TWB Voice is the rarest thing in AI right now. A ten-year humanitarian field network, a 200,000-linguist volunteer base, and a live two-tier licensing posture — all sitting on languages the frontier labs admit they can't serve. This site is the thinking I'd bring on day one. I've done the first twenty thousand dollars of it for free.
For ten years, low-resource languages were a moral argument. Now they're a margin argument. Meta's own MMS paper reports a character error rate of 2.9 across 363 African languages, versus 1.5 to 1.7everywhere else. Only 91% of African languages clear the CER ≤ 5 bar. Everywhere else it's 96 to 99%. That's not a rounding error. That's a published receipt from a frontier lab saying its best model doesn't work for a continent.
Every lab racing to ship voice-native LLMs has the same hole in its roadmap. None of them has a field network. None of them has a decade of trust with the communities whose voices they need. That's the asset TWB Voice has been compounding since 2015, and no amount of capital buys it on a 12-month timeline.
TWB Voice 1.0 shipped 111.45 hours of approved audio across Hausa, Kanuri, and Shuwa Arabic, with 135 contributors mobilised for a single Northeast Nigeria pilot. Yoruba and Igbo adapters shipped through NaijaVoices in July 2025. Gamayun crossed 622,961 translation segments across 15 languages including Rohingya and Tigrinya. This isn't a pitch deck. It's a working platform with 622,961 receipts.
Mozilla moved Common Voice to a gated Data Collective in October 2025. That's the market agreeing, out loud, that two-tier licensing is the future of speech data. TWB Voice arrived at that position from a different starting point and has a cleaner claim to it.
“Humanitarian trust at the input. Genuine low-resource specialism in the middle. Gated commercial licensing at the output. The African-and-crisis-language analogue of AI4Bharat — with a commercial flank Mozilla hasn’t built and can’t easily build.”
The speech-data market splits cleanly on two axes: licensing posture, and mission. Most players cluster in two corners. TWB Voice is almost alone in the third. That's not a coincidence. That's a moat.
Crisis-affected communities will never hand their voices to a for-profit. Frontier labs will never build a 200,000-person volunteer network in a decade they don't have. The quadrant is structurally defensible. Hold it.
200,000 linguists across the TWB Community, 135 contributors mobilised for the inaugural Northeast Nigeria pilot, ten years of humanitarian field presence. Nobody in the commercial speech-data world has ever produced this asset, and nobody is trying. It cannot be bought. It has to be earned, and TWB has been earning it since 2015.
Hausa, Kanuri, Shuwa Arabic, Rohingya, Tigrinya. Not the languages Google or Meta are solving in the next roadmap cycle. Synthetic ASR pipelines already sit at 1.1M combined rows for Hausa, 658k for Luo, 375k for Chichewa. Depth, not breadth, is the game — and depth is what TWB Voice already ships.
The NaijaVoices deal shipped Yoruba and Igbo w2v-bert-2.0 adapters in July 2025. That's not a hypothesis. That's a live commercial collaboration pattern that converts linguist hours into shippable adapters on a timeline frontier labs can plan around. Repeat the pattern eighteen times and you have a business.
Crisis-language ASR and TTS depth. Gated commercial licensing to frontier labs. Infrastructure-partner deals where Lelapa, Sama, or Appen pay to use the platform rather than compete with it. Foundation anchors that renew on measurable outcome metrics.
Building a generic commercial ASR API. Competing with Lelapa on telecoms and fintech contact centres. Building a generic African TTS voice-clone product. Any move that trades the humanitarian moat for commercial volume the market has already commoditised.
TWB Voice wasn't built for a market. It was built for a mission. That's its strength and its awkwardness. The platform fits the humanitarian segment like a glove because that's where it was forged. The academic fit is plausible with small tweaks. The commercial fit is a repositioning job, not a product overhaul. Treating all three as one audience is how good products end up stuck.
Below I've mapped each segment by what they're actually trying to do, what they'll pay, where TWB Voice already wins, and where the gaps sit.
Collect ethically-sourced, community-consented voice data in low-resource languages to power downstream services, hotlines, chatbots, and accountability mechanisms for crisis-affected populations.
Budgets are grant-constrained and cost-reimbursable. Willingness to pay is anchored to what the donor funds, not the value delivered. Revenue comes from paid collection services and funded pilots, not SaaS subscriptions.
Ethical collection frameworks. Community consent workflows. Field-tested across three under-resourced languages in northeast Nigeria. Credibility that no commercial vendor can shortcut, because trust here is earned in years, not quarters.
