Quantifying Consonant Contributions to Word Intelligibility via Acoustic Masking

• Consonant contribution could complement stimulability, ease of articulation, and frequency when prioritizing misarticulated sounds for limited therapy time.The FITI framework – captures contribution through factors including positional prominence, frequency, and functional load, but had been perceptually validated only in English and for a small consonant set [14].Source: §I · p. 1
• Perceptual validation of consonant contribution requires language-specific listener studies, making coverage across languages prohibitively costly.Contribution is shaped by language-specific phonological structure, and studies require native listeners and language-specific stimuli.Source: §I · p. 1
• ASR offers a scalable alternative because it can process large corpora across natural word contexts, and multilingual models extend measurement to many languages.ASR-derived word error rate has been reported to correlate with human intelligibility judgments for synthesized speech and clinician ratings for dysarthric speakers,,,.Source: §I · p. 1
• MMR measures a consonant’s contribution as the proportion of masked instances in which the word is misrecognized.The method masks one consonant at a time in an isolated word and tests whether ASR still recognizes it.Source: §I · pp. 1–2
• Across English, Spanish, German, and Czech, consonant rankings agree only partially, indicating that contribution is language-dependent.The paper also relates MMR to phoneme frequency and functional load.Source: §I · p. 2

• The FITI framework – derives consonant importance from linguistic factors, but its validation covered only two English consonants, /ô/ and /T/ [14].Source: §II · p. 2
• Masking probes segment importance by testing whether recognition survives signal removal; this study uses isolated words to limit sentence-context influence.Prior ASR masking work added noise to phonemes in full sentences, where degradation can reflect the recognizer’s use of sentence context.Source: §II · p. 2

• Silence masking removes all acoustic information in a consonant interval, unlike real mispronunciations that may retain partial cues, so MMR is an upper bound on contribution.Source: §III · p. 2
• Forced alignment identifies each phoneme’s interval and the boundaries of its enclosing word before masking.Source: §III · p. 2
• Only correctly transcribed isolated words are retained, so later misrecognitions can be attributed to masking rather than baseline errors.Homophones count as correct transcriptions.Source: §III · p. 2
• Each consonant is silenced one at a time, and the fraction of masked instances that no longer match the reference word is its MMR.Scores are computed separately for each language and ASR model; homophones count as correct.Source: §III · p. 3

A. Datasets

• Five corpora span English, Spanish, German, and Czech, including the latter three as OneVoice-MSD 2026 focused-track languages.Source: A. Datasets · p. 3
• Common Voice version 22.0 supplies test data for masking and train and development data for estimating linguistic factors.The corpus is crowdsourced read speech with sentence-level transcripts and broad phonemic coverage over commonly used vocabulary.Source: A. Datasets · p. 3
• TIMIT provides an independent American English corpus with hand-verified, time-aligned phone boundaries for comparison with Common Voice-derived linguistic factors.Source: A. Datasets · p. 3

1) Forced alignment:

• TIMIT uses manual phone segmentation, while Common Voice is aligned with MFA; out-of-vocabulary words are transcribed with Epitran before alignment.All audio is processed at 16 kHz.Source: 1) Forced alignment · p. 3

2) Silence masking:

• Each consonant’s masking window is extended by 25% of each adjacent phoneme’s duration to reduce acoustic leakage from coarticulation.At word boundaries, the window is extended on one side only and clipped to the word boundary.Source: 2) Silence masking · p. 3
• The experiments use MMS (encoder-only), Whisper (encoder-decoder), and Qwen3-ASR (LLM-based), all supporting the four languages.The five corpora span English, Spanish, German, and Czech; train+dev tokens estimate linguistic factors, while baseline-correct tokens are masking trials.Source: 3) ASR transcription · p. 3, Table I · p. 3
• In all corpora except Czech, retained trial counts follow MMS < Whisper < Qwen; Qwen drops below Whisper for Czech.The authors note that this is consistent with Qwen3-ASR’s reported weaker performance on lower-resource languages than Whisper [34].Source: 3) ASR transcription · p. 3

1) Lexicon preparation:

• The lexicon is built from lowercased Common Voice train+dev transcripts, converted to phonemes, and restricted to words occurring at least 10 per million tokens.Transcripts are tokenized by whitespace and punctuation, then converted using Epitran and segmented with PanPhon.Source: 1) Lexicon preparation · p. 3

C. Linguistic factors for phoneme contributions

• Across all four languages, phoneme frequency and functional load are positively correlated.The relationship is shown on log–log axes.Source: Figure 2 · p. 4

1) Lexicon preparation:

