From Record to Resource

Abstract

The two companion papers describe a single system: a retrieval-augmented reading interface to the Bleek–Lloyd ǀxam folklore corpus, and a proposal to restore the ǀxam click pronunciation the interface cannot voice. This paper steps back from the one case to ask what it is an instance of. The claim is that the documentary records of endangered and sleeping languages have become tractable in a way they were not five years ago: retrieval-grounded language models can make a written record legible to a non-specialist, and community-grounded speech technology, neural and formant, can help restore or sustain the spoken language, so that the binding constraint is no longer model capability but the availability of data and the governance of its use.

We test that claim against a landscape of cases, drawing on the primary literature and reputable reporting: the Yaaku (Yakunte) and El Molo languages of Kenya; the community-owned Māori speech work of Te Hiku Media; the Inuktut technology programme of the National Research Council of Canada; the revival of Cornish from written records; and the singular, non-replicable revival of Hebrew. We survey what large language models actually do for documentation (interlinear glossing, dictionary work, translation from a grammar book) and what they fail at, and what neural and formant speech synthesis require. Two findings recur and bound the optimism. First, the capability is real but narrow: models assist with, rather than generate, an endangered language, and the gains come from existing parallel data and lexicons rather than from any emergent linguistic reasoning. Second, the decisive factors are not technical: intergenerational transmission, community control, and data sovereignty determine whether a tool helps or merely extracts. We close with what generalises from these cases and what does not, and where ǀxam sits within it.

1. Introduction

The first companion paper described an interface that makes the Bleek–Lloyd ǀxam corpus readable: a retrieval-augmented reading layer over roughly 12,000 to 15,000 pages of nineteenth-century notebooks, with generation tied to retrieved passages and a glossary, browse views, and English-language audio. The second described the limit of that audio: the text-to-speech cannot articulate the ǀxam click consonants (ǀ ǁ ǃ ǂ), and a path to an authentic pronunciation layer through the living sister language Nǀuu and reconstruction from comparative phonetics. Both papers are about one language and one collection.

This paper asks the more general question those two raise but do not answer. ǀxam is a sleeping language: fully documented, no living first-language speakers. It is one of thousands of languages that are endangered, dormant, or sleeping, and many of them sit behind extensive written or recorded documentation that is, like the Bleek–Lloyd material, public but not readable, archived but not usable. The question is whether the recent generation of language technology, large language models and neural speech synthesis, changes what can be done with those records, and if so, where the new limits lie.

The argument of this paper is a bounded one. There is a real shift: the same tooling that has, until recently, served only the world's high-resource languages can now be turned, with care, to the documentary record of a language with few or no speakers. A retrieval-grounded model can make a written record legible. Speech technology, where in-language audio or measured phonetic parameters exist, can help recover or sustain its sound. But the shift is narrower than the popular framing suggests, and the binding constraints have moved rather than vanished. The tools support access, reconstruction, and transmission of an existing record; they do not generate the language anew. Retrieval, not generation, and reconstruction, not invention, remain the ceilings. And the factors that decide whether a project helps a language community or merely mines it are not technical at all: they are about transmission, consent, and control.

Contributions

This paper makes no technical contribution; it is a synthesis. Its contributions are as follows.

1. A cross-case account of language technology applied to endangered and sleeping languages, assembled from the primary literature and reputable reporting, with each factual claim checked against its sources and the common "last speaker" and "learns a language from a book" overstatements corrected.
2. A separation of the readable problem from the audible problem, mapping which tools address each, and what each genuinely delivers as against what is claimed for it.
3. An honest map of the bounds, distinguishing capability gains that are real (in-context use of dictionaries and parallel data; speaker-timbre transfer) from those that are not (grammatical reasoning from prose; producing phonemes absent from training data; generating fluent novel text in a language with no corpus).
4. A statement of what generalises and what does not, locating ǀxam within the landscape and identifying the preconditions under which the pattern of the two companion papers could be repeated elsewhere.

2. A vocabulary and a map

2.1 Endangered, sleeping, dormant, extinct

The words used for languages in decline carry more weight than they appear to, and the popular ones are frequently wrong. A language with no first-language speakers but a complete documentary record is better called sleeping or dormant than extinct: the term records that the material to reawaken it exists. The Endangered Languages Project uses "dormant" and "awakening" for exactly this distinction. The "extinct" label, by contrast, is often premature. UNESCO reclassified Cornish from "extinct" to "critically endangered" in 2010, stating that the older label no longer reflected the situation, because the language had been revived from its written records and had acquired new speakers (see Section 5).

The single most common error in public accounts is the "last speaker" headline. It conflates distinct facts, the last first-language (native) speaker, the last fluent speaker, and the last member of an ethnic group, and it almost always oversimplifies a gradual decline into a single death. The Wukchumni case is the clearest caution: Marie Wilcox, who spent more than twenty years compiling a dictionary of the language and is widely called its "last speaker," was survived at her death in 2021 by at least three fluent speakers, including a daughter, precisely because she had taught them. The framing is not merely imprecise; it inverts the outcome, reporting a transmission success as a terminal loss. This paper therefore states speaker situations as ranges with dates, and avoids the lone-last-speaker formulation except where it is genuinely accurate.

