A simulated oscillator network classified two spoken English vowels with 100% accuracy after it was given extracted formant frequencies, according to the study. A second design that worked from raw vowel waveforms reached 93% on the same binary task.
The result came from a tightly defined comparison. The reported binary evaluation used 45 “ah” samples and 45 “iy” samples. The wider dataset contained 12 spoken English vowel categories, while the 100% score applied to the “ah” versus “iy” task.
The frequency-first network
The first architecture used extracted formants, frequency features used to represent each vowel. It was an oscillator-only computational model built from coupled ring oscillators. The model used ordinary differential equations and positive couplings because the analysis focused on frequency rather than phase.
The network parameters were tuned with Nevergrad and a gradient-free genetic algorithm. That search selected oscillator frequencies and coupling strengths without calculating gradients.
The proposed oscillator-only network was reported to contain 34 neurons, 32 parameters and 50 couplings. Those figures describe the simulated architecture used in the test.
On the binary task, the preprocessed oscillator-only design reached 100% accuracy. The figure applies to the reported “ah” versus “iy” evaluation, not to the full set of 12 vowel categories.
The raw-waveform pipeline scored lower
The second architecture took in raw vowel waveforms. Its oscillator front end compared outputs from adjacent oscillators with XOR gates, smoothed those signals with a moving-average low-pass filter, and converted the changing signal into a feature vector using zero-crossing counts, a tally of sign changes.
The final classifier was a perceptron, making this an oscillator-plus-perceptron system rather than an oscillator-only network. It achieved 93% accuracy on the binary test set.
The two scores came from different pipelines. One began with extracted frequency features, while the other extracted features from the waveform and passed them to a perceptron. The reported results were therefore a descriptive comparison between two proposed architectures.
The wider vowel test was more mixed
The researchers also explored the full 12-class problem. Training and validation performance reached 97%, but the authors reported overfitting and failures to generalize to the test set in some cases.
The 97% figure describes training and validation performance, not a consistently successful 12-vowel test.
Some other vowel classes performed around 80% to 90%, depending on how separable their formant frequencies were. Those results gave a more mixed picture than the perfect binary score.
Efficiency remains a simulation estimate
The paper also offered an efficiency estimate. The authors estimated that the architecture could perform several tens of thousands of classifications per second, with energy per inference in the nJ, or nanjoule, regime.
That figure was a projection from the computational architecture, not a measured hardware benchmark. The study used an ODE-based model of coupled ring oscillators, so the reported speed and energy remained simulation estimates.
The evidence was limited to simulations of oscillator architectures on a spoken-vowel dataset, with the central result coming from the binary “ah” versus “iy” task.
The authors state that the vowel dataset is available through the fancompute/wavetorch GitHub repository and was downloaded in January 2024.
The study received partial support from an Intel Corporation HIMON grant and European Union Horizon EU funding through grant agreement No. 101092096, known as PHASTRAC. The disclosure says Intel’s involvement was limited to literature discovery and study motivation. The authors declared no conflicts of interest.
The publication note describes the document as an accepted version of an Information paper from 2025 and displays placeholder journal-reference and DOI fields.
Paper data and sources
Original title: Oscillator-Based Processing Unit for Formant Recognition
Authors: Tamás Rudner-Halász, Wolfgang Porod, Gyorgy Csaba
Journal/Repository: Rudner-Halász, Tamás, Wolfgang Porod, and Gyorgy Csaba. "Oscillator-Based Processing Unit for Formant Recognition." Information 16, no. 7 (2025): 611
Status: Peer-reviewed
First online: 2026-08-25
DOI: 10.3390/info16070611
Original paper · Full text