Assessing the Reusability of Public Speech Resources for Low-Resource Languages: A Central Kurdish Case Study

1. Introduction and background information

• The release provides 35 hours and 20 minutes of transcribed speech from three speakers, three trained TTS models, and evaluation code; the paper reports MOS scores between 4.01 and 4.08.The review examines documentation, evaluation design, reproducibility, licensing, and what the release records about the language; it does not independently assess audio quality.Source: §1 · p. 1
• Central Kurdish remains poorly represented in public speech technology, and the release adds resources amid a gap in publicly documented and evaluated systems and datasets.Earlier work included a 21-hour Tacotron 2 system without released models or data, 13 hours of studio recordings, and Meta’s multilingual MMS models; the authors note limitations in evaluation, public datasets, and voice-cloning capability.Source: §1 · p. 1
• The review checks the full paper, released model and dataset files, and upstream components, comparing accounts where the paper and model card disagree.It also corrects earlier assumptions when the paper contradicts them.Source: §1 · p. 2

2. What the paper establishes that the repository does not

• The paper says the models fine-tuned the English checkpoint of F5-TTS-base, but the model card says only that they are “based on the F5-TTS architecture.”The upstream repository documents the checkpoint as trained on Chinese and English, and the paper’s F5TTS_Base link differs from the repository’s F5TTS_v1_Base.yaml configuration; which version was used is unstated.Source: §2 · p. 2
• The paper explains that the non-Kurdish vocabulary in vocab.txt led the pipeline to convert Kurdish text to phonemes for compatibility with F5-TTS.The paper says Central Kurdish is almost phonemic and has high-precision G2P tools; Mahmudi and Veisi (2021) report contested orthography, including short /i/ and ی and و, while the converter’s choices are baked into training data and raw text can fail.Source: §2 · p. 2
• The no-derivatives license came from audiobook rights holders, while non-commercial terms are inherited from F5-TTS weights; together, they constrain public redistribution of verification and corrections.The paper says audiobook writers and narrators granted CC BY-NC-ND 4.0 permission; the restriction can allow private verification while preventing publication of adapted models, and outside users cannot redistribute even a two-line number-bug fix.Source: §2 · pp. 2–3
• The review recommends separating licenses by artifact and stating their provenance, so users know which permissions apply and which rights holders could relax them.Three model-card sentences drawn from the paper could explain the pretrained checkpoint, phoneme conversion, and license origins, reducing the risk that users misread the achievement, misuse the text input, or approach the wrong party.Source: §2 · p. 3

• 38,400 frames/GPU, an eight-GPU comment, and eleven epochs in F5TTS_v1_Base.yaml conflict with the paper’s one-GPU run and roughly 275,000 plotted steps.The paper also discusses convergence during the first 100k iterations; on eleven hours of audio, eleven epochs cannot approach the plotted step count.Source: §3 · pp. 3–4
• An upstream diff suggests the hardware comment and epoch count are inherited defaults: only the dataset name and logging block were changed.Source: §3 · p. 4
• Misleading settings can deter community linguists and organizations from reproducing the work or extending Central Kurdish systems to other varieties.Badini extension also involves a variety change and possibly an unsupported script; Hawrami lies outside the Central Kurdish grouping.Source: §3 · p. 4
• The published repository names redirect to renamed repositories, so the artifacts remain traceable despite the name mismatch.The paper’s names were aranemini/tts4all-ckb-dataset and aranemini/ckb-f5-tts; the current names are central-kurdish-tts4all and central-kurdish-tts.Source: §3 · p. 4
• Correcting two configuration lines or labeling inherited defaults would address the mismatch; existing redirects already preserve citation traceability.Source: §3 · p. 4