2.2 The resource hierarchy

The technical literature has its own vocabulary for the same gradient. Joshi et al. (2020) propose a now-standard six-class taxonomy of the world's languages by the data available for them, from class 0, "the Left-Behinds," with almost no digital resources, to class 5, "the Winners," such as English. The point of the taxonomy is that natural-language processing, despite presenting itself as language-agnostic, overwhelmingly serves a handful of high-resource languages, and that the great majority of languages, including every endangered one, sit at the bottom with little or no labelled data. Endangered-language work is the extreme low-resource case: not merely scarce data, but, as Mager et al. (2018) note for the Indigenous languages of the Americas, often typologically distant from the high-resource languages on which models are trained, and frequently under-described. Hedderich et al. (2021) survey the methods that the field brings to bear, data augmentation, distant supervision, and above all cross-lingual transfer from related higher-resource languages. The recurring lesson of that literature, returned to throughout this paper, is that transfer and existing parallel data, not model scale alone, are what move the needle at the bottom of the hierarchy.

2.3 The tools

Four families of technology are relevant, and they should not be conflated.

Large language models and retrieval-augmented generation operate on text. A retrieval-augmented system answers from a corpus by retrieving relevant passages and conditioning generation on them, which is the architecture of the first companion paper. Used this way, the model is an interface to an existing record, not a source of new content.

Automatic speech recognition (ASR) transcribes speech to text. Modern systems are built on self-supervised models such as wav2vec 2.0 (Baevski et al. 2020) and its cross-lingual successor XLS-R (Babu et al. 2021), which pre-train on large amounts of unlabelled audio and can then be fine-tuned on comparatively little transcribed speech.

Neural text-to-speech (TTS) synthesises speech from text by learning from a corpus of recorded audio. It is data-hungry in a specific way discussed in Section 4.

Formant (rule-based) synthesis generates speech from a table of acoustic parameters rather than from recordings (Klatt 1980). It is the corpus-light alternative for languages with no usable speaker audio, at the cost of a robotic result.

The distinction between the last two, and the distinction between transferring a speaker's voice and teaching a model new sounds, turn out to be the crux of what is and is not possible for the sound of an endangered language, and they are treated in Section 4.

3. The readable problem: making written records legible

3.1 Retrieval over a documentary corpus

The first companion paper is one worked instance of a general move: putting a retrieval-grounded reading layer over a documentary corpus so that a non-specialist can navigate it. The design discipline there, every claim tied to a retrieved passage, no generation of novel ǀxam, the colonial conditions of recording kept in view, is not incidental. It is what distinguishes a faithful interface to a record from a model improvising over a culturally sensitive corpus. The general claim is modest and, within its bounds, well founded: where a language has an extensive written record, retrieval-grounded models can make that record readable and searchable for an audience that could not otherwise use it. This is an access gain, or more precisely a readability gain, the corpus was already public, and it does not depend on the model knowing the language.

3.2 What language models do for documentation

Beyond reading, a growing body of peer-reviewed work asks whether language models can assist the linguistic work of documentation itself. The evidence is genuinely encouraging in places and sobering in others, and the honest reading is that models assist rather than replace the documentary linguist.

The most studied task is interlinear glossing, the morpheme-by-morpheme annotation that is the backbone of language documentation. The SIGMORPHON 2023 shared task (Ginn et al. 2023) evaluated systems on six typologically diverse endangered languages (Arapaho, Gitksan, Lezgi, Natügu, Tsez, Uspanteko); the winning system improved on a strong baseline by a wide margin. GlossLM (Ginn et al. 2024a) then assembled the largest interlinear corpus to date, over 450,000 examples across roughly 1,800 languages, and pre-trained a model that exceeded the prior state of the art on in-domain languages. But the same group's study "Can we teach language models to gloss endangered languages?" (Ginn et al. 2024b) found that few-shot prompting of general LLMs (GPT-4o, Gemini 1.5 Pro, and others) beats weak baselines but generally underperforms dedicated supervised systems, and that the hardest genuinely low-resource languages remain largely unsolved: Gitksan morpheme accuracy sits around ten per cent regardless of method. No source in this literature claims the task is solved.

For dictionary work, language models are being used as lexicographic assistants: generating example sentences, modernising orthography, supporting community editing. Merx et al. (2024) document the consistent pattern: example quality "degrades significantly for lower-resourced languages." Community-controlled platforms such as Kuene for Hawaiian neologism (Walker et al. 2025) use the model as a supporting feature while the community committee remains the authority over the words, which is the right division of labour and is treated, in that work, as a governance requirement rather than a courtesy.

3.3 The grammar-book question, and a cautionary correction

The most striking recent claim was that a model could learn to translate a new language from a single grammar book. Tanzer et al. (2024) introduced the MTOB benchmark ("Machine Translation from One Book") for Kalamang, a language of fewer than two hundred speakers with almost no web presence, supplying a model with a grammar, a bilingual wordlist, and a few hundred parallel sentences. Their own result was measured: the best model fell short of a human who had studied the same materials for about ten hours. The claim escalated when the Gemini 1.5 technical report (Google DeepMind 2024) presented near-human-learner performance on the same materials as evidence of "in-context language learning from a single grammar book," a framing amplified well beyond the paper.

