workflow enabling relative protein abundance profiling? We sought to answer both questions analyzing human whole saliva. Addressing question one, we coupled DRC with covalent glycopeptide enrichment and MS/MS. With DRC we identified ∼2 times more N-linked glycoproteins and their glycosylation sites than without DRC, dramatically increasing the known salivary glycoprotein catalog. Addressing question two, we compared differentially stable isotope-labeled saliva samples pooled from healthy and metastatic breast cancer women using a multidimensional peptide fractionation-based workflow, analyzing in parallel one sample portion with DRC and one portion without. Our workflow categorizes proteins with higher absolute abundance, whose relative abundance ratios are altered by DRC, from proteins of lower absolute abundance detected only after DRC. Within each of these salivary protein categories, we identified novel abundance changes putatively associated with breast cancer, demonstrating feasibility and benefits of DRC for relative abundance profiling. Collectively, our results bring us closer to realizing the full potential of DRC for proteomic studies.

Computation in a typical Transformer-based large language model (LLM) can be characterized by batch size, hidden dimension, number of layers, and sequence length. Until now, system works for accelerating LLM training have focused on the first three dimensions: data parallelism for batch size, tensor parallelism for hidden size, and pipeline parallelism for model depth or layers. These widely studied forms of parallelism are not targeted or optimized for long sequence Transformer models. Given practical application needs for long sequence LLM, renewed attentions are being drawn to sequence parallelism. However, existing works in sequence parallelism are constrained by memory-communication inefficiency, limiting their scalability to long sequence large models. In this work, we introduce Ulysses, a novel, portable, and effective methodology for enabling highly efficient and scalable LLM training with extremely long sequence length. Ulysses at its core partitions input data along the sequence dimension and employs an efficient all-to-all collective communication for attention computation. Theoretical communication analysis shows that, whereas other methods incur communication overhead as sequence length increases, Ulysses maintains constant communication volume when sequence length and compute devices are increased proportionally. Furthermore, experimental evaluations show that Ulysses scales to more than 1 million context length and trains 2.5x faster with 4x longer sequence length than the existing method SOTA baseline.

Dynamic Range Compression (DRC) is a key component in all modern Hearing Aids. Attack and Release time constants decide the speed with which the DRC should act to the incoming signal amplitude variation. So an accurate estimation of time constants gives a precise control over the DRC behavior. In this paper we examined various errors that occur in the output of the DRC while using conventional methods which affect attack and release time constants adversely. New methods are proposed for a better estimation of time constants in DRC. Since all the modifications are made in the estimation of attack time, there is no need to change the existing hardware for DRC. The proposed algorithm gives perfect output characteristics with zero error for test signals defined in ANSI S3.22 standards for hearing aid specifications.

For kge 3, a k-rollercoaster is a sequence of numbers whose every maximal contiguous subsequence, that is increasing or decreasing, has length at least k; 3-rollercoasters are called simply rollercoasters. Given a sequence of distinct real numbers, we are interested in computing its maximum-length (not necessarily contiguous) subsequence that is a k-rollercoaster. Biedl et al. (in: ICALP, volume 107 of LIPIcs. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, pp 18:1–18:15, 2018) have shown that each sequence of n distinct real numbers contains a rollercoaster of length at least lceil n/2rceil for n>7, and that a longest rollercoaster contained in such a sequence can be computed in O(nlog n)-time (or faster, in O(n log log n) time, when the input sequence is a permutation of {1,ldots ,n}). They have also shown that every sequence of ngeqslant (k-1)^2+1 distinct real numbers contains a k-rollercoaster of length at least frac{n}{2(k-1)}-frac{3k}{2}, and gave an O(nklog n)-time (respectively, O(n klog log n)-time) algorithm computing a longest k-rollercoaster in a sequence of length n (respectively, a permutation of {1,ldots ,n}). In this paper, we give an O(nk^2)-time algorithm computing the length of a longest k-rollercoaster contained in a sequence of n distinct real numbers; hence, for constant k, our algorithm computes the length of a longest k-rollercoaster in optimal linear time. The algorithm can be easily adapted to output the respective k-rollercoaster. In particular, this improves the results of Biedl et al. (2018), by showing that a longest rollercoaster can be computed in optimal linear time. We also present an algorithm computing the length of a longest k-rollercoaster in O(n log ^2 n)-time, that is, subquadratic even for large values of kle n. Again, the rollercoaster can be easily retrieved. Finally, we show an Omega (n log k) lower bound for the number of comparisons in any comparison-based algorithm computing the length of a longest k-rollercoaster.

