Homophone Normalization Methods

• Homophone normalization is a family of operations that converts same-sounding tokens into canonical written or spoken forms to resolve pronunciation-based ambiguities.
• Techniques include pronunciation-aware embedding, candidate generation with contextual reranking, selective phonetic backoff, and deterministic canonicalization to address challenges in noisy text, ASR, and MT.
• Evaluation metrics such as BLEU and WER reveal trade-offs between improved robustness and preserving valid orthographic variation in different application settings.

Homophone normalization denotes a family of normalization and robustness operations in which forms that are identical or similar in pronunciation are mapped, ranked, or reinterpreted so that a system produces a canonical written form, a canonical spoken form, or a pronunciation-aware intermediate representation. In current research, the target varies by application: noisy text may be normalized into standard orthography, written text may be verbalized for speech systems, ASR or NMT models may be made robust to homophone substitutions without explicitly restoring the original token, and script-level variants may be collapsed into a canonical graphemic form (Doshi et al., 2020, Wu et al., 2016, Liu et al., 2018, Belay et al., 2022). The topic therefore spans text normalization, written-to-spoken verbalization, orthographic canonicalization, pronunciation-aware decoding, and contextual disambiguation.

1. Task scope and representational targets

The literature represented here includes several distinct problem formulations. In social-media and adversarial-noise settings, normalization is lexical and orthographic: a noisy token such as an abbreviation, phonetic spelling, or emphasized form is mapped to standard English (Doshi et al., 2020). In speech applications, text normalization is defined as converting written text into “how the text is to be spoken,” or into “their canonical spoken equivalents,” which makes the output explicitly pronunciation-oriented rather than merely orthographically standardized (Wu et al., 2016, Wong et al., 5 Nov 2025). In MT and ASR robustness work, the objective is often weaker: the model need not emit a corrected source string, but it must remain semantically correct when a source token has been replaced by a same-sounding or similar-sounding alternative (Liu et al., 2018, Qin et al., 2020, Zheng et al., 2020). In script normalization, especially for Amharic and other Ge'ez-script languages, the task is deterministic preprocessing: characters “that have the same sound in a writing script are mapped to one character” (Nigatu et al., 20 Jul 2025).

Setting	Representative papers	Normalized target
Noisy written text	(Doshi et al., 2020, Khan et al., 2020)	Standard lexical form or cluster
Written-to-spoken TN	(Wu et al., 2016, Wong et al., 5 Nov 2025, Oji et al., 2021)	Spoken rendering
Robust ASR/NMT under homophone noise	(Liu et al., 2018, Qin et al., 2020, Zheng et al., 2020, Chung et al., 2023)	Robust representation or decoding path
Script-level canonicalization	(Belay et al., 2022, Nigatu et al., 20 Jul 2025)	Canonical character sequence

This heterogeneity matters because the same term can refer either to explicit correction or to robustness. Some systems select a canonical verbalization from a candidate set, as in WFST-based TN with ranking (Wu et al., 2016). Others replace suspicious characters with syllables and defer semantic recovery to a downstream model, which is normalization only in an intermediate representational sense (Qin et al., 2020). Still others normalize only for evaluation, not for training, as a way to preserve orthographic variation while reducing metric mismatch (Nigatu et al., 20 Jul 2025). A plausible implication is that “homophone normalization” is best treated as a family of operations indexed by target representation and deployment point rather than as a single task.

2. Linguistic sources of homophone variation

Homophone normalization arises wherever pronunciation preserves enough information for humans, but surface form does not preserve enough information for a model. In Mandarin NMT and ASR, the underlying problem is dense many-to-one mapping from character sequences to a comparatively small syllabic inventory. One paper notes that Chinese Pinyin without tones contains about 400 syllables representing about 6,500 characters, and another uses 404 Pinyin syllables with tone discarded (Qin et al., 2020, Liu et al., 2018). This makes homophone substitutions common in ASR, IME input, and noisy text, while leaving lexical identity brittle for text-only encoders.

In informal English, the relevant variation is broader than strict lexical homophony. The noisy forms discussed include elongated spellings, vowel deletion, syllabic or phonetic substitutions such as “2morrow,” symbol substitutions such as “4” and “@”, abbreviations, contractions, and adversarial character-level perturbations (Doshi et al., 2020). Roman Urdu presents a related but script-conversion-specific case: a single Urdu lexical item may appear under many Roman spellings, with vowel omission, silent or omitted “h”, nasal omission, and alternative consonantal Romanizations, so the same pronunciation is distributed across a large orthographic family (Khan et al., 2020).