The correction is important because it bears directly on the central thesis. Two independent peer-reviewed studies ran the ablation the headline omitted. Aycock et al. (2025), an ICLR spotlight, found that "almost all improvements stem from the book's parallel examples rather than its grammatical explanations," and that a small conventional fine-tuned translation model was competitive with the much-larger in-context model. Pei et al. (2025), studying Manchu with an enciphered control to rule out memorisation, reached the same conclusion: "high-quality dictionaries and good parallel examples are very helpful," while "grammars hardly help." The capability is real, but it is the capability to exploit parallel data and lexicons in context, not to reason from grammatical description. For a language whose documentation is a grammar and a wordlist but little parallel text, this is a sharp limit, and it is exactly the situation of many sleeping languages.

3.4 What fails

The failures are as instructive as the successes, and they set the floor. Unaided frontier models score essentially zero on genuinely unseen endangered languages: Zhang et al. (2024) report GPT-4 at 0 BLEU before their in-context method raises it to 10.5, a figure that is itself still poor translation. Stap and Araabi (2023) conclude bluntly that "even large language models like GPT-4 are not well-suited for extremely low-resource languages," and that a smaller fine-tuned model outperforms them. The 2024 AmericasNLP shared task on translation into Indigenous languages saw improvement over baseline for only four of eleven languages, with a best margin of a few ChrF++ points.

Two methodological findings explain why the rosy numbers should be distrusted. LingOly (Bean et al. 2024), a contamination-controlled benchmark of linguistic reasoning across more than ninety mostly low-resource languages, found that even the best model reached under forty per cent on the harder problems, and concluded that "in absence of memorisation, true multi-step out-of-domain reasoning remains a challenge", that is, much apparent multilingual competence is recall of training data, not reasoning. Ahuja et al. (2024) found that most widely used multilingual benchmarks are contaminated with training data, which inflates scores precisely for the languages a model has seen. The corollary cuts in the field's favour only in a bleak way: for the most resource-scarce languages, there is too little text to memorise, so low scores there are genuine rather than deflated.

A final caution separates metric from usefulness. Shu et al. (2024) built a retrieval-augmented English-to-Cherokee system whose semantic similarity scores looked respectable (BERTScore around 0.96), yet human evaluators scored every model zero for Cherokee fluency and grammaticality. A good score on an automatic metric is not a usable sentence. The lesson for any endangered-language deployment is that evaluation must include speakers, and that the absence of speakers, the defining condition of a sleeping language, is therefore also the hardest evaluation problem.

The synthesis for the readable problem is this: language models are a real aid to making records legible and to assisting documentation, strongest when they are given existing dictionaries and parallel data to work from, and weakest, often useless, when asked to supply linguistic competence the data does not contain. This is the empirical basis for the paper's bound: retrieval and assistance, not generation.

4. The audible problem: restoring and sustaining sound

4.1 Speech recognition for living endangered languages

For a language with living speakers, the most consequential recent change is in ASR. Self-supervised pre-training decoupled the data requirement: wav2vec 2.0 (Baevski et al. 2020) reached usable accuracy on English from as little as ten minutes of transcribed audio, after pre-training on a large unlabelled corpus, and XLS-R (Babu et al. 2021) extended this across 128 languages so that a few hours of target-language audio, layered on a multilingual base, goes a long way. The headline low-data figures are English read-speech benchmarks and should not be read as typical for a field language, but the direction is real, and it is what makes community-scale ASR feasible.

Te Hiku Media's Māori work is the exemplar, and it is as much a governance model as a technical one (Section 6). The community gathered its own corpus through a crowd-sourcing campaign, Kōrero Māori, in March 2018: about 310 hours of transcribed speech from roughly 2,500 participants and 200,000 recordings, collected in ten days, incentivised by a community competition. The resulting ASR reached about 86 per cent word-level accuracy initially and around 92 per cent by 2024. The National Research Council of Canada's Indigenous Languages Technology programme is the institutional counterpart: funded from 2017, it released the Nunavut Hansard parallel corpus (Joanis et al. 2020), roughly 1.3 million aligned Inuktitut–English sentence pairs from two decades of legislative proceedings, the largest parallel corpus of an Indigenous language of the Americas, and built research translation, speech, and reading tools on an explicitly open-source, "empowerment-oriented" model. Both cases show ASR working for endangered languages at community scale; both also show that the work is gated by the existence of in-language audio and parallel text, not by the model.

4.2 The two roads to synthesised speech, and their data profiles

Restoring the sound of a language is a different problem, and the distinction between the two synthesis methods is where most public confusion lies.

Neural TTS learns to speak from a corpus of recordings. A high-quality single-speaker voice has conventionally needed on the order of twenty to twenty-four hours of clean studio audio (the LJSpeech standard); transfer from a multilingual base brings acceptable quality down to roughly an hour of target-language data, and dedicated low-resource pipelines lower it further (LRSpeech; Xu et al. 2020). The crucial qualification is phoneme coverage: the recordings must actually contain the language's sounds.

This is the point at which the popular "clone any voice from seconds of audio" capability is routinely misunderstood. Zero-shot voice cloning, VALL-E from a three-second prompt (Wang et al. 2023), YourTTS from under a minute (Casanova et al. 2022), and the commercial systems, transfers a speaker's timbre. It does not add phonemes the underlying model never learned. The proof is in the design of the cross-lingual systems themselves: VALL-E X (Zhang et al. 2023) can render a speaker's voice in another language only because that language was in its training data; the speaker prompt carries identity, while the model supplies the phonology. Gong et al. (2024) confirm the failure mode directly, intelligibility collapses (character error rates above forty per cent) for languages whose sounds are absent from pre-training, even with phonetic input. For a language with unusual phonology, this is the whole game. No amount of voice-cloning supplies clicks the model never heard.