Medical endoscopic video processing requires real-time execution of color component acquisition, color filter array (CFA) demosaicing, and high dynamic range (HDR) compression under low-light conditions, while adhering to strict thermal constraints within the surgical handpiece. Traditional hardware-aware neural architecture search (NAS) relies on fixed hardware design spaces, making it difficult to balance accuracy, power consumption, and real-time performance. A collaborative “power-accuracy” optimization method is proposed for hardware-aware NAS. Firstly, we proposed a novel hardware modeling framework by abstracting FPGA heterogeneous resources into unified cell units and establishing a power–temperature closed-loop model to ensure that the handpiece surface temperature does not exceed clinical thresholds. In this framework, we constrained the interstage latency balance in pipelines to avoid routing congestion and frequency degradation caused by deep pipelines. Then, we optimized the NAS strategy by using pipeline blocks and combined with a hardware efficiency reward function. Finally, color component acquisition, CFA demosaicing, dynamic range compression, dynamic precision quantization, and streaming architecture are integrated into our framework. Experiments demonstrate that the proposed method achieves 2.8 W power consumption at 47 °C on a Xilinx ZCU102 platform, with a 54% improvement in throughput (vs. hardware-aware NAS), providing an engineer-ready lightweight network for medical edge devices such as endoscopes.

BackgroundNext generation sequencing (NGS) enables a more comprehensive analysis of bacterial diversity from complex environmental samples. NGS data can be analysed using a variety of workflows. We test several simple and complex workflows, including frequently used as well as recently published tools, and report on their respective accuracy and efficiency under various conditions covering different sequence lengths, number of sequences and real world experimental data from rhizobacterial populations of glyphosate-tolerant maize treated or untreated with two different herbicides representative of differential diversity studies.ResultsAlignment and distance calculations affect OTU estimations, and multiple sequence alignment exerts a major impact on the computational time needed. Generally speaking, most of the analyses produced consistent results that may be used to assess differential diversity changes, however, dataset characteristics dictate which workflow should be preferred in each case.ConclusionsWhen estimating bacterial diversity, ESPRIT as well as the web-based workflow, RDP pyrosequencing pipeline, produced good results in all circumstances, however, its computational requirements can make method-combination workflows more attractive, depending on sequence variability, number and length.

Directional transfer functions (DTFs), the directional components of head related transfer functions, are generally measured at finite locations in azimuth and elevation. Thus models are needed to synthesize DTFs at finer spatial resolution than the measured data to create complete virtual auditory displays. Here, minimum-phase all-pole and all-zero models were used for modeling both human and cat DTFs. For the human DTFs, model orders were chosen to achieve specific objective error criteria published in previous studies that were based on subjective listening tests. Because subjective listening tests are not always feasible in animals, objective methods must be used to assess the quality of the DTF reconstructions. Here, the same error criteria reported in subjective tests of human DTF reconstructions were used to constrain models of cat DTFs based on the assumption that if humans cannot discriminate reconstructed versus empirical DTFs for a given objective reconstruction error, then cats will not be able to either provided the reconstruction error is held below that criteria. All-pole and all-zero models of orders as low as 25 were able to model DTFs with errors comparable to previous research findings and preserve the main spectral features in both human and cat DTFs. [Work supported by NIH Grant No. R01-DC6865.]

Recent advancement of large-scale pretrained models such as BERT, GPT-3, CLIP, and Gopher, has shown astonishing achievements across various task domains. Unlike vision recognition and language models, studies on general-purpose user representation at scale still remain underexplored. Here we explore the possibility of general-purpose user representation learning by training a universal user encoder at large scales. We demonstrate that the scaling law is present in user representation learning areas, where the training error scales as a power-law with the amount of computation. Our Contrastive Learning User Encoder (CLUE), optimizes task-agnostic objectives, and the resulting user embeddings stretch our expectation of what is possible to do in various downstream tasks. CLUE also shows great transferability to other domains and companies, as performances on an online experiment shows significant improvements in Click-Through-Rate (CTR). Furthermore, we also investigate how the model performance is influenced by the scale factors, such as training data size, model capacity, sequence length, and batch size. Finally, we discuss the broader impacts of CLUE in general.

Accurate estimation of decay coefficients for dynamic range compressors in hearing aids and a hardware level comparison of different architectures

The rising machine learning (ML) models have become the preferred way for landslide detection based on remote sensing images, but the performance of these models in a sample-free area are rarely concerned in many studies. In this study, we used a cross-validation method (training model in one area and validation in another) to compare the model portability of trained ML models applied in an “off-site” area, as a consideration of the landslide detection ability of these models in sample-free areas. We integrate nighttime light imagery, multi-seasonal optical Landsat time-series and digital elevation data, and we employed support vector machines (SVM), artificial neural networks (ANN) and random forest (RF) models to classify the satellite imagery and identify landslides. Samples of two scenarios generated from two subareas of the Jiuzhaigou disaster-stricken region are used for the cross-application and accuracy evaluation of three ML models. The results revealed that when the trained models are applied in areas outside those in which they were developed, the landslide identification accuracy of these three models has declined. Especially for the SVM and ANN models, the accuracy is greatly reduced and there appears a seriously imbalanced user’s and producer’s accuracy. However, although the performance of the RF model is lower than that of SVM and ANN models in their local area, the RF model exhibits stable portability, and retains the original performance and achieves a satisfactory balance between overestimation and underestimation in “off-site” areas. An additional validation from a new area proved that the landslide detection performance of the RF model with stable portability is higher than that of the SVM and ANN models in “off-site” areas. The results suggest that evaluating the model portability through cross-application can be a useful way to determine the most suitable model for landslide detection in “off-site” areas with a similar geographic environment to model development areas, so as to maximize the accuracy of landslide detection based on limited samples.

