Restoring the Sound of a Sleeping Language

Abstract

The Bleek–Lloyd archive preserves the words of ǀxam, a sleeping San language of the South African Karoo, but not its sound. The reading interface described in the companion paper renders the corpus aloud in English but cannot articulate the ǀxam click consonants (ǀ ǁ ǃ ǂ); the click characters are removed before synthesis, so the informant ǁkabbo is voiced "kabbo."

This paper examines that limitation. A survey of current speech technology and the primary literature indicates that no available system can produce ǀxam clicks, and that this follows from training data rather than configuration: neural models reproduce a speaker's timbre, not new phonemes, and none has been shown to generate click consonants absent from its pretraining. Two approaches could in principle address it. The first is a low-data formant-synthesis method, for which the acoustic parameters of Nǀuu clicks have already been measured (in part from Ouma Katrina Esau's own speech); the second is a higher-data neural method. Both would be bridged to ǀxam by linguistic reconstruction from the living sister language Nǀuu. On that basis the paper sets out a phased, consent-first, community-led proposal for an authentic pronunciation layer: the clicks and the corpus's recurring names voiced correctly, attested by living Tuu-family speakers and reconstructed by linguists. The realistic end state is accurate pronunciation, not a synthetic voice narrating in ǀxam, for which no data source exists. The broader claim is that modern speech synthesis, used on this foundation, offers endangered and sleeping languages a practical opportunity to recover a sound that no recording preserved, bounded by available data and grounded in reconstruction rather than invention.

1. The missing sound

ǀxam is a sleeping language: fully documented in the nineteenth-century notebooks of Wilhelm Bleek and Lucy Lloyd, but with no living mother-tongue speakers. A retrieval-augmented interface (described in the companion paper) makes that corpus readable, searchable, and, in English, audible, reading replies aloud with a word-level karaoke highlight.

The audio pre-processing chain, however, removes the click characters ǀ ǁ ǃ ǂ before sending text to the speech engine, because the English-trained voices mishandle them. The interface therefore reads every story aloud while mispronouncing the archive's names: ǁkabbo becomes "kabbo." Because the clicks are integral to the language, this is a substantive limitation, and it is the subject of this paper.

The clicks are phonemes, not optional ornamentation. A click-free rendering of ǁkabbo is not an accented version of the name but a different word, missing its initial consonant. Much of the corpus's distinctive vocabulary, and the names of its informants, are click-bearing; a system that cannot produce ǀ ǁ ǃ ǂ cannot say them.

2. ǀxam, Nǀuu, and the Tuu family

ǀxam belongs to the Tuu family (formerly "Southern Khoisan"), the name Güldemann (2005) proposed for the family, after the noun *tuu 'people'. The family has two branches, Taa and !Ui, and ǀxam sits in the !Ui branch, documented in the Strandberg, Katkop and Achterveld dialects recorded by Bleek and Lloyd.

The relevant point is an asymmetry. ǀxam is documented but silent; its closest living relative, Nǀuu (the Nǁng cluster, also !Ui), is endangered but still spoken by Ouma Katrina Esau, its last fluent speaker. Of the whole Tuu family, the only modern survivors are several Taa varieties in Botswana and Namibia and, in South Africa, the remnants of the Nǁng cluster. For this reason Nǀuu, rather than the geographically closer Bantu languages, is the relevant reference point for ǀxam's sound: the phonetics ǀxam lost can be approached through the phonetics Nǀuu retains. (Kora, sometimes compared to ǀxam for its click inventory, is a Khoekhoe language, not Tuu; that comparison is one of click typology across family lines, not common descent.)

3. Why no tool can voice ǀxam clicks today

This section is drawn from a survey of the primary literature and current product documentation (June 2026); sources are in the References.

3.1 Existing speech systems

A survey of current speech technology finds no dedicated text-to-speech or speech-recognition system for any Khoisan/San language.