• Phoneme frequency is the log10 of token-weighted phoneme occurrences summed across the lexicon.For each word, its token count weights the number of times the phoneme occurs in that word’s phoneme sequence.Source: 2) Phoneme Frequency (freq) · p. 4
• Functional load is the log10 count of minimal pairs in which a phoneme is the differing sound, including substitutions and single insertions or deletions.The study adopts the unweighted minimal-pair measure following FITI.Source: 3) Functional Load (FL) · p. 4

D. Word filtering

• Function words are excluded because their recognition depends less on segment acoustics, given that grammatical context often makes them recoverable.Language-specific stopword lists approximate the content-word set used for both masking and linguistic-factor calculations.Source: D. Word filtering · p. 4

E. Evaluation

• Surface phone labels are normalized to the canonical phoneme inventory used for frequency and functional-load estimates, with unmappable phones excluded.The mapping collapses surface variants such as aspirated stops, place-assimilated nasals, and final-devoiced obstruents; exclusions affect less than 10% of tokens per dataset.Source: E. Evaluation · p. 4
• Partial Spearman correlations isolate the association between MMR and one factor while controlling for the other, whose correlation with that factor could otherwise confound the result.The method applies Spearman correlations to ranked variables.Source: E. Evaluation · p. 4

V. RESULTS

• Marginally, MMR and frequency are negatively correlated, while MMR and functional load show no consistent pattern; functional load and frequency themselves correlate strongly (ρ = +0.80 to +0.95).Their strong intercorrelation can make the marginal associations misleading.Source: §V · p. 4, Table II · p. 5
• Controlling for frequency, the MMR–functional-load partial correlation is positive in every condition and significant in most; controlling for functional load, the MMR–frequency correlation becomes more strongly negative.The Czech functional-load correlation is positive but not significant; Figure 3 displays the MMR–functional-load partial correlations, using Qwen3-ASR for TIMIT-en and Whisper for the other cases.Source: §V · p. 4, Figure 3 · p. 5
• The nonsignificant Czech functional-load estimate is attributed to noisier estimates from a smaller train+dev set and lower ASR performance on Czech.The smaller set affects frequency and functional-load estimates, while lower ASR performance affects MMR.Source: §V · pp. 4–5
• The partial-correlation trends are similar across ASR models, including MMS, suggesting the pattern does not primarily stem from language modeling.Figure 3 uses the model whose MMR correlates most strongly with functional load for each language.Source: §V · p. 5, Figure 3 · p. 5

A. Cross-language analysis on consonant contribution

• The cross-corpus comparison covered the 12 consonants shared by all datasets.Source: A. Cross-language analysis on… · p. 5
• MMR rankings agreed only partially across languages; Spanish correlated significantly with every corpus (ρ = 0.63–0.76), while the English corpora agreed most strongly (ρ = 0.80).The pairwise correlations compare rankings for shared consonants.Source: Figure 4 · p. 5, A. Cross-language analysis on… · p. 5
• German and Czech rankings were not significantly correlated (ρ = 0.50, p = .10), nor were either English corpus and German or Czech.Source: A. Cross-language analysis on… · p. 5

B. Positional asymmetry in consonant contribution

• In English and German, masking word-initial consonants is generally more disruptive, but Spanish and Czech show no clear positional trend.Source: B. Positional asymmetry in consonant… · p. 5, Figure 5 · p. 6
• The onset advantage is not uniform: some consonants are similarly or more disruptive in final position, and further within-language datasets are needed to confirm the patterns.Source: B. Positional asymmetry in consonant… · p. 6

A. Integration with FITI

• Higher functional load predicts greater masking disruption, while more frequent consonants are less disruptive; these patterns were similar across three ASR architectures.The authors interpret frequency’s negative association as reflecting contextual predictability, which can leave a word recoverable from unmasked segments.Source: A. Integration with FITI · p. 6
• Functional load reflects per-instance lexical necessity, whereas frequency reflects how broadly phonemes recur across the vocabulary.Source: A. Integration with FITI · p. 6
• Treatment priorities can follow the objective: frequency favors producing more words intelligibly, while functional load favors recognizability of high-impact consonants.Source: A. Integration with FITI · p. 6

B. Clinical considerations

• A speaker’s stimulability, motor range, and personally important vocabulary can override language-level MMR rankings.Corpus-level frequency and contrastiveness may not reflect names, place names, or routine-specific words.Source: B. Clinical considerations · p. 6

VIII. CONCLUSION

• Acoustic masking paired with ASR produced empirical consonant-contribution rankings across English, Spanish, German, and Czech, which can serve as language-level priors for treatment targets.Source: §VIII · p. 6
• Silence is not equivalent to dysarthric mispronunciation, one-at-a-time masking omits segment interactions, and listener validation is still needed.Source: §VIII · p. 6