In morphologically silent systems, the ambiguity is not noisy spelling but underspecified acoustics. French ASR offers a clean example: spoken /livʁ/ may correspond to livre or livres, and third-person verb forms may be singular or plural in writing despite identical pronunciation. The disambiguating evidence is carried by determiners, pronouns, or subject noun phrases elsewhere in the utterance (Mohebbi et al., 2023). Khmer is reported to be rich in homophones “due to its complex structure and extensive character set,” so writers often struggle to choose the correct same-pronunciation word from context (Born et al., 2024).

In Ge'ez-script languages, the source of ambiguity is historical phonological change plus script persistence. Several characters may now share a sound in Amharic, which encourages preprocessing-time collapsing, but those same characters can remain distinct in Tigrinya or Ge'ez (Belay et al., 2022, Nigatu et al., 20 Jul 2025). Cantonese adds a further variant: the acoustic model may recover the syllable class while failing to recover the intended character because many characters are homophonous and rare words have little supervision (Chung et al., 2023). This suggests that homophone normalization is often a joint consequence of phonological merger, orthographic redundancy, and contextual underdetermination rather than of homophony alone.

3. Algorithmic paradigms

A first major paradigm is pronunciation-aware representation learning. In Mandarin-to-English NMT, a token $a$ with pronunciation sequence $\psi(a)$ is embedded by interpolating textual and phonetic information: $$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$ where $\beta$ controls the relative weight of phonetics and text (Liu et al., 2018). The reported best clean-set BLEU occurs at $\beta=0.95$, not at balanced weighting and not at pure phonetics, which indicates that phonetics should dominate but not erase orthographic information. A related ASR strategy uses homophone-based label smoothing rather than explicit correction: the target prior assigns 0.6 probability to the true character, 0.3 total mass to its homophones, and 0.1 to all remaining characters (Zheng et al., 2020). Both approaches encode the same design principle: same-sounding alternatives should not be treated as uniformly impossible.

A second paradigm is candidate generation plus contextual reranking. In minimally supervised written-to-spoken TN, a WFST grammar deliberately overgenerates possible spoken forms for each written token, and a maximum-entropy ranker selects among candidates using local output n-grams, boundary trigrams, written/spoken skip-grams, and a pass-through bias (Wu et al., 2016). In noisy English normalization, BERT proposes candidate replacements for a masked token and a reranker combines contextual rank with string and phonetic similarity. The relevant formulas are

$$SimScore(X,Y) = 0.65 \cdot PSim(X,Y) + 0.35 \cdot SSim(X,Y)$$

and

$FinalScore(X) = P(X) \cdot SimScore(X,Y)$

where $PSim$ uses Metaphone, Soundex, and Fuzzy Soundex, and $SSim$ uses normalized Levenshtein, Jaro-Winkler, and $n$-gram cosine similarity (Doshi et al., 2020). In Roman Urdu, UrduPhone phonetic encoding, string matching, and contextual similarity are fused inside Lex-Var clustering, so normalization proceeds via lexical-variation groups rather than direct one-step correction (Khan et al., 2020).

A third paradigm is selective phonetic backoff. In SANMT, a detector predicts whether an observed Chinese character is plausible given the syllable sequence $\psi(a)$0, using the score

$\psi(a)$1

If $\psi(a)$2 falls below a threshold, the system replaces the character with its Pinyin syllable and feeds the resulting mixed character/syllable sequence into a syllable-aware Transformer (Qin et al., 2020). This does not restore the original character; it removes misleading lexical identity and backs off to pronunciation. Cantonese homophone extension applies a related idea at decoding time: each beam-search candidate character is expanded with lexicon-derived homophones and rescored with an external LM, while unified writing merges orthographic variants in the training transcripts (Chung et al., 2023).

