Jan Johannsen, J. MacAllister, Tomáš Michálek et al. · 2005
Various authors have pointed out that humans can become quite adept at deriving phonetic transcriptions from speech spectrograms (as good as 90% accuracy at the phoneme level). In this paper, we describe an expert system which attempts to simulate this performance. The speech spectrogram expert (SPEX) is actually a society made up of three experts: a 2-dimensional vision expert, an acoustic-phonetic expert, and a phonetics expert. The visual reasoning expert finds important visual features of the spectrogram. The acoustic-phonetic expert reasons about how visual features relate to phonemes, and about how phonemes change visually in different contexts. The phonetics expert reasons about allowable phoneme sequences and transformations, and deduces an English spelling for phoneme strings. The speech spectrogram expert is highly interactive, allowing users to investigate hypotheses and edit rules.
Cited by 17
Jeff Tucker, Kathleen E. Wage, John R. Buck et al. · The Journal of the Acoustical Society of America · 2025
Spectrograms are used for time-frequency analysis and as preprocessing for signal classifiers and other algorithms. The conventional spectrogram is a tapered short-time Fourier transform, equivalent to a bank of bandpass filters. The taper defines filter-bank characteristics such as bandwidth and sidelobe levels. Although the conventional spectrogram uses minimal computational resources, its design requires a compromise between resolution and interference suppression. Adaptive spectrogram algorithms adjust the filter-bank based on incoming data, thereby allowing different bandwidth/sidelobe trade-offs at each frequency and time. Adaptation can simultaneously improve tonal resolution and reveal quiet sources but typically costs substantially more to implement. This paper presents an adaptive spectrogram designed for applications with limited computational resources, e.g., autonomous vehicles. The performance weighted blended (PWB) spectrogram combines the output of a set of conventional filter-banks designed with different tapers. By adapting its blend weights at each frequency and time, the new algorithm separates loud closely spaced tones and identifies quiet signals. Because it relies on conventional filter-banks, the PWB spectrogram requires significantly less computation than other adaptive algorithms that require expensive matrix computations. Analysis of underwater glider data demonstrates the algorithm's ability to reveal a quiet chirp signal in the presence of vehicle self-noise.
Cited by 1
David Sussillo, Anshul Kundaje, Dimitris Anastassiou · EURASIP Journal on Advances in Signal Processing · 2004
Cited by 75Open access
Xinhui Luo (11880086), Xueming Fang (22075701), Jie Liu (15128) et al. · Figshare · 2025
<div> In this study, we employed a combination of theoretical and experimental analyses to explore the effects of the physico-chemical properties of lignite samples and surfactants on lignite dust’s wettability, thereby improving dust control in coal mines. First, we measured and analysed the coal samples’ industrial composition, elemental composition and chemical structure. It was found that the selected lignite dust has high ash and low moisture content and contains many hydrophobic functional groups, resulting in poor wettability by water. Next, we conducted surface tension tests, contact angle tests and lignite dust settling experiments to screen 12 monomer surfactants, exploring the impact of these solutions on lignite dust wettability. Finally, considering all monomer surfactants’ abilities to reduce surface tension, decrease contact angles and promote dust settling in solutions, we selected five surfactants (AES, MES-30, AEO-9, CDEA and CHSB) for blending based on their excellent performance in tests. We prepared the blends of these five surfactants, each with a mass fraction of 0.5 wt%, in a 1:1 ratio, resulting in 10 blended solutions. We measured the performance of these solutions and revealed that the AES + AEO-9 blend demonstrated a significant synergistic effect, markedly enhancing the capture efficiency of water for lignite dust. </div>
Cited by 0Open access
David Sussillo, Anshul Kundaje, Dimitris Anastassiou · SHILAP Revista de lepidopterología · 2004
Cited by 1Open access
Podder, Prajoy · Figshare · 2025
The user who is interested in using this dataset can cite the following papers:<br><br>P. Podder, M. Zawodniok and S. Madria, "Deep Learning for UAV Classification: Impact of Noise and Multipath Fading in RF Signals," <i>2025 10th International Conference on Smart and Sustainable Technologies (SpliTech)</i>, Bol and Split, Croatia, 2025, pp. 1-6, doi: 10.23919/SpliTech65624.2025.11091682.<br><br>P. Podder, M. Zawodniok and S. Madria, "Deep Learning for UAV Detection and Classification via Radio Frequency Signal Analysis," <i>2024 25th IEEE International Conference on Mobile Data Management (MDM)</i>, Brussels, Belgium, 2024, pp. 165-174, doi: 10.1109/MDM61037.2024.00040.
Cited by 0Open access
Marcelo Araya‐Salas · 2020
A set of tools to generate dynamic spectrogram visualizations in video format.