4. What the evaluation supports, and what the model card claims

• The study’s 88 listeners gave 3,101 ratings across seven categories, rating 40 sentences each on a 1-to-5 human-likeness scale.Each system was assessed on 140 sentences, and listeners were not told that real speech was included.Source: §4 · p. 5
• The model card calls the female audiobook voice suitable for general-purpose use, but the paper says three speakers are insufficient for a Central Kurdish multi-speaker model.The release has two male voices and one female voice, so applications built from it inherit that limited choice of speakers.Source: §4 · p. 5
• 4.08 against 4.06 underlies the model card’s “highest” ranking, but the paper calls the female audiobook system only “marginally better” and claims no significant difference.The paper’s conclusion is that the systems perform similarly, not that one beat another.Source: §4 · pp. 4–5, Table 1 · p. 5
• Each system loses nearly the same amount from its source recordings—0.34, 0.33 and 0.34 MOS—and the TTS ranking follows the recordings’ MOS ranking.Because each condition has one speaker, audiobook source, narrator, recording environment and reading material cannot be disentangled; the paper notes this limitation.Source: §4 · p. 5
• The studio system leads on PESQ, scoring 4.16 on natural test and 4.00 in-domain against 3.74 and 3.04 for the female audiobook system.The card’s approximate figures otherwise largely reflect rounded means of the relevant Table 3 rows.Source: §4 · p. 6

5. A measurement that depends on its own training data

• The fine-tuned SeamlessM4T V2 Large system generated the training transcripts and then scored the synthetic speech, making the CER evaluation dependent on its own training data.CER indirectly measures intelligibility by comparing recognition output from synthetic audio with the input text; the transcripts were manually revised after automatic transcription.Source: §5 · p. 6
• The studio system’s natural test (0.09) and in-domain (0.08) CER values together more than account for its full seven-category-mean gap versus the female audiobook system.On categories furthest from any training domain, it scores 0.02 on semantically unpredictable sentences and 0.03 on expressions, versus 0.02 and 0.04 for the female audiobook system. The model card reports category means of 3.7%, 4.7% and 5.6%; the paper’s claim that CER is below 8% does not match Table 3’s 0.09 row.Source: §5 · p. 6
• Because the recognizer shaped the training text, CER may miss errors matching its expectations, including competing Kurdish spellings, and could favor Central Kurdish read speech when evaluating Badini or Hawrami.Without natural-speech control figures, researchers extending the reusable tts4all_eval pipeline cannot separate model performance from recognizer behavior.Source: §5 · p. 7
• Report natural-speech control rates or each system’s error relative to its own natural baseline, use an independently trained recognizer where possible, and disclose the shared recognizer when reporting CER.The proposed control comparisons are computable from data already in the paper.Source: §5 · p. 7

6. An undocumented preprocessing step

• The inference script replaces unbroken strings of nine or more Western Arabic digits with placeholders that it never restores, and “<” is missing from the vocabulary.The issue is established by inspecting the released code, but its effect on generated audio was not tested; numbers with spaces or separators pass through normally.Source: §6 · p. 7
• The preprocessing chain assumes Central Kurdish in Arabic script; a Latin-script Kurmanji input could silently mishandle ḧ and ẍ, while the demo’s character-count timing estimate can understate phoneme processing time.The study did not test the Kurmanji failure; the proposed fixes are to restore or remove the number placeholder, estimate time from phonemes, and validate vocabulary coverage before synthesis.Source: §6 · p. 8

