A Situational Speech Synthesizer for Yoruba: System Design, Phonological Rule Architecture, and Orthographic Extensions for Contour

Adedayo Oluokun; Dadepo Aderemi; Hafiz Adewuyi; Kola Tubosun

REVIEW 3 major objections 6 minor 2 references

A Situational Speech Synthesizer for Yoruba: System Design, Phonological Rule Architecture, and Orthographic Extensions for Contour

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A rule-based synthesizer with 651 diphones generates correct Yoruba pronunciation for tone-marked names, and a new single-vowel caron/circumflex notation for contour tones is perceptually equivalent to the standard geminated notation.

desk verdict Useful engineering documentation for the first deployed Yoruba TTS, but the perceptual-equivalence evaluation is vacuous: the two notations are normalized to the same audio before synthesis, so the null result says nothing about orthographic adoption. read the letter →

arxiv 2607.18317 v2 pith:CWDOY6Y3 submitted 2026-07-17 cs.SD cs.CLeess.AS

Kola Tubosun , Adedayo Oluokun , Hafiz Adewuyi , Dadepo Aderemi This is my paper · ORCID

classification cs.SDcs.CLeess.AS

keywords Yorubatext-to-speechdiphonesynthesistonecontourtonescaroncircumflexlow-resourceNLP

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper presents TTSYoruba, a deployed text-to-speech system that turns tone-marked Yoruba text into audio by applying hand-written phonological rules to a small recorded inventory of diphone units. Its central claim is that this rule-based approach is sufficient for the domain of personal names, where input already encodes tone. The paper also argues that contour tones (rising and falling) can be written on a single vowel using caron and circumflex marks, and that these marks produce audio indistinguishable from the standard double-vowel spelling. If true, the system offers a template for low-resource tonal languages and gives Yoruba users a keyboard-friendly way to represent contour tones without changing established names. A fifty-listener study reports high intelligibility and moderate naturalness across four name categories.

What carries the argument

The load-bearing mechanism is the five-way tonal file-selection rule, encoded in file suffixes _l, _m, _h, _f, _r. The suffix for a syllable is determined by its graphemic tone and the tone assigned to the immediately preceding syllable: a low after high/rising takes _f (falling), a high after low/falling takes _r (rising), and level tones take their default file. This single rule, applied over a 651-unit diphone inventory, generates the contextual contours of Yoruba sandhi. The nasal disambiguation rules (N1–N4) resolve the three-way ambiguity of the letter 'n' by tone marks, vowel class, and position. The caron/circumflex notation feeds into the same rule system via a normalization step th

What would settle it

A preregistered same-different or ABX listening test in which native listeners hear matched geminated and caron/circumflex forms of the same names; if listeners can reliably tell the two notations apart beyond chance, or if a TOST equivalence test with a pre-specified margin finds the confidence interval outside the margin, the equivalence claim is refuted.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes a complete, explicit phonological rule architecture for Yoruba concatenative synthesis: every consonant-vowel pair is recorded in five tonal variants (low, mid, high, falling, rising), and a file-selection rule picks the variant based on the syllable's graphemic tone and the tone of the preceding syllable. Contextual rising and falling contours are derived from level-tone input, and a three-way rule set disambiguates oral /n/, nasalized vowels, and syllabic nasals. The orthographic contribution extends the pipeline by normalizing caron and circumflex vowels to their geminated equivalents before synthesis, so the same audio is produced whether a user ty

Load-bearing premise

The claim that caron/circumflex notation is perceptually equivalent to geminated notation rests on the absence of a significant difference in a study that was not designed with equivalence margins or a power analysis, so a real but small difference could go undetected.

Editorial extensions

If this is right

Any correctly tone-marked Yoruba personal name can be synthesized automatically, which is how the system has generated audio for nearly 10,500 dictionary entries without per-name recording.
The caron/circumflex notation gives a Unicode, keyboard-compatible way to write contour tones on single vowels; users get identical audio whether they type the geminated or single-vowel form.
The documented rule architecture can serve as a reproducible specification and as a linguistic front-end for a future neural acoustic model.
The reported MOS scores provide a first published baseline for naturalness and intelligibility of Yoruba text-to-speech.
Because the acoustic and linguistic layers are separated, swapping the recorded voice (e.g., to a female speaker) requires no changes to the rule system.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: the perceptual-equivalence claim is supported only by non-significant t-tests without equivalence margins or power analysis; a larger study designed to detect a small but real preference could overturn it.
Editorial inference: the caron/circumflex convention is not limited to TTS; it could be adopted by keyboards, text editors, and pedagogical materials wherever Yoruba contour tones must be written compactly.
Editorial inference: the acknowledged morpheme-boundary nasal errors (e.g., words where a nasalized vowel ends one morpheme before a vowel-initial morpheme) suggest that full coverage of the name corpus will require a morphological parser or a curated exception list rather than more spelling rules.
Editorial inference: the same rule-plus-small-diphone architecture could be ported to other tone-marking languages, though for languages that do not mark tone orthographically an automatic tone-restoration step would be needed first.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