• Meta's Massively Multilingual Speech (MMS), which provides TTS for over 1,100 languages, does not cover Khoekhoe/Nama or any Tuu-family language. Where MMS does produce clicks (Zulu, Xhosa), it learned them from in-language training audio; it did not generalise them.
• ElevenLabs (the engine the reading interface uses) supports no Khoisan language across any of its text-to-speech models (Eleven v3, ~74 languages; Multilingual v2, 29; Flash v2.5, 32). Its only Southern-African synthesis language is Afrikaans (v3 only), which has no clicks. Its speech-to-text model can transcribe Xhosa and Zulu (the Bantu click-borrowing languages) but those cannot be synthesised, and no Khoisan language appears even for transcription. Its phoneme-level controls (SSML IPA <phoneme> tags and .PLS pronunciation dictionaries) are honoured by only some models, and its newest model accepts inline IPA but renders it probabilistically (~80–90% consistently by its own documentation). None of these mechanisms can produce a phoneme the model never learned: feeding the click symbols ǀ ǁ ǃ ǂ most likely yields silence, an ignored glyph, or a nearby pulmonic consonant. Its Impact Program (whose "Culture" pillar names oral histories and heritage) is therefore better understood as a channel through which to ask about click support or a custom voice than as an existing product route.

A click-capable voice for any Khoisan language is thus an unaddressed gap rather than a solved problem.

3.2 Neural TTS: corpus requirements

Modern neural voices are trained on large corpora: roughly 27,000 hours across 16 languages for Coqui XTTS, ~50,000–60,000 hours for the VALL-E family, ~32 hours each across 1,100+ languages for MMS; a single-speaker English voice from scratch (the LJSpeech benchmark) is about 24 hours.

The apparent counter-example is that voices can now be cloned from seconds of audio: 3-second prompts for VALL-E and XTTS, under a minute for YourTTS, under two minutes for ElevenLabs' instant cloning. Such cloning transfers a speaker's timbre, conditioned on phonemes the base model already represents; it does not add a new phonetic inventory. No published model produces click consonants zero-shot from click-free pretraining. The one partial exception, phonological-feature representations that approximate unseen sounds (Staib et al., Interspeech 2020), is contested and has not been shown to yield a full click inventory. A natural neural Nǀuu/ǀxam voice would therefore require click-bearing, in-language audio; the realistic fine-tuning target is on the order of 5–10 hours of clean, transcribed, click-rich speech, collected through a recording programme.

3.3 Formant synthesis: a low-data alternative

The older alternative is better suited to this case. Formant / articulatory synthesis (Klatt-style) builds each phoneme from an acoustic-parameter table rather than a corpus, and is an established method for moribund languages with few or no speakers (Koffi & Petzold, Linguistic Portfolios 11, 2022, demonstrated on the moribund West African language Betine). It requires no corpus, only measured acoustic parameters.

For Nǀuu clicks, those parameters have already been measured. Exter (2011), "The Acoustic Modeling of Click Types" (ICPhS XVII), constructs a source–filter model of abruptly-released clicks and validates it against Nǀuu recordings from three speakers (Ouma Katrina Esau, Ouma Anna Kassie and Ouma Hanna Koper), reporting formant targets (apical click F1 ≈ 1284 Hz, F2 ≈ 4917 Hz; laminal F1 ≈ 1885 Hz, F2 ≈ 4425 Hz; apical roughly an order of magnitude more intense than laminal). The paper is an acoustic analysis rather than a synthesiser, but it provides the kind of parameter set such a synthesiser would draw on, derived in part from the last fluent Nǀuu speaker's own recordings. The related descriptive work (Miller et al., JIPA 39(2), 2009) classifies all 73 Nǀuu consonants, and further published work on click acoustics (Sands) provides additional grounding.

3.4 The Nǀuu → ǀxam bridge is reconstruction, not a second model

There is no ǀxam audio to train on, and there will not be. The path from Nǀuu to ǀxam is linguistic rather than computational: living Nǀuu phonetics (recorded, or read from Exter's tables) mapped onto a reconstructed ǀxam inventory and phonotactics (du Plessis 2018), and assembled into ǀxam words. A formant or concatenative approach makes this mapping relatively direct. A neural Nǀuu voice could in principle be driven with ǀxam phoneme sequences, but ǀxam prosody and coarticulation are unattested, so the suprasegmentals would be reconstructed rather than learned.