7. What can and cannot be reproduced

• The released evaluation package reproduces its bundled four-utterance reference values across seven metrics, although the paper describes only three metric aspects.The package implements intelligibility, naturalness, signal quality, speaker similarity, and prosody; deterministic measures match exactly and GPU-computed measures agree to six or seven significant figures.Source: §7 · p. 8
• The evaluation code has a Kurdish intelligibility-model slot, but the required checkpoint is not distributed, so the intelligibility results cannot be recomputed from released materials.The paper gives no download location for its recognizer; a fine-tuning recipe using Wav2Vec2BertForCTC and the AsoSoft Central Kurdish corpus is published instead.Source: §7 · pp. 8–9
• Running the evaluation package also depends on separately installed or hosted components, while its bundled smoke test covers four French utterances and omits character error rate.Module-level imports require fairseq and utmosv2, and the quality-proxy checkpoint is fetched from another repository; the release includes model definitions and scripts rather than weights.Source: §7 · p. 9
• All 1,500 paper-designated test utterances are present but unmarked in the single training split, so fine-tuning on the released corpus can contaminate evaluation.The paper specifies 500 test samples for each dataset; the paper and dataset card total 17,671 utterances, while the Hugging Face viewer reports 17,738 rows.Source: §7 · pp. 9–10
• None of the seven listening-test sentence sets are published, so the subjective evaluation cannot be repeated by other researchers.The categories include loanwords, idioms, expressive sentences, nonsensical sentences, neutral news, in-domain test sentences, and genuine speech; the authors’ 140 evaluation sentences per system carry no third-party rights.Source: §7 · pp. 9–10
• Training requires undocumented work: users must convert the parquet corpus, infer audio-preparation choices, and guess which checkpoint was selected, with no TTS training script or exact command.The release does not distinguish deliberate settings from defaults, a particular obstacle for linguists and community researchers; documenting the invocation, preparation, and checkpoint choice would make the procedure followable.Source: §7 · pp. 9–10

8. The license and its practical reach

• CC BY-NC-ND 4.0 follows audiobook permissions, while F5-TTS pretrained weights add non-commercial terms; the F5-TTS code and AsoSoft library are MIT-licensed.The passage says the audiobook writers and narrators granted permission under CC BY-NC-ND 4.0; AsoSoft adds no restriction.Source: §8 · p. 10
• The release inherits non-commercial and no-derivatives terms, but its model and dataset cards omit their provenance, contrary to documentation frameworks’ licensing guidance (Mitchell et al., 2019).Hugging Face separates model and dataset repositories, each with its own license field, offering two places to record provenance.Source: §8 · p. 10
• A Kurdish developer may use the models in research, but fine-tuning, publishing bug fixes, adapting varieties, or converting weights may be legally uncertain because no court has defined model-weight derivatives.The study offers no legal opinion, and its argument does not depend on one.Source: §8 · p. 10
• The demo says it runs unmodified checkpoints, yet half-precision conversion and ported logic are arguably modifications under a literal license reading despite ordinary engineering usage.The documentation’s phrasing appears chosen with the license in view, leaving a good-faith developer in a grey zone.Source: §8 · pp. 10–11
• The dataset card lists TTS training and voice cloning despite unclear permission to distribute adapted models, so the researchers could seek a model-only redistribution exception or clarify permitted uses.The proposed exception would allow corrected or dialect-adapted models for non-commercial research while leaving the recordings under their present terms.Source: §8 · p. 11
• Documentation should identify which license terms come from audiobook permissions, pretrained weights, or the authors, and assign licenses separately to recordings, transcripts, weights, and code.The passage proposes these changes to preserve opportunities to build on the work even if raw recordings remain restricted.Source: §8 · p. 11

9. One rights holder named, one absent

• The female audiobook’s writer and narrator are credited, but the male audiobook’s writer or content rights holder is not, leaving its permission chain un verifiable from the paper.The male audiobook comprises 10 hours 51 minutes and 6,044 biography-domain utterances; the passage recommends naming its rights holder and permission scope in the dataset card. Narrator names also differ between the paper and dataset card: Shahin Shahlai/Fateh versus Shahen Shahlai/Fatih.Source: §9 · p. 11