Formant synthesis is the alternative for the hardest case, a language with no usable speaker recordings at all. It generates each sound from a table of acoustic parameters rather than from audio (Klatt 1980), and is the established method for documenting moribund languages: Koffi and Petzold (2022) build a formant synthesiser for Betine, a language of Côte d'Ivoire whose last speakers are dead, working entirely from measured acoustic parameters. The method is corpus-light but not data-free, it requires accurate phonetic measurements, which for an extinct language must come from whatever archival phonetic records survive. The contrast is instructive: Koffi's own later concatenative experiment on the same language needed a recorded databank of 2,318 syllables, exactly the speaker audio that formant synthesis avoids.

4.3 The click gap, and why ǀxam is the hard case

The companion voice paper argued that no current system can voice ǀxam clicks. The massively multilingual speech literature confirms this at the data level. Meta's MMS (Pratap et al. 2024) covers ASR and TTS for 1,107 languages, the largest such effort, by training on New Testament recordings. A direct check of its published language lists shows that no Khoisan or click language is supported for either ASR or TTS; ǀxam and Nǀuu appear in none of its models. The only click-consonant coverage is ASR for Zulu and Xhosa, two Bantu languages, and that ability was learned from in-language audio containing those clicks, not transferred from a click-free model. This is the general principle of Section 4.2 made concrete: the coverage of even a thousand-language model is bounded by which languages have recorded audio, and the Khoisan click languages, including ǀxam's living sister Nǀuu, have essentially none in these corpora.

ǀxam is therefore close to the worst case for the audible problem: a sleeping language, with no recordings of its own, whose distinctive sounds are absent from every large model, and recoverable only by the route the companion paper sets out, living Nǀuu phonetics or measured acoustic parameters, mapped onto a reconstructed ǀxam inventory. The honest ceiling there, accurate pronunciation rather than fluent synthetic narration, is the same ceiling this paper draws generally: reconstruction, not invention.

5. A landscape of cases

The thesis is best tested against specific languages. The cases below are chosen to span the range: two Kenyan languages that frame the problem, the community-technology exemplars, the canonical "revived from documents" precedent, and the famous outlier. Each is stated with the precision Section 2.1 demands. Appendix A collects them in a table.

5.1 Kenya: Yaaku and El Molo

The Yaaku (or Yiaku) people of the Mukogodo Forest in Laikipia, Kenya, speak, or spoke, Yakunte, a Cushitic language. It is among the six Kenyan languages the UNESCO Atlas listed as extinct in its 2010 edition (some later reporting gives 2020 and "critically endangered"; the robust fact is the 2010 listing). The speaker situation is exactly the "last speaker" trap of Section 2.1: headlines have ranged from "only seven" elderly fluent speakers around 2010 to two or three by the early 2020s, the decline is real, but there is no single last speaker, and the most defensible statement is a falling handful of very elderly first-language speakers plus one young revivalist. The language was abandoned in favour of Maa (Maasai) roughly between 1925 and 1936, under the stigma attached to the Yaaku's hunter-gatherer status among cattle-keeping neighbours.

What makes Yakunte relevant here is the revival, which is technology-touched in a modest, instructive way. A community teacher, Juliana Kageni, runs twice-weekly classes for a growing roll of learners (reported between roughly 250 and 390), building on a dictionary compiled by the Leiden-affiliated researcher Fleur Wensveen around 2007 and earlier vocabulary from Heine (1974). The "technology" is not an app, no dedicated Yakunte app is verifiable, but a WhatsApp group of around thirty participants exchanging text and voice messages, and a striking afforestation method in which trees in the Mukogodo Forest are tagged with metal cards bearing the Yakunte word on one side and English on the other: "as the trees grow, the language grows." It is a revival held together largely by one activist, and it is fragile.

El Molo, on the southeastern shore of Lake Turkana, is the contrasting Kenyan case and the more sobering one. Also Cushitic, it lost its last fully fluent speaker, Kaayo, around 1999, having had only about eight fluent speakers as far back as the 1970s. Like Yakunte it is UNESCO-2010-extinct (and "dormant" in the Endangered Languages Project's terms), and like Yakunte its community shifted to a Maa language, here Samburu, in the first half of the twentieth century; a residue of El Molo fishing vocabulary survives embedded in the community's Samburu speech. An Elmolo Development Group formed in 1995 and a revival programme began around 2006, but it was dropped around 2012, and no active technology-supported revival exists today. The two Kenyan languages are closely matched, both Cushitic, both shifted to Maa, both listed extinct in 2010, but one has an active if fragile community-led revival and the other does not. The difference is not technological. It is the presence of organised community effort, which is the variable this paper keeps returning to.

5.2 The Bleek–Lloyd case: ǀxam and Nǀuu

ǀxam sits at the documented-but-silent end of the range. Unlike Yakunte and El Molo, it has no living speakers and no community of heritage speakers shifting to a dominant neighbour; it has instead an unusually rich nineteenth-century written record (the subject of the first companion paper) and a living sister language, Nǀuu, whose last fluent speaker, Ouma Katrina Esau, anchors the pronunciation proposal of the second. It is the case where the readable problem is most fully solved (the record is extensive and now navigable) and the audible problem is hardest (no recordings, unsupported phonology). It thus tests both halves of the thesis at their extremes.