What should you know upfront: this is a genuine engineering contribution—the first publicly deployed Yoruba TTS with a fully documented rule architecture, a 651-unit diphone corpus, and explicit handling of nasal disambiguation and contextual contours. If you work on low-resource TTS or Yoruba NLP, the file-selection rules alone are worth having on record. But the headline orthographic claim, the caron/circumflex notation, is not supported by the evaluation as written. Section 5.4 says caron and circumflex vowels are expanded to their geminated equivalents before grapheme-to-phoneme conversion. That means the Category C and D stimuli in the listener study are, in effect, the same audio files. The paired t-tests in Section 6.4 (p=.307, p=.101) are comparing identical waveforms under two labels. A null result in that comparison is a design consequence, not evidence of perceptual equivalence. Calling the two notations "perceptually equivalent in synthesis" is trivially true, but it does nothing for the Section 5.2 adoption argument. The only direct evidence of a spelling effect is one free-text comment ("off" vs. "this is it"). That's not zero evidence, but it's thin. The equivalence claim needs to either be dropped from the paper or reframed as a statement about input normalization, with the orthographic argument left to the companion paper.

What the paper does well: the phonological rule system is explicit and replicable; the nasal disambiguation rules are concrete enough to implement; the known limitations section is honest about morpheme-boundary failures, upstep, and incomplete circumflex support. The MOS numbers (naturalness 3.90–4.15, intelligibility 4.48–4.60) are reasonable for concatenative synthesis and establish a baseline. The missing Ọrẹ (2014) reference is a minor citation slip. The 100% attention-check pass rate is suspicious but not damning; the check may simply have been easy.

Who's this for? Anyone building rule-based or hybrid TTS for tone-marked languages, and anyone working on Yoruba orthography. It deserves a serious referee—not because the orthographic claim is solid, but because the system documentation and deployment are substantive and the research questions are worth airing. Send it to peer review, but ask the authors to revise Section 6.4: either report the normalization equivalence explicitly and remove the claim of perceptual validation, or run a real experiment that varies written notation while holding audio fixed. As it stands, the paper's central evidence for its own orthographic contribution is circular.

Referee Report

3 major / 6 minor

Summary. This paper presents TTSYoruba, a rule-based concatenative diphone synthesizer for Yoruba deployed on YorubaName.com. The system converts tone-marked text to audio via Unicode normalization, syllabification and nasal disambiguation, a five-way tonal file-selection rule system, and concatenation of 651 diphone units. The authors also propose an orthographic extension: using caron (ˇ) and circumflex (ˆ) on single vowels to represent rising and falling contour tones, with normalization into geminated forms in the TTS pipeline. Evaluation by 50 listeners produces MOS scores of 4.08 naturalness and 4.55 intelligibility, and the paper claims perceptual equivalence between geminated and caron/circumflex notation on the basis of non-significant paired t-tests.

Significance. The engineering contribution is substantial and credible: the system is deployed, the rule architecture is specified in unusual detail (Tables 1, 3, 4), the 651-unit inventory and file-naming scheme are concrete and auditable, and the authors are transparent about known limitations. The MOS results demonstrate a functional, highly intelligible system for Yoruba names. The main problem is the orthographic equivalence claim. Because §5.4 expands caron/circumflex forms to their geminated equivalents before synthesis, the Category C and D stimuli are acoustically identical, so the reported null t-tests cannot validate perceptual equivalence. The evaluation therefore does not support the paper's second central contribution as currently framed.

major comments (3)

[§5.4, §6.4] Caron/circumflex normalization makes the C/D comparison tautological. Section 5.4 states that caron-marked and circumflex-marked vowels are 'expanded to their geminated equivalents' before grapheme-to-phoneme conversion, with Arómọlárǎn normalizing to Arómọláràán and selecting roon_r.wav. Consequently, a Category D stimulus and its matched Category C stimulus are the same synthesized audio. The paired t-tests in Section 6.4 therefore compare identical audio under two notation labels; the null result is a design consequence, not evidence for the perceptual equivalence of the notation. To support the Section 5.2 claim, the experiment should vary the written notation while holding audio fixed (or, if the written form was displayed, use an orthographic-acceptability measure) and should include a pre-registered equivalence margin.
[§6.4] Non-significance is not equivalence. Even setting aside the normalization issue, p=.307 (naturalness) and p=.101 (intelligibility) with d=0.10/0.17 do not establish that the notations are equivalent; equivalence testing (e.g., TOST with a specified bound) and a power/sensitivity analysis are required. Additionally, the paired t-test treats the two C/D pairs per participant as independent, yielding 100 observations from 50 participants; this ignores participant-level clustering. A mixed-effects model or participant-level averaging should be used.
[§5.2, §5.5, §6.4] The orthographic adoption claim is not self-contained. Section 5.2 and Section 5.5 defer the full equivalence table and the formal orthographic argument to Túbọ̀sún (forthcoming). The only direct evidence of a spelling-preference effect in the current manuscript is a single free-text comment about Akẹ́kọ̀ọ́ vs. Akẹ́kọ̌ (Section 6.4). The MOS naturalness and intelligibility scales are not a validated instrument for orthographic acceptability. Either include the formal argument and equivalence data, or clearly scope this paper as the computational implementation and leave the adoption claim to the companion paper.