The pronunciation decisions embedded in the models

• The G2P converter consistently resolves contested orthographic choices, including where to insert unwritten /i/, fixing one account of Kurdish pronunciation in the model weights.The cited work identifies the short vowel /i/ and two letters as unresolved cases in Kurdish orthography.Source: §10 · p. 11
• Audiobook narration and studio dubbing standardize the input register, while converter rules standardize pronunciation, yielding consistent speech that does not vary like a speaker’s.A fluent demo can therefore sound like Kurdish pronunciation generally, although it reflects one rule system and one register.Source: §10 · pp. 11–12
• With singleOutputPerWord=True, the converter keeps only the first candidate pronunciation, so accepted alternatives are absent from training and cannot be recovered through prompting.The source code confirms that word_G2P returns the first of several candidate pronunciations when the flag is true.Source: §10 · p. 12
• The normalizer changes every word-initial ر to ڕ, a recognized Central Kurdish convention, but allows no exceptions for loanwords or speakers who do not make that substitution.The released code enables changeInitialR=True, but the repository does not document the convention’s behavior.Source: §10 · p. 12
• The pipeline strips syllable-boundary markers before training even though the symbol appears in vocab.txt, and the study cannot determine whether retaining them would improve results.The paper identifies prosodic variability as its most challenging case, but does not establish a connection to marker removal.Source: §10 · p. 12
• Disclosing converter settings, discarded syllable information, and read-speech training would help readers calibrate the demo; the paper’s evaluation supports the models on exactly those terms.The release represents one narrator, one register, and one rule system.Source: §10 · p. 12
• For regional documentation or revitalization, preprocessing transcripts through this pipeline yields standard Central Kurdish reading rather than the recorded variety, so suitability depends on the project’s purpose.The pipeline normalizes toward the written standard before training and evaluation.Source: §10 · p. 12

The direction of accommodation

• The pipeline cumulatively removes variation, and phoneme conversion fits Kurdish to the F5-TTS baseline vocabulary rather than adapting that vocabulary to Kurdish.The stages are documented in the paper or released files; vocabulary expansion and continued pretraining were possible alternatives, but the project lacked the data and compute to pursue them.Source: §11 · p. 13
• All three sources are prepared read speech, only three speakers’ undocumented varieties are attested, and recognizer conventions shape transcripts before manual revision.The sources include audiobook narration and a professional dubber reading prepared sentences; audiobook editing removes substantial regional phonology, local vocabulary, code-switching, and spontaneous prosody.Source: §11 · p. 13
• Normalization changes every word-initial ر to ڕ, while G2P resolves contested spellings consistently and retains only the first candidate pronunciation.The normalizer permits no exceptions for loanwords or speakers who do not make the distinction; alternative accepted pronunciations leave no trace in the training data.Source: §11 · p. 13
• The 2,545-entry inventory contains no Kurdish-specific symbols; unrepresented Kurdish forms cannot be modeled, and vocabulary expansion with continued pretraining was not pursued.The paper attributes that choice to likely data and compute constraints, but the resulting representational space was not designed with Kurdish in mind.Source: §11 · p. 13
• Evaluation uses raters of unrecorded varieties and an intelligibility recognizer that produced the transcripts; CC BY-NC-ND bars outsiders from redistributing modified systems.The license therefore prevents others from publishing a system that reverses the pipeline’s decisions.Source: §11 · p. 13
• The release silently emits ḧ, ẍ, and <, none present in vocab.txt, and includes no step to report unrepresentable characters.The missing < also explains why the number placeholder described in Section 6 fails.Source: §11 · p. 13
• The unsupported claim that Central Kurdish is almost phonemic underpins the design, although the paper acknowledges contested orthographic cases where variation remains.The paper gives no citation for the claim, while its own cited dependency reports contestation in those cases.Source: §11 · p. 13

Parallel standards, not a standard and its deviations

• Kurdish’s parallel written standards mean a normalization rule selects among community positions, and the release does not identify the software default making that selection.Sorani, Kurmanji, and Kurdish in former Soviet republics use different scripts or conventions, with no institution authorized to rank them; planned Badini and Hawrami resources may inherit the same hidden assumptions.Source: Parallel standards, not a standard and… · p. 14