A packaged collection-as-a-service offer with transparent per-hour pricing. Off-the-shelf data-sharing agreements. Tighter integration with the MEAL tools humanitarian orgs already live inside.
Generate or access well-annotated corpora in low-resource languages for NLP research, ASR/TTS benchmarking, and peer-reviewed publication. Researchers need reproducibility, provenance, and licensing clarity more than polish.
Labs pay, but in grant-funded chunks, not enterprise contracts. Revenue here is more about strategic positioning, citations, co-authors, advisory board seats, than top-line.
Transparent ethical provenance, which matters enormously at review-committee level. Languages nobody else has. A credible publishing partner brand if you position it right.
Standard dataset cards in Hugging Face and LDC format. DOIs. A research-tier license that's cheap or free for accredited academic use. Machine-readable metadata. Without these, academic buyers drift to Common Voice by default.
Source training data for multilingual voice products in frontier markets, or build speech features for banking, agri-tech, telco, and BPO where English is the second or third language. Enterprise buyers need volume, SLA, and a legal footing that doesn't explode under due diligence.
Enterprise buyers pay $5 to $15 per hour of curated, annotated speech in a rare language, sometimes significantly more for contracted collection with specific demographic or domain parameters. But they pay nothing if the paperwork, QA reports, and delivery cadence don't look like a commercial vendor.
A genuinely differentiated supply. The data does not exist elsewhere at comparable quality. That's a moat, priced correctly.
Commercial pricing tiers. A master services agreement template. Enterprise-grade IP, indemnity, and data lineage guarantees. A sales motion that doesn't rely on a product lead doing founder-led selling between grant reports.
Humanitarian (services-led, grant-funded). Academic (free-to-cheap licensing with attribution). Commercial (dataset licensing plus contracted collection at market rates). Same platform, three price cards, three contract templates, three BD motions.
Today it's a project. Tomorrow it's a SKU: 40 hours of annotated conversational speech in language X, delivered in 90 days, at $Y per hour. That turns every humanitarian deployment into a repeatable commercial offer.
Ethically collected, community-consented, audit-ready is not a nice-to-have in 2026. It's a procurement requirement for any AI buyer who has read an EU AI Act summary. That's TWB Voice's unfair advantage. Price it in.
Five levers drive this business. Move any of them and the three-year revenue picture recalculates. Defaults are the base case. The aggressive case assumes one frontier lab deal closes. The conservative case assumes none do. Both break even.
Grant-dominant. PJMF plus one smaller co-funder. No frontier deal.
Balanced. Grants plus commercial licensing plus fine-tuning. Default case.
Frontier lab anchor deal Y2. Commercial flywheel compounds.
A roadmap that looks clever on a slide and dies in practice is worse than no roadmap at all. Here's the sequence I'd run, phased so each stage de-risks the next. Nothing in Phase 2 starts before Phase 1 has produced evidence. Nothing in Phase 4 gets pitched without Phase 3 numbers in hand.
Get the commercial chassis in place. Stop selling the mission, start selling the product.
Product lead (existing). Fractional commercial lead or senior BD (8 to 12 days/month). Legal counsel for templates (one-off).
Over-indexing on strategy decks and under-indexing on closed deals. Every deliverable tested against a real prospect conversation within 30 days of draft.
Prove the model works with 5 to 10 paying organisations on the books and a repeatable commercial dataset sale.
BD capacity scales to 1 FTE equivalent. Data engineering support to harden the annotation pipeline. Modest marketing spend (events, targeted content, not ads).
Commercial sales cycles are long. First enterprise close may slip. Maintain humanitarian services pipeline as revenue floor through the period.
Twenty-plus paying orgs, automated collection, 15+ languages live.
Formal team structure. Tooling spend on annotation, QA, and pipeline automation. Regional community coordinators for new languages.
Quality degradation under scale. QA metrics baked into the pipeline from Phase 1, not retrofitted at Phase 3.
Graduate from project funding to institutional capital.
CEO-level capacity for fundraising. CFO or financial advisor support for diligence readiness. Clean financial model updated monthly.
Capital-market timing. Treat institutional capital as an option, not a plan. The operation should be viable without it.
This is the Year 1 and Year 2 pipeline as I'd build it on day one. Ranked by strategic fit, not reachability. Some are funders, some are buyers, some are infrastructure partners. A few are all three. Every one of them has a reason to take the meeting this quarter.
You've published on multilingual benchmarks and have the research team for it. TWB Voice gives you the only humanitarian-grade gated corpus for African crisis languages in existence.