4. Precedents and comparable projects

The proposal does not stand alone. A small but maturing body of work has built speech technology for endangered, under-resourced and Indigenous languages, increasingly under community-defined governance. None addresses a click language, or a sleeping language reconstructed through a living sister tongue, so the proposal occupies a genuine gap; but each precedent settles questions of data scale, method, and governance. The full review, with figures and citations, is given in the supplement.

The most instructive precedent is Te Hiku Media's Papa Reo programme in Aotearoa New Zealand, which built a te reo Māori speech recogniser from over 300 hours of community-contributed speech, crowdsourced through its Kōrero Māori campaign (more than 2,500 contributors reading over 200,000 phrases in about ten days), reaching roughly 92 per cent reported accuracy (Te Hiku Media and NVIDIA, 2022). Two findings carry over: a usable low-resource model was built from a few hundred hours rather than the tens of thousands used for commercial English voices; and the data is released under the Kaitiakitanga Licence, which treats data as held in guardianship, with benefit returning to the source community. Te Hiku has publicly declined to surrender its corpus to external models, arguing that indiscriminate scraping of Indigenous-language data reproduces a colonial pattern. The Lauleo project carried this model directly to Hawaiian (ʻŌlelo Hawaiʻi), in collaboration with Te Hiku and the University of Hawaiʻi at Hilo, showing that the approach transfers to a second community under a derived guardianship licence.

On data scale and tooling, the National Research Council of Canada built Inuktut speech recognition from on the order of seventy-five hours of transcribed speech and released the roughly 1.3-million-pair Nunavut Hansard parallel corpus (Joanis et al., 2020), and Australia's Elpis (Foley et al., 2018) lets community language workers train their own recognisers without machine-learning expertise. Mozilla Common Voice confirms that the field treats tens of hours as a meaningful low-resource working point, and that it holds no Khoisan dataset, just as Meta's Massively Multilingual Speech has no Khoisan coverage. The low-resource neural-synthesis literature places intelligible new-language voices at a few hours of paired, transcribed audio, with roughly twenty to thirty minutes marginal, which is the basis for this proposal's five-to-ten-hour neural target.

The decisive negative finding is that no text-to-speech or speech-recognition system exists for any Tuu, Kxʼa or Khoe-Kwadi language, and no formant or articulatory synthesiser exists for any click language. Where mainstream systems produce clicks at all, it is for the Bantu click-borrowing languages (isiXhosa, isiZulu) and only because in-language training audio contained them; the clicks were learned, not generalised. The acoustic groundwork for the relevant language nonetheless exists, measured in part from Esau's own speech (Exter 2011; Miller et al. 2009): the science to build clicks exists in pieces but has not been assembled into a working voice. That is the space this proposal enters.

Taken together the precedents support the proposal's positions: a few hours suffices for fine-tuned synthesis, so the binding constraint is the session capacity of a single elderly speaker, not raw volume; the low-data formant route is better matched to a sleeping language than the data-hungry neural route; the field favours tools community members can operate; and community-defined governance (the Kaitiakitanga Licence, Indigenous data sovereignty, and in this jurisdiction the San Code of Research Ethics) is the established operating frame, not an optional overlay.

5. Proposal: an authentic ǀxam pronunciation layer

5.1 Vision and deliverable

This paper proposes a ǀxam pronunciation resource integrated into the archive interface, to be developed with the Nǀuu-speaking community and linguists specialising in ǀxam and the Tuu family. The realistic end state is accurate pronunciation: the clicks and the corpus's recurring names and terms voiced correctly, attested by living sister-language speakers and reconstructed by linguists. It is not a synthetic voice narrating in ǀxam, for which no data source exists. The deliverable comprises:

• the four clicks, demonstrated by living Tuu-family speakers;
• the corpus's ~40–60 recurring click-bearing names and terms, voiced correctly;
• optionally, a low-data formant-synthesis model for click-bearing words beyond the recorded set,

surfaced as audio against glossary terms and informant names, so that the reading interface also supports listening.