minor comments (6)

[§2.2] The text cites Ọrẹ̀ (2014), but I could not find this entry in the reference list. Please add it or correct the citation.
[§5.2, §7.4] §7.4 says circumflex on open vowels ệ/ộ is only partially implemented, but §5.2 includes ộ ≡ ọ́ ọ̀ in the equivalence list. Clarify the current support status for open-vowel circumflex forms.
[§6.4] The 100% attention-check pass rate is not a strong data-quality guarantee; the threshold (rating a single item 1 or 2 on both dimensions) is lenient. A more sensitive check or catch trials would strengthen the data-quality argument.
[§6.4] The text states that the Category B pattern is 'discussed below', but no such discussion appears in §7. Either add the discussion or reword the sentence.
[§6.2] The listener-study description never explicitly states whether the written form was displayed alongside the audio. Given the text-sensitive comments in §6.4, this should be stated explicitly in §6.2.
[Title] The term 'Situational' in the title is not defined or used anywhere in the paper. Either define it or remove it.

Circularity Check

2 steps flagged · score 6.0 of 10

C/D perceptual-equivalence test is tautological: caron forms are normalized to geminated strings before synthesis, so the paired t-tests compare identical audio by construction.

self definitional [Section 5.4 / Section 6.4]
"A preprocessing normalization step was added to the TTSYoruba pipeline: before grapheme-to-phoneme conversion, caron-marked and circumflex-marked vowels are expanded to their geminated equivalents, after which processing continues normally."

Because every caron/circumflex form is expanded to its geminated equivalent before G2P, each Category D stimulus is synthesized from the same normalized string as its matched Category C stimulus, producing identical audio. The paired t-tests in §6.4 (naturalness t(99)=−1.03, p=.307; intelligibility t(99)=−1.65, p=.101) therefore compare the same audio condition under two labels; the null result is forced by construction and cannot independently establish 'perceptually equivalent in synthesis' or support the §5.2 adoption claim. The equivalence is an input to the pipeline, not an experimental output.
self citation load bearing [Section 5.2 / Section 5.5]
"The full equivalence table across all seven Yorùbá vowels, and the formal orthographic argument for the notation's adoption, are presented in Túbọ̀sún (forthcoming); the present section documents only the computational implementation within the TTSYoruba pipeline."

The paper's orthographic contribution—adopting caron/circumflex as standard single-vowel contour tone markers—rests on a formal equivalence argument that is not contained in this paper and is cited to a forthcoming work by the same first author. The present text supplies only representative equivalences and an implementation detail, so the load-bearing justification for the notation's adoption reduces to a self-citation that is not independently verifiable from the manuscript.

full rationale

The core TTS architecture is a rule-based engineering system: the tone-selection and nasal-disambiguation rules are derived from phonological literature (Pulleyblank, Ajiboye) and corpus analysis, and the MOS study benchmarks the system's intelligibility and naturalness externally. That portion is self-contained and not circular. The circularity is concentrated in the orthographic-equivalence claim. Section 5.4 normalizes caron/circumflex forms to their geminated equivalents before grapheme-to-phoneme conversion, so Categories C and D are the same audio by construction; Section 6.4's paired t-tests therefore compare a condition with itself, and the null result cannot provide independent evidence for 'perceptual equivalence' or for the Section 5.2 adoption claim. Additionally, the formal orthographic argument for the notation is deferred to Túbọ̀sún (forthcoming), a self-citation that is load-bearing for the orthographic contribution. The self-citation is disclosed, but it is not independently checked within this paper. Overall, one central evaluation reduces by construction and the orthographic proposal relies on an unverified self-citation, while the main synthesizer design remains independent; hence a partial-circularity score of 6.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No numerical free parameters are fitted; the design choices are categorical rules derived from linguistic literature and corpus analysis. The caron/circumflex are pre-existing Unicode symbols repurposed as orthographic markers, not new physical entities. The main axioms are domain assumptions about Yoruba phonology and the completeness of the recorded inventory.