Command A multilingual expansion is live. You need training data for languages Meta hasn't fixed yet.
Your own model card admits the low-resource gap. We can close it for the thirty languages you'll never collect yourself.
Whisper v4 roadmap needs non-English evaluation data more than compute.
Your CER 2.9 result is the single best piece of market intelligence we have. Let us ship you the long tail MMS v2 won't reach.
You've already validated the gap publicly. The procurement conversation starts from yes.
The $364k anchor was phase one. Phase two is a multi-year renewal tied to commercial-licensing outcome metrics.
Your 2026 AI-for-good portfolio review. Bring them a success story with a revenue line.
You ship product. We ship infrastructure. Vulavula runs better on TWB Voice adapters than on anything you'd build in-house.
Post-Seed growth phase, burn rate matters, buy is cheaper than build.
Data Collective is the posture. TWB Voice is the content. Co-market the two and we both win the narrative.
Mozilla's post-October-2025 positioning needs case studies. We are the clearest one.
Maternal-health voice bots in Hausa and Kanuri need ASR that works. We're the only source that can ship this in 18 months.
Global Health discovery portfolio funds exactly this in 2026-2027.
Yoruba and Igbo shipped in July 2025. Let's repeat the template for six more Nigerian languages over the next 18 months.
Relationship warm, delivery proven, next tranche just needs scope and signature.
Co-authored benchmarks, shared evaluation infrastructure, credit to the community. Cheap to build, expensive not to.
Masakhane's governance maturation opens the door to institutional partnerships it couldn't hold a year ago.
Not a customer. A cross-learning partner. You solved the Indic version of our problem. Trade playbooks.
Their model is the one we want to run for Africa. The call is free and overdue.
Alvin is already on the team. Formalise the Makerere bridge as a joint research track on East African languages.
Funder narrative loves named university partnerships. Low cost, high signalling value.
Canadian development funder with an active AI-for-development portfolio. Natural multi-year co-funder with PJMF.
IDRC's 2026 strategic plan names responsible AI as a priority theme.
Claude's multilingual roadmap needs evaluation data, not just training data. We can build the benchmark.
Claude multilingual performance is a known gap and a 2026 priority.
European sovereign-AI positioning needs African languages more than it needs one more English benchmark.
Voxtral is live, the roadmap is open, European AI Act framing aligns with humanitarian licensing.
You charge $0.15 to $0.35 per hour. You can't serve Kanuri. We can. White-label the long tail.
Enterprise customers asking for language coverage you don't have is a 2026 sales blocker.
Same pitch as AssemblyAI, different sales cycle. $0.46 per hour economics work better if you don't have to build the data yourself.
Nova-3 language expansion roadmap is public and obvious.
You do paid annotation at scale. We do volunteer collection in communities you can't reach. Complementary, not competitive.
Sama's post-2023 ethical-sourcing repositioning needs partners that reinforce the narrative.
Same logic as Sama. We're the field-collection layer your commercial clients keep asking about.
Appen's Q3 earnings keep flagging low-resource language demand as a growth segment they can't fulfil.
Dataset hub hosting, model hub adapters, co-branded evaluation leaderboard for African ASR. Distribution, not dollars.
HF's Africa strategy is in build phase. Flagship partners are still being chosen.
You've spent 200 years getting into low-resource languages. We can help your audio-Scripture pipelines work on the same budget.
Biblica's digital-first 2026-2030 strategy has a line item for voice AI. The brief practically writes itself.
Three strategies, each internally coherent, each fundable on its own. The base case mixes all three. If forced to pick one, I'd pick Commercial Gated, because the moat compounds fastest. But the point of three scenarios is that no single one has to be right.
Blended $50 per hour. Per-language fine-tuning packaged as adapter-as-a-service at $45k. One seven-figure frontier-lab anchor deal in Year 2 changes the whole financial picture. The moat is the gated corpus plus the consented-collection pipeline, neither of which a competitor can clone on a venture timeline. Three-year revenue $4.5M, mostly unrestricted. This is the real business.
Every serious pitch has five or six objections that sink unserious ones. Here are the ones I'd expect, and the answers I'd give in the room.
No. CLEAR Global is a mission-locked 501(c)(3) with EIN 27-3840123 and an Irish charity registration. Volunteer contribution is opt-in, consented, and any commercial licensing flows back to source-community revenue share and mission delivery. The governance looks nothing like a for-profit annotation vendor, and the 2015-to-present track record is the receipt. Commercial revenue funds more language coverage, not shareholder returns.