5.2 Data requirements

The data needed differs by approach (sources in the References; a fuller data and cost model is given in the supplement):

Existing Nǀuu audio is a useful starting point but not a ready TTS corpus: the ELAR archive holds approximately 22 hours across the last 10 speakers (deposit dk0089), and a Nǀuu revitalisation project holds recordings of Ouma Katrina Esau and her students, including time-aligned audio of the 160-page Nǀuu reader (deposit dk0505). This is multi-speaker documentation, valuable for acoustic reference and reconstruction but not a clean single-voice TTS corpus.

5.3 Phased plan

• Phase 1 (low cost, in-app). Wire freely-licensed IPA click recordings into the glossary as "hear the click" controls. Demonstrates the interface and gives collaborators a concrete reference.
• Phase 2 (one speaker or phonetician). Record the ~40–60 recurring click-bearing names and terms with an Nǀuu speaker or phonetician; attach as per-term audio.
• Phase 3 (grant-scale). A formant-synthesis pronunciation model (low data, drawing on Exter's parameters) and/or a neural Nǀuu voice (5–10 h corpus), developed with linguists providing phonological reconstruction and Khoesan expertise, in collaboration with the Nǀuu-speaking community, under an endangered-language documentation grant and/or industry support. This phase would be the principal grant deliverable.

5.4 Roles

An indicative division of labour, by where the rights and expertise would sit. Nothing here is arranged: no individuals or institutions have been approached or have agreed to take part, and none are named for that reason. This is a hypothetical structure, not a commitment by anyone.

• Linguistic direction and the recording programme. ǀxam phonological reconstruction, broader Khoesan expertise, and community liaison, together with the Nǀuu speakers whose voices the recordings would capture. An academic institution with the relevant expertise would typically act as applicant of record and hold this work; cultural stewardship would sit with the speaker community.
• The speaker community. The Nǀuu-speaking community would own the recordings and any trained voice, and would set the terms of participation.
• The interface layer. The reading and listening surface, and the integration of the pronunciation audio into the archive system (the present author's part).

5.5 Ethics and data governance

Data governance is integral to this work, not an addendum. It involves an elderly last speaker, and the constraint is human rather than computational: a recording programme is limited by her voice and session capacity and must be paced accordingly. It should follow the CARE Principles for Indigenous Data Governance (Collective benefit, Authority to control, Responsibility, Ethics; Carroll et al., 2020): free, prior and informed consent, community ownership of the recordings, and benefit returning to the community. The governing principle is engagement rather than extraction, which is also why the reading system has so far deferred a synthesised voice pending a partnership of this kind.

Because the speaker and her community are San, the binding instrument is the San Code of Research Ethics (South African San Council, 2017), the first research-ethics code issued by an Indigenous African community, which requires prior approval from the San Council and commits researchers to fairness, respect, care and honesty (Schroeder et al. 2019; Callaway 2017). Beneath it sits free, prior and informed consent (UNDRIP, 2007, Articles 19 and 32), which here demands that the speaker understand, in Nǀuu, what a synthetic voice is and can be made to do. CARE governs the data lifecycle that follows; OCAP (First Nations Information Governance Centre) supplies a concrete possession model under which the community, not the institution, holds the master recordings and the trained voice; and any academic, cultural or commercial value generated should be the subject of an explicit benefit-sharing agreement, on the model of the San's own Hoodia precedent. The full treatment, with citations, is in the supplement.

