When it comes to systems that require approximation reasoning, fuzzy logic has reliably shown to be an effective technique for conveying uncertainty. Power optimization [11], control engineering [13], and fault-tolerant systems [15] are just a few of the many areas that have made use of it since its creation by Zadeh [2]. By shifting from Type-1 to IT2-FLS, we were able to reduce limitations in handling higher-order uncertainties and increase flexibility through the FOU [14].

Robotic route planning [10], offshore wind turbine torque management [14], and energy systems have all made use of advanced fuzzy systems [11]. Fuzzy clustering and evolutionary algorithms are two optimization techniques that researchers have combined with Type-2 fuzzy models to create models that can withstand changes in data and context. The use of toolkits like PyIT2FLS has made deployment easier in open-source settings [4].

When it comes to processing signals and languages, fuzzy logic has found applications in areas such as medical diagnostics [12], pattern recognition [22], and biometric decision fusion [5]. Type-1 and Type-2 systems were compared for their accuracy in decision-making tasks [12], and the inferential capacities of BK subproduct logic were evaluated in other studies [24], [25]. Despite these developments, no studies have looked into post-ASR transcription correction, which is particularly problematic for low-resource languages that have a lot of ambiguity.

The data-intensive and often unintelligible statistical or deep learning approaches used by most ASR correction algorithms today are a serious problem. Sundanese and other low-resource languages lack rule-based fuzzy correction frameworks that account for their unique phonological features. Therefore, this research aims to investigate the underexplored and ignored subject of language-specific post-ASR correction using fuzzy modeling, despite the fact that the methodology and principles of fuzzy modeling are well-defined.

The dataset utilized in this study was sourced from Kaggle, specifically the Sundanese ASR corpus compiled by Erik Mardani [https://www.kaggle.com/datasets/erikmardani/sundanese-asr]. It comprises 827 Sundanese audio files in. wav format, accompanied by manually annotated transcriptions in .txt. Each audio sample comprises brief spoken snippets featuring quotidian idioms and specified items.

We picked 30 representative examples from the complete set, encompassing varied phonological patterns characteristic of Sundanese, including reduplications, compound phrases, and consonant variants. All audio files were resampled to 16 kHz and normalized utilizing Librosa. These samples were utilized for both Whisper transcription and fuzzy correction assessment.

As shown in Table 1, the following examples are drawn from the Sundanese speech dataset hosted on Kaggle. Each sample includes the audio file ID, the transcription generated by the Whisper ASR model, the manually verified ground truth, and the corresponding WER. These examples highlight the typical challenges encountered in ASR for low-resource languages - such as misrecognized named entities, verb distortions, and morphological inconsistencies. These observed patterns formed the basis for designing the fuzzy rule-based correction system proposed in this research.

The initial transcription was produced via OpenAI's Whisper-base paradigm, implemented through the Whisper Python package in Google Colab. The model operated in multilingual mode with automatic language identification. Despite Whisper's support for over 90 languages, it encounters difficulties with Sundanese owing to insufficient training exposure, leading to elevated WER.

The generated texts were archived with their respective accurate transcriptions for subsequent assessment and rectification.

A fuzzy rule-based correction layer was designed to improve the quality of transcriptions generated by Whisper. This module explicitly targets typical Sundanese verbal patterns that are frequently misrecognized by ASR systems. The criteria were manually built based on observed transcription errors in Whisper outputs, emphasizing phonological problems such as consonant change, reduplication, and complex named entities. For example, in the Whisper output "Ahmad Albar Allahok ninggal di Prins", the system misread the intended "Ahmad Albar olohok ningali Prince"—misrecognizing verbs (ningali → ninggal) and names (Prince → Prins) because of phonetic proximity and language unfamiliarity. Another instance includes "Pasya Ungu olohok ningali The Beatles", where "Pasya" is a phonetic departure from "Pasha", a typical Indonesian name. Each rule is encoded using an IT2-FLS, implemented via the open-source PyIT2FLS package [4]. The fuzzy rules model phonetic similarity through Gaussian-shaped membership functions with interval-valued uncertainty, enabling the system to adeptly translate imprecise or ambiguous signals into more precise language representations. The fuzzy inference engine employs a Mamdani-style framework, and defuzzification is executed via Karnik–Mendel type-reduction to yield a precise correction choice.