5.3 Te Hiku Media and te reo Māori

Māori is not endangered in the way the others are, te reo has hundreds of thousands of speakers and state support, but Te Hiku Media's work is the clearest demonstration of the audible thesis combined with the governance one. The community built its own ASR (Section 4.1) and, critically, retained control of the data and the models under a bespoke licence (Section 6). When a data vendor offered to pay for Māori audio, and when OpenAI's Whisper was found to have been trained on Māori and Hawaiian audio of undisclosed provenance, Te Hiku's response was not to withdraw from the technology but to insist on owning it. It is the positive existence proof: community-controlled, technically current, and explicitly non-extractive.

5.4 Cornish: revival from documents

Cornish is the load-bearing precedent for any project that hopes to reawaken a language from its records. Community transmission ended (the "last speaker," Dolly Pentreath in 1777, is itself a legend that oversimplifies a longer decline), but the language was reconstructed from its written corpus in the twentieth century, by Jenner (1904) and Nance's Unified Cornish (1929), and has since acquired a body of second-language speakers, a small number of children raised in it, and, in 2010, UNESCO's reclassification from "extinct" to "critically endangered." Cornish is the proof that a documented language can be brought back into use. It is also a caution about what "back into use" means: revived Cornish is a reconstruction, with necessary gaps filled by neologism and inference, not a seamless continuation of the old language. That is precisely the honest ceiling this paper draws for ǀxam sound, and the parallel is exact: reconstruction restores use, not the original unbroken thing.

5.5 Hebrew: the outlier that proves the rule

Hebrew is the case everyone cites and the one that generalises least. It is the only instance of a language without native speakers acquiring millions of them. But the conditions were singular and are not reproducible: Hebrew was never fully dormant, retaining continuous liturgical and textual use with its corpus intact; it was carried by a nationalist state-building project; and it served a diverse immigrant population that needed a common vernacular and had no shared rival. The revival is also not the pure recovery it is often taken for, Zuckermann argues that Modern Hebrew is a hybrid heavily shaped by the European mother tongues of its revivers. Scholars treat Hebrew as an outlier, not a template. Its lesson for endangered-language technology is deflationary: the decisive factors in the one total success were social and political, not linguistic or technical, which is the through-line of this entire landscape.

5.6 The documented-silent cases, briefly

Several further cases fill in the documented-but-silent end and reinforce the Section 2.1 cautions. Eyak (Alaska) lost its last first-language speaker, Marie Smith Jones, in 2008, but is dormant rather than dead: extensively documented by Krauss, it has a self-taught second-language speaker and an online revitalisation programme. Yaghan (Chile) lost its last fluent speaker, Cristina Calderón, in 2022, with a dictionary compiled by her granddaughter. Livonian (Latvia) lost its last native speaker in 2013 but has roughly forty second-language speakers, an institutional home (the University of Latvia Livonian Institute, 2018), and reportedly at least one new first-language child. Ainu (Japan) has effectively no fluent first-language speakers but gained legal recognition in 2019 and a national museum in 2020. In each, the documentary record persists and the question is transmission, not preservation, which is the distinction the next sections turn on.

6. Governance: the binding constraint

The recurring finding of the case landscape is that the decisive variable is not the model. It is whether the work is controlled by, and returns benefit to, the language community. A mature set of governance frameworks now codifies this, and they are not optional add-ons to the technical work; they are conditions on whether it is legitimate at all.

The frameworks most relevant here form a layered structure. At the level of international norms, the UN Declaration on the Rights of Indigenous Peoples (UNDRIP, adopted 13 September 2007) and its principle of free, prior, and informed consent (FPIC) set the backdrop, though UNDRIP is a declaration rather than binding law. At the level of data specifically, the CARE Principles for Indigenous Data Governance (Carroll et al. 2020), Collective benefit, Authority to control, Responsibility, Ethics, were formulated by the Global Indigenous Data Alliance to complement the FAIR principles: where FAIR is data-centric, CARE is people- and purpose-centric. The Canadian First Nations OCAP principles (Ownership, Control, Access, Possession), established in 1998, make the same point as a matter of governance. Most directly relevant to this corpus, the San Code of Research Ethics (2017), issued by the South African San Council and the South African San Institute, was the first research-ethics code authored by an Indigenous people in Africa. It commits researchers to respect, honesty, justice and fairness, and care, together with a process of community approval, and it was prompted by exactly the extractive harms the technology now risks repeating: the patenting of San traditional knowledge of the Hoodia plant, and a 2010 genomic study whose conclusions San leaders condemned as "private, pejorative, discriminatory and inappropriate." Any ǀxam-related work sits squarely within the reach of this code.