assumptions (4)

domain assumption High tones are the most stable and Mid tones do not participate in lexical contour formation in Yoruba.
Used in Section 2.1 and Table 1 to justify the five-file inventory with no mid-tone contour variants (_mf, _mr).
domain assumption The three-way nasal classification (oral onset, nasalization marker, syllabic nasal) is correct and covers the name corpus.
Section 4.3 Rules N1-N4; acknowledged to fail for morpheme-boundary cases like Ìtànìfẹ́, so not fully validated.
domain assumption Contextual contour assignment rules (high-after-low => rising; low-after-high => falling) correctly model tonal sandhi for personal names.
Section 4.2 and Table 3; derived from Pulleyblank (2004) but not experimentally validated for all name environments.
domain assumption A diphone inventory of 651 units recorded by one male speaker is sufficient to synthesize any correctly tone-marked Yoruba name.
Section 3; assumes the phonological inventory and its coverage are complete; no coverage evaluation is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Situational Speech Synthesizer for Yoruba: System Design, Phonological Rule Architecture, and Orthographic Extensions for Contour." pith.science (2026). https://pith.science/paper/CWDOY6Y3

@misc{pith2026260718317,
  author       = {Pith},
  title        = {Pith review of: A Situational Speech Synthesizer for Yoruba: System Design, Phonological Rule Architecture, and Orthographic Extensions for Contour},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWDOY6Y3}},
  note         = {Machine review of arXiv:2607.18317}
}

read the original abstract

We present TTSYoruba, a rule-based concatenative diphone speech synthesizer for Yoruba, deployed at online as part of the YorubaName.com open dictionary of Yoruba personal names. The system takes tone-marked Yoruba text as input and produces audio output by applying a hand-crafted phonological rule system to a recorded inventory of 651 diphone units spanning five tonal variants of every consonant-vowel combination in the language. We describe the phonological architecture of the system in detail, including our complete tonal file-selection logic, our treatment of the three-way nasal disambiguation problem (oral /n/, nasalized vowel, and syllabic nasal), and the derivation of contextual rising and falling tones from level-tone input. We also present, as an orthographic contribution, the adoption of the caron and circumflex, which are symbols with prior standing in Yoruba phonological transcription, as standard single-vowel contour tone markers, integrated into the TTS normalization pipeline and the WriteYoruba keyboard input tool. The system's performance was evaluated through a listener study (N=50), with detailed results on Mean Opinion Scores (MOS) presented in Section 6. Keywords: Yoruba, text-to-speech, low-resource languages, diphone synthesis, contour tones, African language NLP, rule-based synthesis

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

2 extracted references · 1 linked inside Pith

[1]

20 Ajíbóyè, O. (2020). Functions and distribution of n in Yorùbá orthography. Ihafa: A Journal of African Studies , 11 (1), 58–68. Akinlabí, A. (2004). The sound system of Yorùbá. In N. Lawal, M. N. O. Sadiku, & A. Dopamu (Eds.), Understanding Yoruba Life and Culture (pp. 453–468). Africa World Press. Awóbùlúyì, Q. (1994). The development of Standard Yoru...

2020
[2024]

9296–9303)

(pp. 9296–9303). ELRA and ICCL. https://arxiv.org/abs/2307.16071 Ọlátúbọ̀ sún, K. (2012). Studies of initial tonal acquisition by American English speakers learning Yoruba [Master's thesis, Southern Illinois University Edwardsville]. Zenodo. https://doi.org/10.5281/zenodo.20833023 Olmsted, D. L. (1951). The phonemes of Yoruba. WORD , 7 (3), 245–249. https...

arXiv 2012

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

20 Ajíbóyè, O. (2020). Functions and distribution of n in Yorùbá orthography. Ihafa: A Journal of African Studies , 11 (1), 58–68. Akinlabí, A. (2004). The sound system of Yorùbá. In N. Lawal, M. N. O. Sadiku, & A. Dopamu (Eds.), Understanding Yoruba Life and Culture (pp. 453–468). Africa World Press. Awóbùlúyì, Q. (1994). The development of Standard Yoru...

2020

[2] [2024]

9296–9303)

(pp. 9296–9303). ELRA and ICCL. https://arxiv.org/abs/2307.16071 Ọlátúbọ̀ sún, K. (2012). Studies of initial tonal acquisition by American English speakers learning Yoruba [Master's thesis, Southern Illinois University Edwardsville]. Zenodo. https://doi.org/10.5281/zenodo.20833023 Olmsted, D. L. (1951). The phonemes of Yoruba. WORD , 7 (3), 245–249. https...

arXiv 2012