Cited by 3Open access
Bahr, Niklas · Figshare · 2023
This animation shows the temporal and spectral propagation of a soliton fission process of a third order soliton. Additional, the spectrogram for each propagation stage is displayed. dashed line: Zero-Dispersion wavelength dotted line: resonant frequency in normal dispersion regime solid line: group-velocity parabola.
Cited by 0Open access
Qingfeng Ji, J. S. Zhang, Yuxin Wang · arXiv (Cornell University) · 2024
Transformer structures have demonstrated outstanding skills in the deep learning space recently, significantly increasing the accuracy of models across a variety of domains. Researchers have started to question whether such a sophisticated network structure is actually necessary and whether equally outstanding results can be reached with reduced inference cost due to its complicated network topology and high inference cost. In order to prove the Mixer's efficacy on three datasets Speech Commands, UrbanSound8k, and CASIA Chinese Sentiment Corpus this paper applies amore condensed version of the Mixer to an audio classification task and conducts comparative experiments with the Transformer-based Audio Spectrogram Transformer (AST)model. In addition, this paper conducts comparative experiments on the application of several activation functions in Mixer, namely GeLU, Mish, Swish and Acon-C. Further-more, the use of various activation functions in Mixer, including GeLU, Mish, Swish, and Acon-C, is compared in this research through comparison experiments. Additionally, some AST model flaws are highlighted, and the model suggested in this study is improved as a result. In conclusion, a model called the Audio Spectrogram Mixer, which is the first model for audio classification with Mixer, is suggested in this study and the model's future directions for improvement are examined.
Cited by 0Open access
François Auger, E. Chassande‐Mottin, Patrick Flandrin · Zenodo (CERN European Organization for Nuclear Research) · 2012
Publication in the conference proceedings of EUSIPCO, Bucharest, Romania, 2012
Cited by 2Open access
Bajwa, Muhammad Khurram Zahur · Zenodo (CERN European Organization for Nuclear Research) · 2025
This dataset was obtained by applying a processing pipeline on two datasets Mozilla Common Voice and Cambridge IELTS Listening Audios. Mozilla Common Voice is a public, crowd sourced speech dataset collected from volunteers worldwide. It contains recordings in multiple languages, accents, and speaking styles, each paired with a text transcript. For this study, only English recordings were used, covering different accents, genders, and natural speaking variations. The original audio (48~kHz) was down sampled to 16~kHz to match the experimental setup. This dataset offers a diverse and natural collection of human speech. Cambridge IELTS recordings are studio-produced, scripted conversations and monologues with clear pronunciation and minimal background noise, making them similar to synthetic or machine-like voices in forensic analysis. All files were resampled to 16~kHz and split into 2-second segments for feature extraction. Together, these datasets provide a balanced mix of natural human speech and structured chatbot-like audio, enabling reliable testing of classification models. Furthermore, we apply deep learning models on MEL spectrogram data in order to detect if an audio segment is coming from a human or a chatbot.
Cited by 0Open access
Robert Eklund, Gustav Peters, Gopal Ananthakrishnan et al. · 2014
Cited by 0
Maria Åkesson, Julia Adlercreutz, Maria Sandsten · 2025
The reassignment technique is used to increase localization for signal components in the time-frequency representation. The technique gives perfect localization for linear chirp signals and constant frequency signals. Despite the recent interest in using the reassignment technique to determine phase-shifts, no general expression of cross-spectrogram reassignment vectors exists. This submission aims to give proof of the generalization of the reassignment vectors to the reassigned cross-spectrogram. The benefit of reassignment for determining phase shift between simulated signals will be illustrated. We exemplify with the reassigned cross-spectrogram of a respiratory frequency and a heart rate variability signal.
Cited by 0
Yuan Gong, Yu-An Chung, James Glass · 2021
In the past decade, convolutional neural networks (CNNs) have been widely adopted as the main building block for endto-end audio classification models, which aim to learn a direct mapping from audio spectrograms to corresponding labels.To better capture long-range global context, a recent trend is to add a self-attention mechanism on top of the CNN, forming a CNN-attention hybrid model.However, it is unclear whether the reliance on a CNN is necessary, and if neural networks purely based on attention are sufficient to obtain good performance in audio classification.In this paper, we answer the question by introducing the Audio Spectrogram Transformer (AST), the first convolution-free, purely attention-based model for audio classification.We evaluate AST on various audio classification benchmarks, where it achieves new state-of-the-art results of 0.485 mAP on AudioSet, 95.6% accuracy on ESC-50, and 98.1% accuracy on Speech Commands V2.