The technology community has its own instruments and critiques converging on the same place. Te Hiku Media's Kaitiakitanga Licence operationalises guardianship: it forbids commercial use without permission and binds the data to the community's benefit. The Masakhane project's participatory model (Nekoto et al. 2020) argues that "'low-resourced'-ness is a complex problem going beyond data availability and reflects systemic problems in society," and builds African-language resources with the communities rather than for them. Against these, the critical literature names the danger plainly. Couldry and Mejias (2019) frame the wider phenomenon as data colonialism, the appropriation of human life as raw material for extraction. Bird (2020, 2024) argues that treating Indigenous languages as "data ready for the taking" re-enacts the causes of language endangerment, and calls for a relational rather than extractive practice. Hutchinson (2024) raises the provenance and consent problems of the religious-text corpora, the Bible translations, on which much low-resource speech technology, MMS included, is built. Cooper et al. (2024) conclude from interviews with researchers in Aboriginal and Torres Strait Islander contexts that it is "how you do things that matters", that process, not artefact, is the measure.

There is a concrete cautionary tale that ties the technical and the ethical together. When OpenAI's Whisper was found to contain over a thousand hours of Māori audio of undisclosed origin, Te Hiku's response, published as "OpenAI's Whisper is another case study in colonisation," was that they had themselves fine-tuned a better Māori model but would never deploy it, because to do so would violate their own licence and their obligations to the community whose voices it was built from. Mahelona's summary, "data is the last frontier of colonization," and Jones's, that the same powers that "physically beat" the language out of his grandparents "now want to sell our language back to us as a service," state the stakes more sharply than any framework. The governance question is not downstream of the technical one. For an endangered-language project it is the first question.

7. What generalises, and what does not

The two companion papers describe a pattern: a retrieval-grounded reading layer over a documentary record, and a community-sourced or reconstructed pronunciation layer over its sound. The landscape above lets us say, with some precision, when that pattern can be repeated and when it cannot.

What generalises. The readable half generalises well. Wherever a language has an extensive written record, a retrieval-grounded interface can make that record legible and searchable for a non-specialist, with the design discipline of the first companion paper, grounding, refusal to generate, provenance kept visible. This is the most portable part of the work, because it depends on the record, not on the language being known to any model. The audible half generalises conditionally: where a language has living speakers, or a living close relative, or adequate archival phonetics, its sound can be recovered or sustained, by ASR and neural TTS in the speaker-rich case (Te Hiku, NRC), by reconstruction from a sister language in the ǀxam case, or by formant synthesis in the no-audio case (Betine). The governance frameworks generalise across all of it: the San Code, CARE, OCAP, and the Kaitiakitanga model are not specific to one language and define the conditions of legitimate work everywhere.

What does not. Generation does not generalise, because it was never available: no model can produce fluent novel text in a language with no substantial corpus, and for sleeping languages this will not change, the data does not exist and cannot be manufactured. Grammatical reasoning from description does not transfer (Section 3.3): a grammar and a wordlist, the typical documentation of a sleeping language, are not enough for a model to translate, because the gains come from parallel data that such languages lack. Phonology absent from training data cannot be voice-cloned into being (Section 4.2), which is why ǀxam clicks remain the hard case. And, most importantly, technology does not generalise into transmission. The clearest lesson of the landscape, from Wukchumni's surviving fluent speakers to the divergence between Yakunte and El Molo to the social engine behind Hebrew, is that what keeps a language alive is people speaking it to one another, above all across generations. The scholarly consensus is consistent and deflationary: technology is necessary-but-insufficient, a support for transmission and never a substitute for it. A tool that produces readable text or recovered sound but does not feed a living community of use has preserved a record, not a language.

Where ǀxam sits follows from this. It is a language whose readable problem is largely solved and whose audible problem is solvable in the bounded sense of accurate pronunciation reconstructed from Nǀuu. What it does not have, and what no interface can supply, is a community of speakers to transmit to. The honest description of the Bleek–Lloyd work is therefore that it makes an extraordinary record legible and, in time, audible, an act of access and reconstruction worth doing on its own terms, while being clear that this is preservation of the record, and that the living future of Tuu-family speech, if there is one, runs through the Nǀuu community and Ouma Katrina Esau's school, not through any model.

8. Conclusion

The claim this paper set out to test was that language technology has made the documentary records of endangered and sleeping languages tractable, and that the binding constraints are now data and governance rather than capability. The evidence supports a bounded version of it. Retrieval-grounded models genuinely make written records readable; ASR and neural TTS genuinely recover and sustain the sound of languages that have speakers or close relatives or archival phonetics; and formant synthesis covers the hardest no-audio case. These were not available, at this quality or coverage, five years ago, and the opportunity they create for the records of endangered languages is real.

But the constraints have moved, not disappeared, and they are unforgiving in specific ways. The capability is to assist, not to generate; to exploit existing parallel data and lexicons, not to reason from grammatical description; to transfer a known voice, not to invent unheard sounds. And the factor that decides whether a project serves a language community or merely extracts from it is governance: consent, control, and the return of benefit, codified in the San Code, CARE, OCAP, and the Kaitiakitanga model, and demonstrated, positively, by Te Hiku Media. The deepest lesson of the landscape is the most deflationary one for a technologist: the survival of a language is a matter of transmission between people, and no tool substitutes for that.

For ǀxam, the conclusion is clear-eyed. The technology makes the Bleek–Lloyd record readable and, with the proposal of the companion paper, audible in the bounded sense of authentic pronunciation. That is a real and worthwhile act of access and reconstruction. It is preservation of a remarkable record, offered in service of the communities and scholars who care for it, and not a claim to have brought the language back. Within those limits, and only within them, the tools of this decade are an opportunity that the records of the world's silenced languages have not had before.