A fourth paradigm is deterministic canonicalization. Amharic MT work uses a “frequency-based homophone normalizer” that replaces each set of same-sounding or functionally equivalent characters with a single most frequently used representative before training (Belay et al., 2022). ParsiNorm performs analogous canonicalization for Persian speech processing, but its target is often spoken text: dates, times, phone numbers, abbreviations, URLs, symbols, and currencies are verbalized according to rule-based templates (Oji et al., 2021). PolyNorm replaces hand-built normalization grammars with prompt-based few-shot LLM inference, using an instruction prompt, in-context examples, and the target unnormalized input, with locale explicitly included in the prompt (Wong et al., 5 Nov 2025).

4. Contextual disambiguation as the central mechanism

A consistent finding across domains is that phonetic equivalence is insufficient for normalization unless context constrains the candidate set. In noisy English, “min” can mean “minutes” or “Minimum,” and “bc” can mean “because” or “British Columbia,” so the system masks the token and asks BERT to generate context-appropriate candidates before phonetic reranking (Doshi et al., 2020). In multilingual TTS normalization, “Dr” may correspond to “Doctor” or “Drive,” “12:30” may have multiple spoken realizations, and “17º/17°” can be read differently by language and context; the prompt therefore instructs the model to normalize “based on the context,” and locale is supplied explicitly (Wong et al., 5 Nov 2025).

French homophone disambiguation shows the same principle in a cleaner syntactic form. The spoken target often contains no acoustic evidence distinguishing singular from plural, so the model must use determiners, pronouns, or subject noun phrases as agreement cues. Controlled experiments show that encoder-only XLSR models incorporate these cues in middle layers, whereas Whisper encoder-decoder models rely mainly on decoder self-attention over the textual prefix rather than on strong speech-encoder cue representations (Mohebbi et al., 2023). This suggests that where contextual evidence resides in the architecture determines where homophone normalization can be attached: mid-layer encoder representations in CTC-style models, or decoder-side rescoring in encoder-decoder ASR.

The converse result appears in spoken word embeddings. A homophone-based inspection of Speech2Vec argues that if a model sees only an isolated spoken word, it lacks the contextual signal needed to separate homophones such as ate/eight or meat/meet. Reproduced speech-derived embeddings place homophones extremely close, with mean cosine similarity $\psi(a)$3, whereas the released “official” embeddings behave like text-derived vectors and place homophones farther apart than random pairs (Chen, 2022). This directly supports the broader claim that phonetic proximity is a candidate-generation signal, not a final decision rule.

Chinese cloaked-toxicity unveiling makes this architecture explicit. A homophone graph and toxic lexicon first generate all substring–candidate pairs satisfying characterwise graph adjacency, producing a high-recall but high-false-positive set. Filtering then compares contextual plausibility of the observed string $\psi(a)$4 and the candidate toxic word $\psi(a)$5 through

$\psi(a)$6

For autoregressive LLMs, the paper derives this difference from full-sequence probabilities: $\psi(a)$7 The method is iterative: it replaces the best-supported candidate first and then recomputes all remaining scores (Ma et al., 28 May 2025). This suggests a general recipe for homophone normalization with causal LMs: generate phonetic candidates first, then score whole-sentence substitutions rather than isolated word probabilities.

5. Evaluation regimes and empirical behavior

Evaluation varies sharply with task definition. Social-media normalization of English uses human-centered ratings rather than token-level gold labels, reporting 86.71% for OOV masking and 83.22% for word-by-word masking on 2,627 SMS messages (Doshi et al., 2020). Written-to-spoken TN typically uses WER, BLEU, SER, or CER; multilingual PolyNorm reports consistent WER reductions against a production-grade rule-based system across eight languages, with examples such as German 10.74$\psi(a)$8 to 4.17$\psi(a)$9, English 9.84$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$0 to 4.28$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$1, Italian 15.02$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$2 to 4.56$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$3, and Japanese 17.49$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$4 to 7.88$$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$5 (Wong et al., 5 Nov 2025). The benchmark, however, usually assumes one preferred spoken form, so ambiguity is often evaluated against a single conventional reference rather than a set of acceptable variants.

Robustness-oriented MT papers measure clean and noisy BLEU. In joint textual-phonetic NMT, a text-only Transformer obtains 45.97 BLEU on the NIST06 dev set, while the joint model reaches 48.91 at $$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$6; on homophone-corrupted NoisySet1 and NoisySet2, the same comparison is 41.33/37.11 versus 45.71/42.66 (Liu et al., 2018). SANMT reports 44.20 BLEU on clean LDC tests and 32.17 on artificial noisy tests for a standard baseline, compared with 45.41 and 44.58 for the detector-plus-syllable-aware system; on a real-noise ASR benchmark, the same LDC-trained systems score 18.85 and 25.36 respectively (Qin et al., 2020). In ASR, homophone-based label smoothing reduces CER from 7.9% to 7.5% in a hybrid CTC/attention Chinese recognizer (Zheng et al., 2020).

