Marcelo Finger

Augusto Camargo, Marcelo Finger

Modern audio processing networks are commonly deployed on accelerators whose peak throughput is obtained through dense linear algebra, whereas conventional acoustic frontends -- a Short-Time Fourier Transform (STFT) followed by sparse Mel aggregation -- remain structurally heterogeneous. This mismatch can introduce memory-bandwidth, dispatch, and intermediate-allocation overheads on contemporary accelerator backends. This work introduces MelT, a single-stage frontend framework in which Mel-spaced Non-Uniform Discrete Fourier Transform (NDFT) bases are precomputed and applied to time-domain acoustic frames through dense General Matrix Multiplication (GEMM) operations. The contribution is not the NDFT operator itself; rather, it is the formulation of Mel-spaced NDFT projection as a GEMM-native audio frontend and its evaluation as a hardware-efficient alternative to conventional STFT+Mel pipelines. Evaluated across platforms ranging from Apple A18 Pro edge hardware to NVIDIA H100 datacenter acceleration, MelT attains up to a $3.75\times$ speedup in inference latency and a $3.52\times$ reduction in energy consumption while maintaining downstream classification accuracy.

Daniel Peixoto Pinto da Silva, Edresson Casanova, Lucas Rafael Stefanel Gris et al.

During the outbreak of COVID-19 pandemic, several research areas joined efforts to mitigate the damages caused by SARS-CoV-2. In this paper we present an interpretability analysis of a convolutional neural network based model for COVID-19 detection in audios. We investigate which features are important for model decision process, investigating spectrograms, F0, F0 standard deviation, sex and age. Following, we analyse model decisions by generating heat maps for the trained models to capture their attention during the decision process. Focusing on a explainable Inteligence Artificial approach, we show that studied models can taken unbiased decisions even in the presence of spurious data in the training set, given the adequate preprocessing steps. Our best model has 94.44% of accuracy in detection, with results indicating that models favors spectrograms for the decision process, particularly, high energy areas in the spectrogram related to prosodic domains, while F0 also leads to efficient COVID-19 detection.

Marcelo Matheus Gauy, Marcelo Finger

This work explores speech as a biomarker and investigates the detection of respiratory insufficiency (RI) by analyzing speech samples. Previous work \cite{spira2021} constructed a dataset of respiratory insufficiency COVID-19 patient utterances and analyzed it by means of a convolutional neural network achieving an accuracy of $87.04\%$, validating the hypothesis that one can detect RI through speech. Here, we study how Transformer neural network architectures can improve the performance on RI detection. This approach enables construction of an acoustic model. By choosing the correct pretraining technique, we generate a self-supervised acoustic model, leading to improved performance ($96.53\%$) of Transformers for RI detection.

Felipe R. Serras, Marcelo Finger

In this work, we carried out a study about the use of attention-based algorithms to automate the categorization of Brazilian case law documents. We used data from the Kollemata Project to produce two distinct datasets with adequate class systems. Then, we implemented a multi-class and multi-label version of BERT and fine-tuned different BERT models with the produced datasets. We evaluated several metrics, adopting the micro-averaged F1-Score as our main metric for which we obtained a performance value of F1-micro=0.72 corresponding to gains of 30 percent points over the tested statistical baseline. In this work, we carried out a study about the use of attention-based algorithms to automate the categorization of Brazilian case law documents. We used data from the \textit{Kollemata} Project to produce two distinct datasets with adequate class systems. Then, we implemented a multi-class and multi-label version of BERT and fine-tuned different BERT models with the produced datasets. We evaluated several metrics, adopting the micro-averaged F1-Score as our main metric for which we obtained a performance value of $\langle \mathcal{F}_1 \rangle_{micro}=0.72$ corresponding to gains of 30 percent points over the tested statistical baseline.

Maria Clara Ramos Morales Crespo, Maria Lina de Souza Jeannine Rocha, Mariana Lourenço Sturzeneker et al.

This paper presents the first publicly available version of the Carolina Corpus and discusses its future directions. Carolina is a large open corpus of Brazilian Portuguese texts under construction using web-as-corpus methodology enhanced with provenance, typology, versioning, and text integrality. The corpus aims at being used both as a reliable source for research in Linguistics and as an important resource for Computer Science research on language models, contributing towards removing Portuguese from the set of low-resource languages. Here we present the construction of the corpus methodology, comparing it with other existing methodologies, as well as the corpus current state: Carolina's first public version has $653,322,577$ tokens, distributed over $7$ broad types. Each text is annotated with several different metadata categories in its header, which we developed using TEI annotation standards. We also present ongoing derivative works and invite NLP researchers to contribute with their own.