<p><b>Speech enhancement was, is, and will be the key technology for digital speech transmission. When developing speech enhancement algorithms, the intelligibility or the quality of the processed speech needs to be evaluated. Intelligibility is more fundamental than quality. The evaluation of speech intelligibility can be carried out through subjective listening tests and objective metrics. Carrying out subjective listening tests is time-consuming and costly. Using the objective metrics is more efficient. Thus, there has been an increasing research in speech intelligibility prediction.</b></p> <p> Speech intelligibility can be measured in terms of the information received by a listener. This thesis aims at developing a mutual information-based speech intelligibility predictor (SIP) and using the mutual information-based predictor to assist the fitting process in hearing instruments. To achieve this goal, this thesis carried out three studies. </p> <p>First, it studied the modeling of the transmitted message. For mutual information-based SIPs, the most important thing is to determine the transmitted message. This thesis studied two approaches of modeling of the message: one is the continuous-valued sound, and the other is the discrete-valued linguistic message. Two corresponding SIPs were developed. By comparing their predicted intelligibility results with the psychometric curves, which are the subjective intelligibility scores, it shows that the modeling of discrete-valued message gives a better match to the psychometric curves.</p> <p> Second, based on the modeling of the discrete-valued message, this thesis proposed a mutual information-based SIP. Since the discrete-valued message cannot be obtained from a speech signal, the proposed SIP calculates the mutual information between the clean speech and the received speech, instead of calculating the mutual information between the message and the received speech. The proposed SIP considers frequency correlation for both the clean speech and the received speech. The evaluation results show that the proposed SIP performs better than the existing state-of-the-art mutual information-based SIPs.</p> <p>Third, this thesis proposed an automatic fitting tool for the nonlinear frequency compression (NFC) operator, which is a frequency lowering operator used in hearing instruments. The automatic fitting tool adjusts the parameter in the NFC by maximizing the mutual information between the message and the frequency-lowered speech. To evaluate the automatic fitting tool, the parameter was also searched by listening tests. The results show that the parameter determined by the automatic fitting tool is consistent to the parameter determined by the listening tests.</p>

Neural networks have topped performance measures across a wide variety of computational tasks. These performances are prevalent within the domain of human perception type tasks such as classification or generation of images, audio, or text. Recurrent neural networks are neural network architecture of choice for time-domain data such as spoken or written human language. Recurrent neural networks have also shown promise for tasks in the specialized signal domain of music. This thesis explores using recurrent neural networks to model the particular musical qualities generated by analog electronic musical equipment. The electric guitar amplifier is used by electric guitar players to shape the timbre of their musical instrument as they play. Most professionals consider analog amplifiers designed to provide acoustic distortion with vacuum tubes as having the best sound and feel for musicians. We attempted to model the sound transformation of a vacuum tube based electric guitar amplifier using a convolutional recurrent neural network architecture. For our experiment, we trained recurrent neural networks of various architectures using inputs of electric guitar signals and the subsequent signal processed through a typical vacuum-tube based amplifier and audio recording equipment. Training data was collected according to the recommended specifications of previous experiments. The amp simulation models were compared against the original amplifier with signal analysis and subjective listening tests. Sound recording techniques for capturing the best input and output data for the current state-of-the-art were analyzed. Analysis of various model input configurations and hyperparameter settings also showed several major limitations in the ability of the model to accurately reproduce the acoustic properties of a signal chain with more complex distortion characteristics than those previously tested. The limitations were analyzed and features of a new model were proposed. Subjective listening tests with three expert musical listeners showed that even though listeners could identify the quality degradation of the model, some listeners found the signals provided by specific models to be of more interest to their musical tastes. These models tended to model an amplifier set to a smooth gentle tone, which the model made less harmonically rich. However, models trained on amplifier tones with dense harmonic distortion characteristics were uniformly judged as very poor quality. These models could not accurately reproduce the intense compression and non-linear distortion qualities and ended up sounding abrupt with a hissing buzz.