Acknowledgements

This paper, like its two companions, builds on Pippa Skotnes's recovery of the Bleek–Lloyd archive, without which the ǀxam record would be neither legible nor reachable. It is indebted, too, to the communities and scholars whose work it surveys, above all those who have insisted, often against the grain of the technology, that an endangered language belongs to the people who speak it.

References

Endangered languages, revitalisation, and the case studies

Endangered Languages Project (ELCat). Language entries and classification methodology ("dormant", "awakening"). https://www.endangeredlanguages.com

Heine, B. (1974/75). Notes on the Yaaku language. Afrika und Übersee.

UNESCO (2010). Atlas of the World's Languages in Danger, 3rd edition. Paris: UNESCO.

On Yaaku/Yakunte: reporting in Al Jazeera (2018), the Christian Science Monitor (2022), Global Voices (2024), and the Daily Nation (2022); and the Yaaku community project (yaaku.org). Speaker figures and the UNESCO date are treated as a range in the text owing to source disagreement.

On El Molo: Tosco, M. (2012), and the Endangered Languages Project El Molo entry; Glottolog (Campbell et al. 2022).

On Cornish: Jenner, H. (1904). A Handbook of the Cornish Language. London. Nance, R. M. (1929). Unified Cornish. UNESCO reclassification reported 2010; speaker figures from the 2011 Census (Office for National Statistics).

On Hebrew: Zuckermann, G. (2020). Revivalistics: From the Genesis of Israeli to Language Reclamation in Australia and Beyond. Oxford University Press.

On Eyak, Yaghan, Wukchumni, Livonian, and Ainu: obituaries and reporting in NPR, the Boston Globe, and Global Voices; the Alaska Native Language Center (Eyak); the University of Latvia Livonian Institute; Japan's Ainu Promotion Act (2019); and the documentary Marie's Dictionary (Global Oneness Project, 2014). Wukchumni speaker status corrected per the surviving fluent speakers reported at Wilcox's death (2021).

Natural-language processing for endangered and low-resource languages

Joshi, P., Santy, S., Budhiraja, A., Bali, K., and Choudhury, M. (2020). The State and Fate of Linguistic Diversity and Inclusion in the NLP World. ACL 2020, 6282–6293.

Mager, M., Gutierrez-Vasques, X., Sierra, G., and Meza-Ruiz, I. (2018). Challenges of Language Technologies for the Indigenous Languages of the Americas. COLING 2018.

Hedderich, M. A., Lange, L., Adel, H., Strötgen, J., and Klakow, D. (2021). A Survey on Recent Approaches for Natural Language Processing in Low-Resource Scenarios. NAACL 2021, 2545–2568.

Ginn, M., Moeller, S., Palmer, A., Stacey, A., Nicolai, G., Hulden, M., and Silfverberg, M. (2023). Findings of the SIGMORPHON 2023 Shared Task on Interlinear Glossing. SIGMORPHON 2023.

Ginn, M., Tjuatja, L., He, T., Rice, E., Neubig, G., Palmer, A., and Levin, L. (2024a). GlossLM: A Massively Multilingual Corpus and Pretrained Model for Interlinear Glossed Text. EMNLP 2024.

Ginn, M., Hulden, M., and Palmer, A. (2024b). Can we teach language models to gloss endangered languages? Findings of EMNLP 2024.

Merx, R., Vylomova, E., and Kurniawan, K. (2024). Generating bilingual example sentences with large language models as lexicography assistants. arXiv:2410.03182.

Walker, S., Wu, W., Torres Fischer, B., and Kimura, L. (2025). Kuene: A Web Platform for Facilitating Hawaiian Word Neologism. ComputEL-8.

Tanzer, G., Suzgun, M., Visser, E., Jurafsky, D., and Melas-Kyriazi, L. (2024). A Benchmark for Learning to Translate a New Language from One Grammar Book. ICLR 2024.

Google DeepMind (2024). Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv:2403.05530.

Aycock, S., Stap, D., Wu, D., Monz, C., and Sima'an, K. (2025). Can LLMs Really Learn to Translate a Low-Resource Language from One Grammar Book? ICLR 2025 (Spotlight). arXiv:2409.19151.

Pei, R., Liu, Y., Lin, P., Yvon, F., and Schütze, H. (2025). Understanding In-Context Machine Translation for Low-Resource Languages: A Case Study on Manchu. ACL 2025. arXiv:2502.11862.

Zhang, K., Choi, Y. M., Song, Z., He, T., Wang, W. Y., and Li, L. (2024). Hire a Linguist!: Learning Endangered Languages with In-Context Linguistic Descriptions. Findings of ACL 2024.

Stap, D., and Araabi, A. (2023). ChatGPT is not a good indigenous translator. AmericasNLP 2023.

Bean, A., et al. (2024). LingOly: A Benchmark of Olympiad-Level Linguistic Reasoning Puzzles in Low-Resource and Extinct Languages. NeurIPS 2024 Datasets and Benchmarks. arXiv:2406.06196.

Ahuja, S., Gumma, V., and Sitaram, S. (2024). Contamination Report for Multilingual Benchmarks. arXiv:2410.16186.

Shu, P., Chen, J., Liu, Z., et al. (2024). Transcending Language Boundaries: Harnessing LLMs for Low-Resource Language Translation. arXiv:2411.11295.