What makes a high-resource language high-resource

• Unlike English or German, Central Kurdish lacks comparable references for checking converter decisions, leaving this release’s automation effectively unexaminable.The review calls for linguistic scholarship to distinguish settled pipeline commitments from contested ones and help researchers assess transfer to other languages.Source: Parallel standards, not a standard and… · pp. 14–15
• Kurdish receives an additional grapheme-to-phoneme step for vocabulary compatibility, while English and Chinese reach F5-TTS without that irreversible transformation.The paper says F5-TTS ordinarily works directly from text, and the Kurdish converter’s underlying questions remain open.Source: Parallel standards, not a standard and… · p. 14

How a default becomes a fixture

• With no alternative, the model card’s “suitable for general-purpose applications” recommendation could make a female audiobook voice a default beyond the demonstrated scope.The no-derivatives license prevents developers from publicly sharing dialect adaptations, although the restriction originates with audiobook rights holders.Source: Parallel standards, not a standard and… · p. 15

What this criticism is not

• The authors probe variation, acknowledge three speakers are insufficient for a multi-speaker model, and use multilingual fine-tuning because Kurdish data cannot support training this quality from scratch.The review’s criticism is narrower: the released artifacts do not expose how the pipeline’s standardization and one narrator’s reading register shape the output.Source: Parallel standards, not a standard and… · p. 15

What would close this

• The review recommends documenting speaker varieties, rater dialects, normalization stages, converter settings, discarded syllable markers, and vocabulary coverage, and labeling the release as Central Kurdish read-speech synthesis.It also urges relevant linguistic references and asks rights holders whether dialect-adapted models can be redistributed for non-commercial research; these changes require no new data or training run.Source: Parallel standards, not a standard and… · pp. 15–16

12. Which variety, and whose judgement

• Although Kurdish spans a continuum of more than 35 million speakers (Central Kurdish: approximately 8 million), the release omits speaker varieties and rater backgrounds, obscuring interpretation of a 0.02 ranking margin.The paper reports 88 evaluators but not their linguistic backgrounds; the margin could reflect listeners’ variety as well as system differences.Source: §12 · pp. 16–17
• The paper makes consequential linguistic claims but grounds Kurdish linguistics sparsely, leaving non-specialists without a route to judge its assumptions or transfer conditions.Its 36 references include two linguistic works, Sheyholislami 2015 and 2021; Mahmudi and Veisi 2021 is an exception that engages orthographic questions, but is cited for its tool rather than its analysis.Source: §12 · pp. 16–17
• All three corpora use prepared read speech, which smooths away spontaneous regional phonology, local vocabulary, code-switching and prosodic patterns.The evaluation probes code-switching, idioms and prosodic variability; prosodic variability was most challenging, with MOS of 3.88, 4.01 and 3.95.Source: §12 · p. 17
• Transfer to Badini or Hawrami remains open because the demonstrated recipe depends on substantial prepared speech resources that may not exist for those varieties.The audiobook-data finding should not be treated as a general solution for varieties without audiobooks.Source: §12 · p. 17
• Because modified versions cannot be redistributed, researchers may be unable to publish corrections or adaptations, risking default reliance on the existing voices and conventions.The restriction comes from rights holders, but still limits what the next developer can share.Source: §12 · p. 17
• Recording speaker variety and region, reporting rater dialect distribution, and briefly situating the work in Kurdish linguistics would make its assumptions and transfer conditions assessable.The team already holds the variety and rater information; the paper could also cite work establishing settled Sorani phonology, contested orthography and conventional variety divisions.Source: §12 · p. 17