Synthesizing multistyle speech has been an important topic of research in recent years. In this research, 11 commonly encountered speech styles have been synthesized by varying pitch, duration, intensity, and the glottal excitation on the Klatt Synthesizer 88, KLSYN88. These 11 speech styles include angry, clear, 50% tasking, 70% tasking, fast, Lombard, loud, normal, question, slow, and soft. All of the styles proved to be intelligible and appropriately styled based on subjective listening tests. The parameter variations of the glottal excitation are based on the results of statistical analyses [K. Cummings, Analysis, Syn., and Rec. of Stressed Speech, Ph.D thesis, Georgia Inst. of Tech., 1992] that demonstrated the importance of glottal excitation changes in styled speech. These statistical analyses demonstrated that the glottal excitation of each of the eleven styles is significantly and identifiably different. One utterance of the work ‘‘hot’’ was synthesized in the normal style. The other ten styles of this word were synthesized by changing only the glottal excitation parameters of the normal utterance. Subjective listening tests with untrained subjects demonstrated that the eleven synthetic styles were perceivably and identifiably different.

The human singing voice is a remarkable instrument that compounds an immense amount of expressivity onto a single dimension. Apart from semantics and melody (pitch, duration and dynamics), accent, age, gender and emotion are all carried in the singing voice. While a single singing voice on its own is aesthetically pleasing to the ear, the addition of concurrent voices of different pitch is commonly known to be capable of producing a pleasing effect far greater than the sum of that produced by each contributing voice. This motivates the use of harmony in singing. Unfortunately, accompaniment voices are difficult to sing, even for professional singers. Thankfully singing synthesis has made it viable for this task to be undertaken by machines. The overall objective of this thesis is to advance today’s understanding of singing harmony and ultimately develop novel techniques for its synthetic reproduction. This is broken down into three parts. The first focuses on a psychophysical basis of harmony, the second focuses on the synthesis of the singing voice, while the third combines the first two to focus on the synthesis of harmonized singing. The first contribution is an attempt to find a psychoacoustic basis of harmony and presented in chapter 2. Apart from stationary harmony (chords, or sonorities: the aesthetics of a group of concurrent notes at one point of time), this also includes transitional harmony (chord progression, or resolution: the aesthetics of a similar group of notes progressing to another). In order to explain both stationary and transitional harmony, it introduces a theory of harmony based on the notions of interharmonic and subharmonic modulations. Acoustic measures of stationary and transitional harmony are proposed and the answers to five fundamental questions of psychoacoustic harmony are presented, both based on this theory. Correlations with existing music theory and perception statistics support this contribution with both stationary and transitional harmony. The second contribution is in the synthesis of the singing voice and presented in chapter 3. Modern singing synthesis methods are at best capable of word- level runtime synthesis, with only two known ones dedicated to realtime synthesis. This means that they are applicable only towards offline music production. A large part of the art of music and singing, however, is in realtime performance. With both of the existing realtime singing synthesis methods bounded by a phone- coverage to realtime-capability tradeoff, a need for one that overcomes it remains. A novel realtime singing synthesis system, SERAPHIM, is proposed as an answer to this. Apart from overcoming this phone-coverage to realtime-capability trade- off, subjective listening tests also showed that listeners preferred voices synthesized by SERAPHIM as opposed to other realtime systems. The third contribution is in the synthesis of singing harmony and presented in chapter 4. With this contribution, a novel method for singing harmony synthesis is proposed. Current implementations can be classified into pitch-inaccurate rule- based systems, timing-inaccurate inference-based systems, and hybrid systems that trade off between pitch inaccuracies and timing inaccuracies. This means that existing systems are vulnerable to either pitch errors, timing errors or both in different degrees of compromise. The challenge in the task was to overcome this compromise to develop a robust technique that is simultaneously resilient to both pitch and timing errors while producing harmonious accompaniment. Our strategy was to leverage on the pitch-accurate inference-based method while eliminating timing inaccuracies by use of machine-synchronization. Spectrograms revealed that harmonized voices produced by this method contain the least dissonances amongst existing methods. Subjective listening tests also showed that harmonized voices produced by this method are perceived to be the best sounding, both by vocal experts and by casual listeners. All in all, the work presented in this thesis contributes to the advancement of the psychoacoustic understanding and machine synthesis of singing harmony across one journal paper, three conference papers and three patents.

Magnetic resonance imaging (MRI) plays a significant role in medical detection technologies but is difficult to avoid suffering from un-uniformity, low brightness and noise pollution. These factors might affect the doctor`s diagnosis and can lead to misdiagnosis in practical applications. In this paper, a new MRI images enhancement strategy is proposed which consists of a dynamic range compression (DRC) operation and a details enhancement (DE) operation. The DRC can efficiently compress the high dynamic range to a suitable range and protect the difference between the pixels, and DE adopts an improved bilateral filtering combined with Laplacian pyramid to remove the noise and enhance the details. A series of experiments results of real MRI images demonstrate that the method is more effective than other methods.