5.6 Funding routes (pursued in parallel)

• ELDP (Endangered Languages Documentation Programme; Arcadia-funded, hosted at the Berlin-Brandenburgische Akademie der Wissenschaften): documentation of an endangered/sleeping language. The current round closes 1 October 2026.
• Lacuna Fund: if framed as an open African-language click-speech dataset (it funds African speech/ASR datasets).
• ElevenLabs Impact Program: its Culture pillar explicitly names oral histories and heritage; approved nonprofits receive a free Pro tier, and it is the natural channel to ask ElevenLabs directly about click support or a custom voice.
• Anthropic's Claude for Nonprofits (discounted access for the chat layer). Note that Claude Corps is US-501(c)(3)-only, so a bid led by a non-US institution would not qualify for that specific programme.
• Endangered Archives Programme (British Library, Arcadia-funded): for any associated archival digitisation.

Relevant institutional context: the Bleek–Lloyd collection is a UNESCO Memory of the World inscription, and the Digital Bleek & Lloyd is an established multi-institution digitisation of it.

5.7 Risks and limits

• Neural-clicks risk. A fine-tuned neural voice may still fail to produce clean clicks if the in-language data is insufficient. The formant approach is the lower-risk option, since it does not depend on a model learning a new phoneme class. This is the largest unverified technical risk and warrants early empirical testing.
• Data-collection risk. The most important recordings can come only from a small, ageing group of speakers; the programme has to be designed around that constraint from the start.
• Ceiling. Even if fully funded and successful, the result is accurate pronunciation (clicks and key names voiced correctly), not fluent generative ǀxam. This should be stated clearly to funders and community partners at the outset.

The principal risks, graded as planning judgements (likelihood and impact low / medium / high) rather than measured probabilities:

6. Conclusion

The reading interface makes the words of a sleeping language legible but not its sound, because available speech tools cannot produce its clicks and the click characters are removed before synthesis. No current system can voice ǀxam clicks, and this reflects the absence of training data rather than a configuration that could be adjusted. The components of a solution do, however, exist separately: an acoustic model of Nǀuu clicks validated against Ouma Katrina Esau's speech (Exter 2011), a published phonological reconstruction of ǀxam (du Plessis 2018), and the living phonetics of Nǀuu. What is missing is a programme that combines them. This paper has set out such a programme: community-led, consent-first, and phased from a low-cost glossary control to a grant-scale pronunciation model, so that the archive might eventually be heard as well as read.

The wider point is that the synthesis tooling now exists to give endangered and sleeping languages back their sound. Neural and formant synthesis, of the kind that already renders this archive aloud in English, is no longer the preserve of high-resource languages; applied with the right linguistic and community foundation, it offers a genuine opportunity to restore a pronunciation that no recording captured and no living person can any longer produce unaided. That opportunity is bounded, and the bounds matter: the realistic end state is accurate pronunciation reconstructed from living sister languages, not a synthetic voice narrating freely in ǀxam, for which no data exists. Within those bounds the approach, living phonetics plus linguistic reconstruction plus modern synthesis, is replicable for other languages documented in writing but lost in sound.

Acknowledgements

This work, like the reading interface it extends, builds on Pippa Skotnes's recovery of the Bleek–Lloyd archive: the Digital Bleek & Lloyd and the scholarship around it (Sound from the Thinking Strings, 1991; Claim to the Country, 2007; When the World Was, 2025, among others), without which the corpus would not be legible or reachable.

References

Speech technology and synthesis

• Casanova, E. et al. (2024). XTTS: A Massively Multilingual Zero-Shot Text-to-Speech Model. arXiv:2406.04904.
• Casanova, E. et al. (2022). YourTTS. ICML 2022 / arXiv:2112.02418.
• Chen, S. et al. (2024). VALL-E 2. arXiv:2406.05370. (and VALL-E, arXiv:2301.02111.)
• Pratap, V. et al. (2023). Scaling Speech Technology to 1,000+ Languages (Meta MMS). arXiv:2305.13516.
• Staib, M. et al. (2020). Phonological Features for 0-shot Multilingual Speech Synthesis. Interspeech 2020 / arXiv:2008.04107.
• Koffi, E. & Petzold, M. (2022). A Tutorial on Formant-Based Speech Synthesis for the Documentation of Critically Endangered Languages. Linguistic Portfolios 11(1), Art. 3.