Script-level normalization can yield large gains, but the direction is not uniform. In Amharic–English MT, frequency-based homophone normalization improves the strongest reported M2M100 system from 34.12 to 37.79 BLEU for Amharic→English and from 29.65 to 32.74 for English→Amharic (Belay et al., 2022). Cantonese ASR shows a similar benefit from orthographic canonicalization: on an out-of-domain rare-word set, beam-search+LM gives 39.33% CER, homophone extension reduces this to 35.82%, unified writing to 33.78%, and the combination to 31.87% (Chung et al., 2023). Yet a later Ge'ez-script study finds that normalization gains are inconsistent, that broader HSL normalization is often harmful, and that post-inference normalization of predictions and references can still improve BLEU by up to 1.03 without collapsing orthographic variation during training (Nigatu et al., 20 Jul 2025).

These results indicate two distinct empirical patterns. First, pronunciation-aware modeling and context-sensitive reranking reliably improve robustness under homophone noise. Second, deterministic canonicalization can either help by reducing sparsity or hurt by erasing meaningful orthographic structure, depending on language, transfer setting, and evaluation protocol. This suggests that metric gains from normalization are not interchangeable with linguistic adequacy.

6. Limits, controversies, and open directions

A recurrent misconception is that homophone normalization is equivalent to phonetic matching. The literature instead shows that phonetics is usually a high-recall prior, while the decisive signal comes from context. PolyNorm does not define a dedicated homophone category; its ambiguity handling is embedded in sentence-level TN. SANMT and joint phonetic NMT improve downstream robustness without reconstructing the exact intended source form. Speech2Vec-style isolated spoken embeddings collapse homophones rather than resolve them. These results indicate that pronunciation alone is normally too coarse for explicit normalization (Wong et al., 5 Nov 2025, Qin et al., 2020, Chen, 2022).

A second controversy concerns standardization. In Amharic NLP, homophone normalization has often been treated as harmless preprocessing, but later work argues that it imposes an implicit orthographic standard, reduces exposure to alternative spellings, and can damage transfer to related Ge'ez-script languages where the same character mappings are invalid or meaning-changing (Nigatu et al., 20 Jul 2025). The proposed compromise is post-inference normalization for evaluation: $$\pi([a, \psi(a)]) = (1-\beta)*\pi(a) + \beta*\pi(\psi(a))$$7 while training remains on unnormalized data. This formulation is presented as a faithful restatement of the paper’s procedure rather than as a printed equation, and it reflects a broader shift from training-time canonicalization toward evaluation-time equivalence classes.

A third limitation is resource scarcity. Khmer work reports a clear need for context-sensitive homophone correction but no specialized tool; a survey of 108 native speakers found that many frequently encounter homophone problems, often doubt their word choice, and want automatic correction, contextual analysis, feedback, and integration with Microsoft Word or Google Docs (Born et al., 2024). Roman Urdu normalization requires language-specific phonetic encoding such as UrduPhone because off-the-shelf Soundex-like schemes either overmerge or miss major transliteration patterns (Khan et al., 2020). Persian speech normalization similarly depends on curated symbol, abbreviation, and semiotic-class resources rather than on a generic text normalizer (Oji et al., 2021).

Open directions are therefore comparatively clear. Several papers imply the need for benchmarks that isolate lexical homophones, heteronyms, and pronunciation-sensitive ambiguity rather than treating them as incidental cases inside broader TN or MT tasks (Wong et al., 5 Nov 2025, Mohebbi et al., 2023). Others suggest hybrid architectures in which phonetic or homophone-graph candidate generation is followed by contextual ranking, possibly with pronunciation lexicon constraints, LM rescoring, or post-inference normalization rather than irreversible training-time collapsing (Ma et al., 28 May 2025, Nigatu et al., 20 Jul 2025). A plausible implication is that future homophone normalization systems will be judged less by whether they erase variation and more by whether they preserve valid variation while resolving only those pronunciation-based ambiguities that context actually licenses.