Marcelo Matheus Gauy, Marcelo Finger

The goal of speech emotion recognition (SER) is to identify the emotional aspects of speech. The SER challenge for Brazilian Portuguese speech was proposed with short snippets of Portuguese which are classified as neutral, non-neutral female and non-neutral male according to paralinguistic elements (laughing, crying, etc). This dataset contains about $50$ minutes of Brazilian Portuguese speech. As the dataset leans on the small side, we investigate whether a combination of transfer learning and data augmentation techniques can produce positive results. Thus, by combining a data augmentation technique called SpecAugment, with the use of Pretrained Audio Neural Networks (PANNs) for transfer learning we are able to obtain interesting results. The PANNs (CNN6, CNN10 and CNN14) are pretrained on a large dataset called AudioSet containing more than $5000$ hours of audio. They were finetuned on the SER dataset and the best performing model (CNN10) on the validation set was submitted to the challenge, achieving an $F1$ score of $0.73$ up from $0.54$ from the baselines provided by the challenge. Moreover, we also tested the use of Transformer neural architecture, pretrained on about $600$ hours of Brazilian Portuguese audio data. Transformers, as well as more complex models of PANNs (CNN14), fail to generalize to the test set in the SER dataset and do not beat the baseline. Considering the limitation of the dataset sizes, currently the best approach for SER is using PANNs (specifically, CNN6 and CNN10).

Aline R. Ioste, Alan M. Durham, Marcelo Finger

We introduce CoLN, Combined Learning of Neural network weights, a novel method to securely combine Machine Learning models over sensitive data with no sharing of data. With CoLN, local hosts use the same Neural Network architecture and base parameters to train a model using only locally available data. Locally trained models are then submitted to a combining agent, which produces a combined model. The new model's parameters can be sent back to hosts, and can then be used as initial parameters for a new training iteration. CoLN is capable of combining several distributed neural networks of the same kind but is not restricted to any single neural architecture. In this paper we detail the combination algorithm and present experiments with feed-forward, convolutional, and recurrent Neural Network architectures, showing that the CoLN combined model approximates the performance of a hypothetical ideal centralized model, trained using the combination of the local datasets. CoLN can contribute to secure collaborative research, as required in the medical area, where privacy issues preclude data sharing, but where the limitations of local data demand information derived from larger datasets.

Marcelo Matheus Gauy, Natalia Hitomi Koza, Ricardo Mikio Morita et al.

We contrast high effectiveness of state of the art deep learning architectures designed for general audio classification tasks, refined for respiratory insufficiency (RI) detection and blood oxygen saturation (SpO$_2$) estimation and classification through automated audio analysis. Recently, multiple deep learning architectures have been proposed to detect RI in COVID patients through audio analysis, achieving accuracy above 95% and F1-score above 0.93. RI is a condition associated with low SpO$_2$ levels, commonly defined as the threshold SpO$_2$ <92%. While SpO$_2$ serves as a crucial determinant of RI, a medical doctor's diagnosis typically relies on multiple factors. These include respiratory frequency, heart rate, SpO$_2$ levels, among others. Here we study pretrained audio neural networks (CNN6, CNN10 and CNN14) and the Masked Autoencoder (Audio-MAE) for RI detection, where these models achieve near perfect accuracy, surpassing previous results. Yet, for the regression task of estimating SpO$_2$ levels, the models achieve root mean square error values exceeding the accepted clinical range of 3.5% for finger oximeters. Additionally, Pearson correlation coefficients fail to surpass 0.3. As deep learning models perform better in classification than regression, we transform SpO$_2$-regression into a SpO$_2$-threshold binary classification problem, with a threshold of 92%. However, this task still yields an F1-score below 0.65. Thus, audio analysis offers valuable insights into a patient's RI status, but does not provide accurate information about actual SpO$_2$ levels, indicating a separation of domains in which voice and speech biomarkers may and may not be useful in medical diagnostics under current technologies.

Augusto Camargo, Wesley Carvalho, Felipe Peressim et al.

