SERious-TIME: Speech Emotion Recognition Insight-Oriented Understanding of Signals through Timing and Latency Evaluation

Section 1. Dataset Title and Concept Dataset Name: SERious-TIMEFull Form: Speech Emotion Recognition Insight-Oriented Understanding of Signals through Timing and Latency Evaluation SERious-TIME is a structured experimental dataset designed to capture and analyze the latency behavior of acoustic feature extraction in Speech Emotion Recognition (SER) systems. The term “SERious” is intentionally constructed to reflect both the domain (SER) and the analytical depth of the dataset, where “I.O.U.S.” represents Insight-Oriented Understanding of Signals, emphasizing the goal of deriving precise, low-level performance insights from signal processing operations. The “TIME” component explicitly denotes the dataset’s focus on timing measurements, latency profiling, and real-time performance evaluation. The dataset supports empirical investigation into how different implementation strategies, specifically Python, NumPy, and compiled approaches such as Cython, behave under varying computational conditions. It is designed to enable reproducible analysis of per-feature latency, distributional characteristics such as P50, P95, and P99, and the interaction between algorithmic structure and execution efficiency. Section 2. Purpose and Scope of the Dataset The SERious-TIME dataset is created to support controlled, empirical evaluation of computational performance in Speech Emotion Recognition (SER), with a specific focus on acoustic feature extraction latency. The dataset is intended to quantify how different implementation strategies behave under real-time constraints, where timing reliability is as critical as average performance. The primary purpose of this dataset is to enable systematic comparison between pure Python, vectorized NumPy, and compiled implementations such as Cython, using consistent input signals and measurement protocols. It captures execution time distributions rather than single-point measurements, allowing analysis of median latency (P50), variability (standard deviation), and tail behavior (P95, P99), which are essential for real-time system validation. The scope of the dataset covers three representative categories of acoustic features commonly used in SER: arithmetic reductions, branching-based computations, and sub-frame decompositions. By including multiple frame sizes ranging from very small to large audio segments, the dataset supports analysis of scalability effects and identification of crossover points where one implementation becomes more efficient than another. The dataset is not intended for model training or emotion classification tasks. Its role is strictly computational benchmarking, focusing on the performance characteristics of feature extraction pipelines. It is designed for researchers and engineers working on real-time audio processing, embedded systems, and performance-critical machine learning pipelines who require reproducible, fine-grained latency analysis. Section 3. Experimental Design and Data Generation The SERious-TIME dataset is generated through a controlled measurement experiment designed to isolate the computational cost of acoustic feature extraction. The experimental setup ensures that all implementations operate on identical input data, under identical execution conditions, enabling direct and statistically valid comparison of latency behavior. Input signals consist of synthetic audio frames generated from a uniform distribution in the range -1, 1, using a fixed random seed to guarantee reproducibility. This approach eliminates variability caused by real-world audio characteristics while preserving the computational structure of feature extraction algorithms. Frame sizes span a wide range from small to large sample counts, representing practical configurations from short real-time windows to full-length audio segments. Each feature extraction function is executed repeatedly using a nested timing protocol. Multiple inner-loop repetitions are used to amortize measurement overhead, while a larger number of outer runs capture variability across executions. This produces a distribution of latency values for each configuration, rather than a single average, allowing robust statistical analysis. The dataset records execution times in microseconds for each run, along with derived statistics such as mean, standard deviation, and percentile values (P50, P95, P99). Measurements are performed after a warm-up phase to eliminate cold-start effects. All implementations are validated for numerical correctness prior to timing, ensuring that performance comparisons are made on functionally equivalent outputs. The experiment includes multiple implementation variants, specifically pure Python, NumPy-based vectorization, and compiled approaches equivalent to Cython. This design enables analysis of how execution models, memory behavior, and algorithm structure interact to produce different latency characteristics under identical workloads. Section 4. Dataset Contents The SERious-TIME dataset consists of six structured CSV files, each representing a distinct segment of the experimental results. These files collectively capture the full range of latency measurements, statistical summaries, and configuration-specific outputs required to reproduce and analyze the findings reported in the study. Each CSV file is organized in a tabular format where rows correspond to individual measurement instances or aggregated configurations, and columns represent timing metrics and experimental parameters. Core attributes include implementation type, feature type, frame size, and execution time measurements in microseconds. Additional columns provide derived statistics such as mean latency, standard deviation, and percentile values (P50, P95, P99), enabling direct evaluation of both central tendency and tail behavior. The dataset is divided to reflect different analytical perspectives of the experiment. This includes baseline timing results for each feature and implementation, frame-size sweep data for crossover analysis, and statistical outputs used for hypothesis testing. Separate files are also used to capture detailed distributions or repeated-run measurements, ensuring that raw experimental variability is preserved rather than reduced to summary statistics. Section 5. Experimental Results and Key Findings The SERious-TIME dataset captures consistent and statistically significant differences in latency behavior across implementation strategies for acoustic feature extraction in Speech Emotion Recognition. The results confirm that compiled approaches and vectorized libraries both deliver substantial performance gains over pure Python, but their relative advantage depends on feature characteristics and input size. For arithmetic reduction features such as RMS Energy, NumPy achieves the lowest mean latency at large frame sizes due to efficient SIMD utilization and optimized C-level kernels. However, for smaller frame sizes typical of real-time processing, compiled implementations exhibit comparable or superior performance due to lower fixed overhead. A clear crossover behavior is observed, where the dominant implementation shifts depending on the number of samples processed. For branching-based features such as Zero-Crossing Rate, compiled implementations consistently outperform NumPy across all tested frame sizes. This is attributed to the absence of intermediate array allocations and the use of single-pass loops, which reduce memory overhead and eliminate garbage collection effects. As a result, both mean latency and tail latency are significantly lower. The dataset also reveals that latency distribution characteristics differ across implementations. NumPy exhibits higher variability and elevated P99 latency due to occasional memory allocation and garbage collection events, while compiled implementations maintain more stable execution profiles. This distinction is critical in real-time systems, where worst-case latency rather than average performance determines system reliability. Overall, the results demonstrate that no single implementation strategy is universally optimal. Performance is governed by the interaction between algorithm structure, memory behavior, and input size. The dataset provides empirical evidence supporting a conditional selection approach, where implementation choice is tailored to feature type and operational constraints. Section 6. Reproducibility The SERious-TIME dataset is designed to support full reproducibility of the experimental results presented in the associated study. All data files are generated under controlled conditions with fixed input generation, consistent measurement protocols, and validated implementation outputs, ensuring that the dataset can be used to independently verify reported latency behavior, statistical results, and observed performance trends. Each CSV file contains sufficient information to reconstruct the experimental configurations, including feature type, implementation method, and frame size. The timing data can be directly reused to recompute summary statistics, validate percentile distributions, and reproduce comparative analyses across implementations without requiring additional preprocessing. In addition to the dataset, a complete replication script is described and provided within the paper. This script enables full reconstruction of the experimental pipeline, including data generation, execution, and measurement procedures. At the time of publication, the replication script is not yet included in the Zenodo archive but is planned for subsequent release. The dataset therefore serves as the primary artifact for immediate validation, while the script provides a pathway for full end-to-end reproducibility once publicly available. The dataset is intended for benchmarking, methodological validation, and system design analysis. It can be used to evaluate alternative implementations such as JIT compilation