Cited by 905
Rui Qin, Zhifen Zhang, Hwa Kian Chai et al. · Mechanical Systems and Signal Processing · 2025
• The paper presents a time–frequency analysis method with filter adaptive iteration. • The proposed method allows for a balance of overall and detailed information. • FBRS is more specific in depicting vital but localized time–frequency information. • Simulation data and three real cases validate the proposed method. Extracting valuable features from a signal’s Region of Interest (ROI) is critical for advanced signal analysis. However, existing methods often require domain-specific design assumptions and fixed filter structures, limiting their adaptability and generalization. To address these limitations, this paper proposes the Frequency Band Recalibration Spectrogram (FBRS), a novel time–frequency analysis method based on the wavelet packet energy distribution. FBRS adaptively optimizes filter bandwidth and dynamically generates filter distributions based solely on a signal’s intrinsic frequency characteristics. By iteratively reconstructing wavelet packet components and refining frequency resolution, FBRS adaptively constructs a signal-driven time–frequency representation by dynamically recalibrating frequency bands based on intrinsic energy distribution. While this results in improved resolution in important regions, the core contribution lies in its adaptive and data-driven spectrogram generation mechanism. To verify the effectiveness of the proposed method, extensive experiments were conducted on simulated signals and three real signals from multiple practical application scenarios. The experimental results demonstrate that FBRS is theoretically innovative and exhibits strong adaptability and descriptive resolution in various signal analysis scenarios. The adaptability and flexibility of the proposed method enable it to cross different domains, providing a new perspective and a powerful tool for signal analysis. The code of the proposed method has been made public in https://github.com/Qinr1026/Frequency-Band-Recalibration-Spectrogram .
Cited by 0Open access
Podder, Prajoy · Figshare · 2025
The user who is interested in using this dataset can cite the following papers:<br><br>P. Podder, M. Zawodniok and S. Madria, "Deep Learning for UAV Classification: Impact of Noise and Multipath Fading in RF Signals," <i>2025 10th International Conference on Smart and Sustainable Technologies (SpliTech)</i>, Bol and Split, Croatia, 2025, pp. 1-6, doi: 10.23919/SpliTech65624.2025.11091682.<br><br>P. Podder, M. Zawodniok and S. Madria, "Deep Learning for UAV Detection and Classification via Radio Frequency Signal Analysis," <i>2024 25th IEEE International Conference on Mobile Data Management (MDM)</i>, Brussels, Belgium, 2024, pp. 165-174, doi: 10.1109/MDM61037.2024.00040.
Cited by 0Open access
Bahr, Niklas · Figshare · 2023
This animation shows the temporal and spectral propagation of a soliton fission process of a third order soliton. Additional, the spectrogram for each propagation stage is displayed. dashed line: Zero-Dispersion wavelength dotted line: resonant frequency in normal dispersion regime solid line: group-velocity parabola.
Cited by 0Open access
Kangping Gao (11154309), Xinxin Xu (690528), Jiabo Li (1836313) et al. · Figshare · 2021
Cited by 0Open access
Richard A. Altes · The Journal of the Acoustical Society of America · 1983
Echolocation behavior and neuronal representations of sound suggest that animal sonar can be modeled in terms of a spectrogram correlation process. Recent experimental results, however, indicate that bats may be sensitive to the fine structure of the signal-echo cross-correlation function (J. A. Simmons, Science 204, 1336). This sensitivity cannot be explained by cross correlation of spectrograms. Spectrogram deconvolution or an analysis-by-synthesis procedure can be used to reconstruct a signal from its spectrogram, aside from an arbitrary phase shift. Cross correlation of two such reconstructed signals, with the same arbitrary phase shift assigned to both, yields the required sensitivity to cross-correlation fine structure. The waveform reconstruction procedure is accurate only if spectrogram filters are not smooth window functions as in man-made processors, but have large time-bandwidth product, e.g., sharp rise and slow decay in the frequency domain, as in critical band models of hearing. This filter property insures that the filter ambiguity function has the same or larger support in the range-Doppler plane as the signal ambiguity function, a necessary condition for spectrogram deconvolution.
Cited by 0Open access
Yandong Ru (18806183), Gaoyang An (18806186), Zheng Wei (1552159) et al. · Figshare · 2024
<div><p>CNN has demonstrated remarkable performance in EEG signal detection, yet it still faces limitations in terms of global perception. Additionally, due to individual differences in EEG signals, the generalization ability of epilepsy detection models is week. To address this issue, this paper presents a cross-patient epilepsy detection method utilizing a multi-head self-attention mechanism. This method first utilizes Short-Time Fourier Transform (STFT) to transform the original EEG signals into time-frequency features, then models local information using Convolutional Neural Network (CNN), subsequently captures global dependency relationships between features using the multi-head self-attention mechanism of Transformer, and finally performs epilepsy detection using these features. Meanwhile, this model employs a light multi-head attention mechanism module with an alternating structure, which can comprehensively extract multi-scale features while significantly reducing computational costs. Experimental results on the CHB-MIT dataset show that the proposed model achieves accuracy, sensitivity, specificity, F1 score, and AUC of 92.89%, 96.17%, 92.99%, 94.41%, and 96.77%, respectively. Compared to the existing methods, the method proposed in this paper obtains better performance along with better generalization.</p></div>
Cited by 0Open access