In this paper, we studied whether models based on BiLSTM and BERT can predict hashtags in Brazilian Portuguese for Ecommerce websites. Hashtags have a sizable financial impact on Ecommerce. We processed a corpus of Ecommerce reviews as inputs, and predicted hashtags as outputs. We evaluated the results using four quantitative metrics: NIST, BLEU, METEOR and a crowdsourced score. A word cloud was used as a qualitative metric. While all computer-generated metrics (NIST, BLEU and METEOR) indicated bad results, the crowdsourced results produced amazing scores. We concluded that the texts predicted by the neural networks are very promising for use as hashtags for products on Ecommerce websites. The code for this work is available at https://github.com/augustocamargo/text-to-hashtag.

Marcelo Matheus Gauy, Marcelo Finger

An acoustic model, trained on a significant amount of unlabeled data, consists of a self-supervised learned speech representation useful for solving downstream tasks, perhaps after a fine-tuning of the model in the respective downstream task. In this work, we build an acoustic model of Brazilian Portuguese Speech through a Transformer neural network. This model was pretrained on more than $800$ hours of Brazilian Portuguese Speech, using a combination of pretraining techniques. Using a labeled dataset collected for the detection of respiratory insufficiency in Brazilian Portuguese speakers, we fine-tune the pretrained Transformer neural network on the following tasks: respiratory insufficiency detection, gender recognition and age group classification. We compare the performance of pretrained Transformers on these tasks with that of Transformers without previous pretraining, noting a significant improvement. In particular, the performance of respiratory insufficiency detection obtains the best reported results so far, indicating this kind of acoustic model as a promising tool for speech-as-biomarker approach. Moreover, the performance of gender recognition is comparable to the state of the art models in English.

Guilherme Lamartine de Mello, Marcelo Finger, and Felipe Serras et al.

In this paper we present PeLLE, a family of large language models based on the RoBERTa architecture, for Brazilian Portuguese, trained on curated, open data from the Carolina corpus. Aiming at reproducible results, we describe details of the pretraining of the models. We also evaluate PeLLE models against a set of existing multilingual and PT-BR refined pretrained Transformer-based LLM encoders, contrasting performance of large versus smaller-but-curated pretrained models in several downstream tasks. We conclude that several tasks perform better with larger models, but some tasks benefit from smaller-but-curated data in its pretraining.

Daniel Oliveira de Brito, Letícia Gabriella de Souza, Marcelo Matheus Gauy et al.

This technical report investigates the performance of pre-trained audio models on COVID-19 detection tasks using established benchmark datasets. We fine-tuned Audio-MAE and three PANN architectures (CNN6, CNN10, CNN14) on the Coswara and COUGHVID datasets, evaluating both intra-dataset and cross-dataset generalization. We implemented a strict demographic stratification by age and gender to prevent models from exploiting spurious correlations between demographic characteristics and COVID-19 status. Intra-dataset results showed moderate performance, with Audio-MAE achieving the strongest result on Coswara (0.82 AUC, 0.76 F1-score), while all models demonstrated limited performance on Coughvid (AUC 0.58-0.63). Cross-dataset evaluation revealed severe generalization failure across all models (AUC 0.43-0.68), with Audio-MAE showing strong performance degradation (F1-score 0.00-0.08). Our experiments demonstrate that demographic balancing, while reducing apparent model performance, provides more realistic assessment of COVID-19 detection capabilities by eliminating demographic leakage - a confounding factor that inflate performance metrics. Additionally, the limited dataset sizes after balancing (1,219-2,160 samples) proved insufficient for deep learning models that typically require substantially larger training sets. These findings highlight fundamental challenges in developing generalizable audio-based COVID-19 detection systems and underscore the importance of rigorous demographic controls for clinically robust model evaluation.

Vinicius Anjos de Almeida, Vinicius de Camargo, Raquel Gómez-Bravo et al.

Background: Medical coding structures healthcare data for research, quality monitoring, and policy. This study assesses the potential of large language models (LLMs) to assign ICPC-2 codes using the output of a domain-specific search engine. Methods: A dataset of 437 Brazilian Portuguese clinical expressions, each annotated with ICPC-2 codes, was used. A semantic search engine (OpenAI's text-embedding-3-large) retrieved candidates from 73,563 labeled concepts. Thirty-three LLMs were prompted with each query and retrieved results to select the best-matching ICPC-2 code. Performance was evaluated using F1-score, along with token usage, cost, response time, and format adherence. Results: Twenty-eight models achieved F1-score > 0.8; ten exceeded 0.85. Top performers included gpt-4.5-preview, o3, and gemini-2.5-pro. Retriever optimization can improve performance by up to 4 points. Most models returned valid codes in the expected format, with reduced hallucinations. Smaller models (<3B) struggled with formatting and input length. Conclusions: LLMs show strong potential for automating ICPC-2 coding, even without fine-tuning. This work offers a benchmark and highlights challenges, but findings are limited by dataset scope and setup. Broader, multilingual, end-to-end evaluations are needed for clinical validation.