Fig. 1 illustrates the overall process, demonstrating how Whisper's raw transcription is processed by a fuzzy inference module informed by Sundanese-specific language rules to yield an enhanced output.

The parameters of the fuzzy membership functions (central, spread, and uncertainty bounds) were improved using a Genetic Algorithm (GA) to enhance correction accuracy. The GA was executed via PyGAD, featuring a population size of 30 and spanning 50 generations. The objective function minimized the WER between the corrected output and the ground truth. The optimization enabled the fuzzy model to enhance its generalization of adjustments across various samples.

We adopted WER as the principal parameter to evaluate the efficacy of the Whisper ASR and the suggested corrective method. WER was calculated utilizing the jiwer library, accounting for replacements, insertions, and deletions, to calculate WER shown as Eq. (1).

S + D + I

WER = , (1)

where S is the number of substitutions, D deletions, I insertions, and N the total number of reference words, comparisons were made between raw Whisper outputs and the fuzzy-corrected versions.

We assessed Whisper transcription accuracy across 30 representative audio samples using the WER metric. Table 1 shows that numerous common ASR error patterns include misrecognition of named entities, verb forms, and phonologically related words.

The average WER before adjustment was 0.838, with a standard deviation of 0.229, indicating significant variation between utterances. Fig. 2 depicts the range of WER values, which reveal that the majority of transcriptions had WER values ranging from 70% to 100%, demonstrating the necessity for a corrective mechanism.

In order to test how well the fuzzy rule-based system worked, we applied the correction mechanism to the initial ASR dataset. Observed error patterns in the dataset informed the application of a curated set of Sundanese-specific fuzzy linguistic rules to uncertain tokens. The rules were optimized using a GA and formed using interval Type-2 fuzzy logic, as described in Section 2.

The average WER for all evaluated samples decreased from 83.8% to 62.8% after fuzzy correction was applied, showing a general improvement of about 25%. This significant improvement proves that fuzzy correction is effective in reducing frequent misrecognitions in ASR outputs that are resource-constrained. Table 2 shows a comparison of the WER before and after fuzzy correction for 10 samples that are considered representative. There is a clear trend of fewer mistakes in all types of speech.

The application of a curated set of Sundanese-specific fuzzy linguistic rules to uncertain tokens. The rules were optimized using a GA and formed using interval Type-2 fuzzy logic, as described in Section 3.

The average WER for all evaluated samples decreased from 83.8% to 62.8% after fuzzy correction was applied, showing a general improvement of about 25%. This significant improvement proves that fuzzy correction is effective in reducing frequent misrecognitions in ASR outputs that are resource-constrained. Table 2 shows a comparison of the WER before and after fuzzy correction for 10 samples that are considered representative. There is a clear trend of fewer mistakes in all types of speech.

To improve the accuracy of the fuzzy rule-based correction system, a GA was used to optimize the interval Type-2 fuzzy membership functions. These parameters comprise the center, spread, and FOU values for each linguistic variable in the fuzzy rule base. Python's PyGAD package was used for the optimization process. The GA was set up with a population size of 30, an 8% mutation rate, and a run time of 50 generations. Each chromosome encoded a real-valued parameter vector that represented the fuzzy system's setup. The objective function reduced the WER between the corrected transcription and the original ground truth. To allow for normal GA maximization reasoning, the fitness score was set to a negative WER number.

The optimization set consisted of 20 Whisper transcription samples with a WER greater than 0.7. Every chromosome in each generation was decoded into a fuzzy system and tested on samples. The average WER across all samples was used as the fitness metric. Over generations, the GA improved the fuzzy rule parameters to better generalize to phonetic distortions and morphological inconsistencies in Sundanese ASR output.

The optimization results are displayed in Fig. 4, which depicts the GA's convergence curve. A persistent increasing trend in fitness indicates good learning and convergence, demonstrating that optimization helped to increase correction performance. The average WER decreased from 83.8% to 62.8% after optimization, with a reduction in standard deviation from 22.9% to 17.2%, demonstrating improvements in both accuracy and robustness.