Before Search Comes Normalization

TL;DR for operators

A voice product cannot safely assume that Arabic speech recognition output and canonical Quranic text are already comparable strings. The paper constructs a 2,290-pair mapping between Uthmani word forms and Standard Arabic, then embeds that resource in a deterministic normalization and retrieval system.¹

The practical result is substantial but bounded. After normalization, 5,669 of 6,236 verses—90.9%—become identical across the two written forms. The validator passes 122 of 124 cases in its released evaluation suite. Yet both failures come from the same architectural choice: an ASR substitution can accidentally create an exact match for the wrong verse, and the system accepts that candidate before comparing alternatives.

A second limit cannot be repaired by a stronger matcher. Across the complete Quran, 1,030 verses belong to groups that share the same normalized four-word opening. Some short inputs therefore do not contain enough text to identify a unique verse.

For operators, the paper supports a layered design: use domain-specific normalization to reconcile scripts, deterministic retrieval to control cost and audit behavior, longer-context requests when the text is ambiguous, and separate acoustic models when Tajweed assessment depends on pronunciation rather than transcription. The reported 98.4% suite score is evidence about this implementation, not proof of complete recitation assessment.

The transcript and the reference are not yet the same language interface

A Quranic voice application starts from a familiar workflow. Speech recognition produces Arabic text. The product then needs to identify which canonical passage was recited before it can compare the recitation with a reference or provide feedback.

The apparent simplicity disappears when the transcript and the reference use different orthographic conventions. Canonical Quranic text is commonly stored in Uthmani script, while general-purpose ASR is more likely to produce Standard Arabic spelling. Even when the words are semantically the same, the underlying strings can differ.

The paper measures how much dedicated normalization can close that gap. Its seven-step pipeline reduces the Uthmani and Standard representations of 5,669 of the Quran’s 6,236 verses to identical strings. That 90.9% result matters because it is achieved before verse identification. Retrieval receives a substantially cleaner comparison problem.

The key technical reason is a superscript alef encoded as U+0670. Its Standard-Arabic realization depends on the word. Among the paper’s 2,290 mapped word forms, 1,335 require a full alef while 362 require no replacement; the remaining forms contain other orthographic differences. A rule that simply deletes the character everywhere therefore cannot reproduce Standard spelling.

The authors instead align the complete 6,236-verse corpus in Uthmani and Simple-Arabic forms. Most verses can be aligned word by word; structurally mismatched cases receive sequence and dynamic-programming alignment. The resulting resource contains 2,290 unique word pairs derived from 8,881 U+0670 occurrences and covers 2,290 of 2,291 distinct U+0670-bearing forms.

For a product team, this is best understood as an interoperability layer. The problem is not cosmetic text cleanup. Domain knowledge is being converted into a deterministic translation between two interfaces used by different components of the system.

Retrieval becomes a control problem once exact matches can be wrong

After normalization, the validator searches in four stages. It begins with exact token matching, then expands through Quranic Arabic Corpus lemmas and roots, then permits relaxed token coverage, and finally uses character-level fuzzy similarity.

That ordering appears conservative: accept strong evidence first, invoke weaker heuristics only when necessary. The two failures in the released 124-case suite show the risk in treating the first exact match as decisive.

A substitution error can transform the transcript into a complete token match for a different verse. Because the exact-match layer terminates first, the system accepts the wrong candidate before a competing near-match is evaluated. One failure appears in the single-verse substitution category and the other in a multi-verse case with errors. The mechanism is the same.

This is more informative than the headline score. The validator passes 122 of 124 cases, including all 62 corpus-generated perturbation cases, for 98.4% suite accuracy. But the failure analysis identifies a concrete policy choice behind the errors: candidate precedence.

For systems that provide authoritative or sensitive feedback, retrieval order is therefore part of the decision policy. An exact match is evidence, but it need not be sufficient evidence to terminate search. Candidate comparison, confidence margins, or contextual checks may be warranted when an alternative explanation remains plausible.

That is a Cognaptus inference from the reported failure mechanism, not a tested improvement in this paper. The study demonstrates the failure; it does not establish which alternative ranking strategy would perform best.

Some ambiguity belongs to the text, not the matcher

The paper also performs a full-corpus census that changes how residual retrieval errors should be interpreted.

After normalization, 1,030 of 6,236 verses—16.5%—share the same four-word opening with at least one other verse. They form 369 ambiguous groups. The most frequently repeated normalized four-word opening appears in 88 verses across 20 surahs.

This is not primarily a normalization defect. Nor is it evidence that fuzzy matching needs another threshold. When several canonical passages begin with the same observed text, four words simply do not contain enough information to determine which passage the user intends.

The operational response changes accordingly. Instead of forcing a prediction, a product can request more recitation, retain several candidates, or use legitimate conversational context. More sophisticated matching cannot recover information that has not yet been observed.

The distinction is valuable beyond this application. A retrieval system should separate uncertainty caused by implementation from uncertainty caused by the information available to it. The former may justify engineering work. The latter requires additional evidence.

The benchmark supports the architecture, not a complete recitation judge

The released evaluation is encouraging but should be read at the scale of the evidence.

The 124-case harness was created by the system authors and combines 62 fixed-seed corpus-generated perturbations with 62 hand-written assertions. Development thresholds and evaluation are not separated by an independent held-out split. Sensitivity checks show that suite accuracy remains 98.4% across several multi-verse, coverage, and fuzzy-threshold settings. A fuzzy threshold of 0.65 raises the same suite to 99.2%, but the authors decline to adopt it because selecting it from the evaluation results would tune against the benchmark.

That restraint is appropriate. These sensitivity tests are robustness checks around the released suite, not a second independent validation exercise.

The deployment evidence is narrower still. From 1,407 user utterances, 34 transcripts satisfy an independent labelling rule requiring the normalized ASR text to form an exact contiguous span of exactly one verse. The validator identifies all 34 correctly. Because this procedure favors correctly transcribed inputs, 34 of 34 should not be interpreted as evidence of robustness to the noisy cases for which fuzzy search is most relevant.

There is also a capability boundary that text processing cannot cross. If a Tajweed error affects properties such as madd duration, ghunna quality, or idgham without changing the ASR transcript, the validator has no signal with which to detect it. That requires acoustic evidence.

Normalization is not complete either: 567 verses still differ across Uthmani and Standard forms after the full pipeline, involving 628 divergent word pairs across 263 distinct types. The mapping is also specific to the Hafs an Asim textual tradition represented by the source corpora.

Spend complexity where the failure actually occurs

The paper’s strongest operational contribution is not the claim that deterministic methods can replace learned systems. It is a clearer decomposition of where different kinds of complexity belong.

Use lexical normalization when the mismatch is orthographic and corpus-specific. Use retrieval logic when several textual candidates remain plausible. Ask for longer context when the available words are intrinsically ambiguous. Apply confidence or abstention policies when evidence is weak. Use acoustic modelling when the property being judged exists in sound rather than text.

That decomposition can reduce inference cost and improve auditability because not every stage requires a generative model. It also makes failure ownership clearer: a normalization miss, a retrieval-ranking error, an underdetermined short input, and an acoustic Tajweed mistake are four different problems.

The paper provides credible, reproducible evidence for the first two layers and a useful measurement of the third. It does not establish a complete recitation-assessment system. That narrower claim is more actionable: deterministic domain infrastructure can remove much of the avoidable uncertainty before a product spends additional modelling capacity on the uncertainty that remains.

Cognaptus: Automate the Present, Incubate the Future.