Click phonetics

• Exter, M. (2011). The Acoustic Modeling of Click Types. Proc. ICPhS XVII, Hong Kong, 655–658.
• Miller, A. L., Brugman, J., Sands, B., Namaseb, L., Exter, M. & Collins, C. (2009). Differences in airstream and posterior place of articulation among Nǀuu clicks. JIPA 39(2), 129–161.
• Sands, B. (1991). Evidence for click features: Acoustic characteristics of Xhosa clicks. UCLA WPP 80, 6–37; and Sands, Click Consonants (Brill).

ǀxam, Nǀuu, classification and reconstruction

• Güldemann, T. (2005). Studies in Tuu (Southern Khoisan). University of Leipzig Papers on Africa No. 23.
• du Plessis, M. (2018). Kora: A Lost Khoisan Language of the early Cape and the Gariep. Unisa Press.
• ELAR deposits dk0089 (Güldemann & Witzlack-Makarevich) and dk0505 (Shah & Brenzinger).

The archive and its recovery (Pippa Skotnes)

• Skotnes, P. (2007). Claim to the Country: The Archive of Wilhelm Bleek and Lucy Lloyd. Jacana / Ohio University Press.
• Skotnes, P. (1991). Sound from the Thinking Strings. Axeage Private Press.
• Skotnes, P. (2025). When the World Was. Tafelberg.
• Skotnes, P. (gen. ed.). The Digital Bleek & Lloyd / ǃkhwe-ta ǃxōë.

Endangered-language speech-technology precedents (Section 4)

• Te Hiku Media and NVIDIA (2022). Māori speech AI (te reo Māori ASR: 300+ hours, ~92% reported accuracy; Kaitiakitanga Licence). NVIDIA Blog.
• University of Hawaiʻi at Hilo (2025). The Lauleo project (ʻŌlelo Hawaiʻi; with Te Hiku Media, Kanaeokana, Awaiaulu).
• Joanis, E., Knowles, R., Kuhn, R. et al. (2020). The Nunavut Hansard Inuktitut–English Parallel Corpus 3.0. LREC 2020.
• Foley, B. et al. (2018). Building Speech Recognition Systems for Language Documentation: the Elpis pipeline (CoEDL). SLTU.
• Ardila, R. et al. (2020). Common Voice: A Massively-Multilingual Speech Corpus. LREC 2020.

Governance

• South African San Council (2017). San Code of Research Ethics; and Schroeder, D., Chatfield, K., Singh, M., Chennells, R. and Herissone-Kelly, P. (2019). The San Code of Research Ethics, in Equitable Research Partnerships (Springer). See also Callaway, E. (2017), Nature 543, 475-476.
• United Nations (2007). Declaration on the Rights of Indigenous Peoples (UNDRIP); free, prior and informed consent, Articles 19 and 32.
• Carroll, S. R. et al. (2020). The CARE Principles for Indigenous Data Governance. Data Science Journal 19(1), 43.
• First Nations Information Governance Centre. The First Nations Principles of OCAP (Ownership, Control, Access, Possession).
• Kukutai, T. and Taylor, J. (eds) (2016). Indigenous Data Sovereignty: Toward an Agenda. ANU Press; and Te Mana Raraunga, the Māori Data Sovereignty Network.
• UNESCO Memory of the World: The Bleek Collection (inscribed 1997).

Appendix: Sourcing notes

Corrections and caveats established during research, reflected throughout:

1. Exter (2011) is an acoustic analysis, not a synthesiser; it supplies validated parameter targets for two click types (apical and laminal), derived from a deliberately rough articulatory model: useful seed data, not a turnkey table for the full Nǀuu/ǀxam click set.
2. The "seven click releases" sometimes attributed to ǀxam is a reconstruction inference (from comparison with Kora; du Plessis 2018), not a directly observed inventory.
3. Several speech-technology figures (e.g. ElevenLabs cloning thresholds, the YourTTS hour total) are search-surfaced quotes of official sources rather than directly re-fetched; treat as high-confidence, not exact.