Marcelo Finger

One of the main reasons to employ a description logic such as EL or EL++ is the fact that it has efficient, polynomial-time algorithmic properties such as deciding consistency and inferring subsumption. However, simply by adding negation of concepts to it, we obtain the expressivity of description logics whose decision procedure is {ExpTime}-complete. Similar complexity explosion occurs if we add probability assignments on concepts. To lower the resulting complexity, we instead concentrate on assigning probabilities to Axioms (GCIs). We show that the consistency detection problem for such a probabilistic description logic is NP-complete, and present a linear algebraic deterministic algorithm to solve it, using the column generation technique. We also examine and provide algorithms for the probabilistic extension problem, which consists of inferring the minimum and maximum probabilities for a new axiom, given a consistent probabilistic knowledge base.

Marcelo Finger

In this paper we show several similarities among logic systems that deal simultaneously with deductive and quantitative inference. We claim it is appropriate to call the tasks those systems perform as Quantitative Logic Reasoning. Analogous properties hold throughout that class, for whose members there exists a set of linear algebraic techniques applicable in the study of satisfiability decision problems. In this presentation, we consider as Quantitative Logic Reasoning the tasks performed by propositional Probabilistic Logic; first-order logic with counting quantifiers over a fragment containing unary and limited binary predicates; and propositional Lukasiewicz Infinitely-valued Probabilistic Logic

Felipe Salvatore, Marcelo Finger, Roberto Hirata

At present, different deep learning models are presenting high accuracy on popular inference datasets such as SNLI, MNLI, and SciTail. However, there are different indicators that those datasets can be exploited by using some simple linguistic patterns. This fact poses difficulties to our understanding of the actual capacity of machine learning models to solve the complex task of textual inference. We propose a new set of syntactic tasks focused on contradiction detection that require specific capacities over linguistic logical forms such as: Boolean coordination, quantifiers, definite description, and counting operators. We evaluate two kinds of deep learning models that implicitly exploit language structure: recurrent models and the Transformer network BERT. We show that although BERT is clearly more efficient to generalize over most logical forms, there is space for improvement when dealing with counting operators. Since the syntactic tasks can be implemented in different languages, we show a successful case of cross-lingual transfer learning between English and Portuguese.

Fabiano Ferreira Luz, Marcelo Finger

Semantic parsing can be defined as the process of mapping natural language sentences into a machine interpretable, formal representation of its meaning. Semantic parsing using LSTM encoder-decoder neural networks have become promising approach. However, human automated translation of natural language does not provide grammaticality guarantees for the sentences generate such a guarantee is particularly important for practical cases where a data base query can cause critical errors if the sentence is ungrammatical. In this work, we propose an neural architecture called Encoder CFG-Decoder, whose output conforms to a given context-free grammar. Results are show for any implementation of such architecture display its correctness and providing benchmark accuracy levels better than the literature.

Fabiano Ferreira Luz, Marcelo Finger

Semantic parsing is the process of mapping a natural language sentence into a formal representation of its meaning. In this work we use the neural network approach to transform natural language sentence into a query to an ontology database in the SPARQL language. This method does not rely on handcraft-rules, high-quality lexicons, manually-built templates or other handmade complex structures. Our approach is based on vector space model and neural networks. The proposed model is based in two learning steps. The first step generates a vector representation for the sentence in natural language and SPARQL query. The second step uses this vector representation as input to a neural network (LSTM with attention mechanism) to generate a model able to encode natural language and decode SPARQL.

Adolfo Neto, Celso A. A. Kaestner, Marcelo Finger

The KE inference system is a tableau method developed by Marco Mondadori which was presented as an improvement, in the computational efficiency sense, over Analytic Tableaux. In the literature, there is no description of a theorem prover based on the KE method for the C1 paraconsistent logic. Paraconsistent logics have several applications, such as in robot control and medicine. These applications could benefit from the existence of such a prover. We present a sound and complete KE system for C1, an informal specification of a strategy for the C1 prover as well as problem families that can be used to evaluate provers for C1. The C1 KE system and the strategy described in this paper will be used to implement a KE based prover for C1, which will be useful for those who study and apply paraconsistent logics.