Two answers. First, Common Voice is read-speech only, crowdsourced without LDC-grade metadata depth, and until October 2025 was fully open. TWB Voice 1.0 ships humanitarian-collected conversational and read data with contributor-level metadata and a gated licensing posture from day one. Second, and more importantly, Mozilla's own October 2025 migration to the Data Collective validates the direction we were already travelling. They're catching up, not pulling ahead.
Meta shipped v1. The published CER gap held. Meta has no plan to send field teams to Maiduguri to collect Kanuri, no plan to build Shuwa Arabic from 12 hours to 200, no plan to touch Rohingya or Tigrinya at all. The long tail is structurally outside MMS's economic model. Every time Meta improves its head-of-distribution model, the demand for clean long-tail data goes up, not down. Meta becomes the customer, not the threat.
Lelapa is a product company. Vulavula is a contact-centre and fintech voice platform for South African commercial markets. TWB Voice is infrastructure for data collection, curation, and licensing. The two businesses don't overlap, and Lelapa is a natural Year 1 target client, not a competitor. I'd expect the first conversation with Jade Abbott inside 30 days of starting.
Only the aggressive scenario does. Base case breaks even in Year 2 on a mix of grant renewal, custom collection, and fine-tuning revenue, with no seven-figure licensing deal required. Conservative case still clears $1.8M by Year 3 on grants alone. The frontier lab deal is the upside, not the pin holding the plan together. Read the model.
The timeline is tight but achievable because three of the five objectives share underlying research. Doing market intelligence and product-market fit assessment in parallel surfaces strategic client mapping for free. Here's how the weeks actually run.
Kick-off with the TWB Voice Product Lead. Read everything: the April market research report, user research outputs, co-design workshop notes, existing financials. Start 8 to 12 stakeholder interviews across CLEAR Global staff, existing pilot partners, and external voice-AI players. Begin market intelligence synthesis in parallel with PMF segment analysis. Single-source Notion page updated live so the commissioning manager always sees current state.
Market intelligence and PMF findings converge into a positioning hypothesis. Start building the dynamic financial model in Google Sheets: unit economics, revenue scenarios, cost structure, sensitivity. First draft client and partner registry from the market scan. Mid-point review with Product Lead at end of Week 4. This is the checkpoint where directional calls get made.
Translate the strategy into phases, milestones, and resource asks. Pricing scenarios stress-tested against the financial model. Commercial and humanitarian client registries refined with tailored value propositions. Second stakeholder check-in.
Write the business plan as a decision-ready document, not a strategy essay. Build the investor/donor deck in parallel, pulling directly from the financial model so numbers always match. Circulate penultimate drafts to the Product Lead for markup.
Incorporate feedback. Final review session. Walkthrough of the financial model so the internal team can drive it without me. Clean handover package.
Written end-of-week summary every Friday, visible to the Product Lead and anyone they want to loop in.
45-minute fortnightly call, plus kick-off and final review.
Live Notion workspace with all drafts, interview notes, and the working financial model link. No version-controlled Word docs flying around email.
Here's the number, here's what's in it, and here's how I'd prefer to be paid. Same terms you published. No surprises.
Senior strategy consulting rates in London sit roughly £800 to £1,500 per day for work of this scope and seniority. I'm priced to reflect the work, not the logo. You're paying for a strategy operator with 14 years of B2B demand generation, go-to-market, and commercial model-building experience, delivered solo without a pyramid of associates to pass the work down to. That's the trade. Fewer layers, less overhead, higher intensity.
Happy to extend into July if scope naturally grows. Happy to scale down if priorities shift mid-engagement. The contract should serve the work, not the other way around. No pre-approved incidentals anticipated; the engagement is remote-first by design.
Advance within 30 days of contract execution, as per CLEAR Global's stated terms.
Interim payments released as each of the four deliverables lands: business plan, model, registry, pitch deck.
Final instalment on Commissioning Manager approval of the full package.
Matches the structure published in CLEAR Global's terms of reference. No renegotiation needed.
The brief asks for four outputs. Rather than describe them in the abstract, I've started building each one. This site is the first. The rest are queued behind it.
The strategic through-line. How TWB Voice moves from grant-dependent to commercially defensible while protecting social impact.
A Google Sheets model your team can drive after I'm gone. Change an assumption, watch the forecast respond.
A prioritised list of prospective clients, partners, and funders with tailored value propositions for each.
A presentation optimised for impact investors and bilateral donors. Model-linked, narrative-clean.