13. What the release does well

• 2.31 and 1.51 BLEU are the zero-shot translation baseline, versus 27.23 and 18.50 with synthetic training data; the FLEURS-audio upper bound is 92.42.Table 5 gives the translation figures; Table 4 gives the upper bound from actual FLEURS audio.Source: §13 · p. 17
• 95% confidence intervals cover every metric across seven categories and three systems, and the shipped evaluation configuration reproduces its reference values across all seven metrics.The intervals put the 0.02 margin in proportion; the published evaluation code runs.Source: §13 · pp. 17–18
• The 35-hour transcribed Central Kurdish corpus, with permission recorded from a named author and both narrators, can support recognition and phonetic study beyond the released models.The corpus’s limits are three speakers, read register, and undocumented varieties.Source: §13 · p. 18
• Each condition uses one speaker reading prepared material, so category-level confidence intervals describe sentence-type variation, not variation across speakers, varieties, or registers.The comparison concerns three specific systems on read speech and 88 listeners whose backgrounds are unrecorded.Source: §13 · p. 18
• The Giganet studio dataset had not, to the authors’ knowledge, been systematically evaluated or used for text-to-speech; the independent demo displays model input phonetics but accepts no uploaded reference audio.Displaying the phonetic string makes preprocessing inspectable, while declining reference-audio uploads prevents arbitrary-voice cloning.Source: §13 · p. 18

Conclusion

• The release does not consistently match the paper: its model card makes a firmer claim, its configuration lists unused hardware, and its dataset omits the evaluation split.The preprocessing code also includes a step absent from the paper that does not do what it was written to do.Source: §14 · p. 18
• The study demonstrates three single-speaker systems trained on read material, with unrecorded listener varieties and a ranking resting on two hundredths of a point—not a general method for low-resource synthesis.The release presents the work more broadly to researchers developing speech synthesis for other languages.Source: §14 · p. 18
• Incomplete documentation is a minor inconvenience to engineers but can block linguists working on Central Kurdish varieties that have little coverage and few alternatives.The varieties with least coverage depend on researchers least equipped to reconstruct missing documentation.Source: §14 · p. 18
• Most listed corrections are inexpensive and use information the team already holds; they would make results checkable without changing them.Examples include marking the test split, aligning the model card with the paper, correcting the configuration, publishing evaluation sentences, and recording speaker varieties and rater backgrounds.Source: §14 · p. 18
• Audiobook rights remain unresolved, including text ownership and permission to redistribute corrected or dialect-adapted models for non-commercial research; the model card also needs alignment with the paper’s wording.Permission from the rights holders, while recordings remain under existing terms, will determine whether future Kurdish voices can build on this work or must begin again.Source: §14 · pp. 18–19
• Unresolved training details include which hardware and checkpoint produced the release, the checkpoint-selection basis, and the missing TTS command and data-preparation documentation.The configuration specifies eight GPUs and eleven epochs, while the paper reports one GPU and Figure 1 extends to approximately 275,000 steps; the TTS command, manifest format, and any further audio preparation are unreported.Source: §14 · p. 19
• Evaluation remains difficult to reproduce or interpret because the test split and evaluation sentences are unidentified, scoring has a recognizer-related caveat, and recognizer access may be restricted.The scoring recognizer generated model-training transcripts and had higher error on genuine studio than audiobook speech (0.09 against 0.01 and 0.02, with a wide interval on the studio figure); the recognizer model listing returned HTTP 401 without credentials.Source: §14 · p. 19
• The public materials do not identify the three speakers’ Central Kurdish varieties or the 88 raters’ dialect backgrounds, nor specify grapheme-to-phoneme settings or whether syllable-marker removal was deliberate.Source: §14 · p. 19

Model

• The model card and repository files were retrieved at commit cdca52a, while the dataset card and viewer row count were also recorded on 8 September 2026.The model card quotations in the study refer to that revision.Source: Model · p. 20
• Related resources include Kurdish ASR and text resources, with the dataset page describing Central, Northern, and Southern Kurdish and announcing Hawrami and Badini as forthcoming.Source: Model · p. 20
• The demonstration was independently built, and upstream F5-TTS code is MIT-licensed while its pretrained weights are CC-BY-NC; TTS4ALL_Eval is GNU AGPL v3.0.Source: Model · p. 20
• The audiobook works, the male audiobook text’s rights holder, and whether the account belongs to the paper’s lead author are not established.Source: Model · p. 20