Ebrahimi, A., et al. (2024). Findings of the AmericasNLP 2024 Shared Task on Machine Translation into Indigenous Languages. AmericasNLP 2024.

Speech technology

Klatt, D. H. (1980). Software for a cascade/parallel formant synthesizer. Journal of the Acoustical Society of America 67(3): 971–995.

Baevski, A., Zhou, H., Mohamed, A., and Auli, M. (2020). wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations. NeurIPS 2020. arXiv:2006.11477.

Babu, A., Wang, C., Tjandra, A., et al. (2021). XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale. arXiv:2111.09296.

Pratap, V., Tjandra, A., Shi, B., et al. (2024). Scaling Speech Technology to 1,000+ Languages. Journal of Machine Learning Research 25. arXiv:2305.13516.

Xu, J., Tan, X., Ren, Y., Qin, T., Li, J., Zhao, S., and Liu, T.-Y. (2020). LRSpeech: Extremely Low-Resource Speech Synthesis and Recognition. KDD 2020.

Wang, C., Chen, S., Wu, Y., et al. (2023). Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (VALL-E). arXiv:2301.02111.

Zhang, Z., et al. (2023). Speak Foreign Languages with Your Own Voice: Cross-Lingual Neural Codec Language Modeling (VALL-E X). arXiv:2303.03926.

Casanova, E., Weber, J., Shulby, C., Candido Junior, A., Gölge, E., and Ponti, M. (2022). YourTTS: Towards Zero-Shot Multi-Speaker TTS and Zero-Shot Voice Conversion for Everyone. ICML 2022.

Gong, C., Cooper, E., Wang, X., et al. (2024). An Initial Investigation of Language Adaptation for TTS Systems under Low-resource Scenarios. arXiv:2406.08911.

Koffi, E., and Petzold, M. (2022). A Tutorial on Formant-Based Speech Synthesis for the Documentation of Critically Endangered Languages. Linguistic Portfolios 11(1), Article 3.

Joanis, E., Knowles, R., Kuhn, R., Larkin, S., Littell, P., Lo, C.-k., Stewart, D., and Micher, J. (2020). The Nunavut Hansard Inuktitut–English Parallel Corpus 3.0. LREC 2020.

Te Hiku Media. Papa Reo and the Kōrero Māori corpus. https://papareo.nz ; ASR accuracy reported by NVIDIA (2024).

Governance, ethics, and the critique of extraction

United Nations (2007). Declaration on the Rights of Indigenous Peoples (UNDRIP). Adopted 13 September 2007.

Carroll, S. R., Garba, I., Figueroa-Rodríguez, O. L., et al. (2020). The CARE Principles for Indigenous Data Governance. Data Science Journal 19: 43.

First Nations Information Governance Centre. The First Nations Principles of OCAP. https://fnigc.ca

South African San Council and South African San Institute (2017). San Code of Research Ethics. Cape Town. (TRUST project / EU Horizon 2020.)

Kukutai, T., and Taylor, J. (eds.) (2016). Indigenous Data Sovereignty: Toward an Agenda. ANU Press.

Couldry, N., and Mejias, U. A. (2019). The Costs of Connection: How Data Is Colonizing Human Life and Appropriating It for Capitalism. Stanford University Press. See also "Data Colonialism," Television & New Media 20(4): 336–349.

Bird, S. (2020). Decolonising Speech and Language Technology. COLING 2020, 3504–3519. Bird, S. (2024). Must NLP be Extractive? ACL 2024.

Nekoto, W., et al. (2020). Participatory Research for Low-resourced Machine Translation: A Case Study in African Languages (Masakhane). Findings of EMNLP 2020, 2144–2160.

Hutchinson, B. (2024). Modeling the Sacred: Considerations when Using Religious Texts in Natural Language Processing. Findings of NAACL 2024, 1029–1043.

Cooper, N., Heldreth, C., and Hutchinson, B. (2024). "It's how you do things that matters": Attending to Process to Better Serve Indigenous Communities with Language Technologies. EACL 2024, 204–211.

Mahelona, K., Leoni, G., Duncan, S., and Thompson, M. (2023). OpenAI's Whisper is another case study in Colonisation. Papa Reo blog.

Appendix A: The cases at a glance

The table summarises the languages discussed, with the precision Section 2.1 requires. "Status" follows the documentary record where sources disagree on labels; speaker figures are given as the defensible range with the caution that "last speaker" framings are frequently inaccurate.

Appendix B: A note on sourcing

The factual claims in this paper were assembled from the primary literature (conference proceedings and journals, cited above) and from reputable reporting, and were checked for the recurring failure modes of this subject: the "last speaker" overstatement (Section 2.1), the "learns a language from a book" overstatement (Section 3.3), the conflation of automatic metrics with usability (Section 3.4), and vendor claims presented as peer-reviewed findings (flagged in the text, for example the Gemini 1.5 grammar-book result and organisational ASR accuracy figures). Where sources disagree, on the UNESCO classification date for Yakunte, on speaker counts in several cases, the disagreement is stated rather than resolved by picking a number. Two specific corrections to commonly repeated claims are made explicitly: that Wukchumni did not die with Marie Wilcox, and that Meta's MMS, despite its thousand-language coverage, supports no Khoisan or click language for either speech recognition or synthesis.