Kaddu: Speech Translation Data Without a Text Intermediate

Kaddu (Wolof kàddu) means word or speech.

Kaddu is a test of a data collection method. The method is not tied to any one language, prompt language or region. Wolof is the first case because the team is based where it is spoken.

The problem

Every standard pipeline for building a speech dataset passes through written text in the target language at some point. Read-speech corpora need a validated sentence set for contributors to read aloud. Transcribed corpora need people who can write the language consistently after listening. Text-first translation pipelines need bilingual contributors who write the translation before anyone records it.

For a large share of the world's languages, that text step is the constraint. The pool of people who can write the language is small, the orthography is contested or unstandardised, and the writing or transcribing work is slow and skilled. Cost concentrates there, and that is where scaling stops. Existing resources for such languages tend to be small read-speech corpora built by specialist teams, which is consistent with the constraint rather than evidence against it.

The method

We call the method reverse elicitation. A contributor sees a short prompt in a high-resource language and speaks its meaning in the target language. The stored pair is target-language speech with source-language text. That is the supervision format for speech-to-text translation, a task that current multilingual speech models such as Whisper and SeamlessM4T support directly. The data can be used for fine-tuning without a target-language transcript ever existing.

What the method removes:

• The target-language text step, and with it the orthography requirement and the requirement that contributors be literate in the target language.
• The cost of reference text. Source-language sentences are effectively unlimited, so the reference is free and the cost moves entirely to contributor speaking time.

What it does not remove:

• Contributors must read the prompt language. The literacy requirement moves from the target language to the source language. In many contexts that is a much larger population, but it is not everyone.
• Validation. A second native speaker must still listen and confirm that the recording means what the prompt says. If that costs as much as transcription would have, the method saves nothing. This is measured, not assumed.
• The register problem. Prompted speech is careful and sentence-shaped. Models trained on it should not be expected to handle spontaneous speech, and no such claim will be made.
• Semantic drift. The contributor may paraphrase, drop a clause or misread. The reference is a target the speaker approximates, not ground truth. Whether the approximation is close enough on average to train on is the central empirical question.

Two things make the method worth testing rather than obvious. Pretrained multilingual speech models now make fine-tuning on tens of hours viable for a language the base model has not seen, which was not true when direct speech-to-translation learning was first studied. And the contributor requirement drops from "literate and bilingual" to "speaks the language and reads the prompt language", which should make crowdsourcing easier and per-item contributor effort lower. Both of those are hypotheses. The second is measured directly.

An optional second channel lets literate contributors later type what a recording says, in whatever spelling they use. The corpus is complete without it. It is measured as a separate condition, not assumed to help.

What is being tested

The Phase 1 protocol asks five questions:

1. Does a model fine-tuned on reverse-elicited data beat the untuned baseline on held-out speakers and held-out prompts?
2. What is the data-efficiency curve from 5 to 30 hours?
3. Does typed transcription add measurable signal over prompt-only supervision?
4. What is the contributor effort per usable minute of audio, and where does volunteer return rate collapse?
5. How much does performance drop on prompts from domains absent from training?

The primary outcome is human-rated meaning preservation, judged blind by native speakers. Automatic metrics are reported but not decisive. The primary cost metric is contributor effort: wall-clock minutes of contributor time per usable minute of audio, including retakes, rejections and upload. That unit is comparable across collection pipelines in a way that money is not.

A decision rule is fixed before collection. It separates two questions that are usually conflated: whether the data trains a usable model, and whether volunteer collection of that data scales. Three outcomes are possible. If the model works and collection scales, the same protocol is applied to further languages. If the model works but collection does not scale, the method is sound and the incentive design is not. If the model does not work at 30 hours, the learning curve says whether the problem is volume or the method, and the negative result is published.

How the project runs

The project is organised in phases.

• Phase 0, Foundations. The conceptual argument, the frozen protocol, the collection app, the data platform, and the consent and data protection framework.
• Phase 1, Wolof feasibility. A preregistered study: 30 usable hours target, three supervision conditions, two model families, evaluation on held-out speakers and held-out prompts, and effort measured throughout.
• Phase 2. Decided by Phase 1 against the rule above.

The protocol is frozen and hashed before any data is collected, and the hash is committed publicly. Hyperparameters are fixed on a development split before the main runs. Any analysis not in the protocol is reported as exploratory. All results are published, including null results. Datasets, model checkpoints and software are released openly.

Where it stands

As of 26 September 2026, Phase 0 is active. The conceptual paper and the feasibility protocol each have a first draft in outputs/. Six items in the protocol remain open before it can be frozen: the effort ceiling, the comparator model, the repurposed language token, the collection end date, the dataset licence, and the validator recruitment source. The app, the data platform and the consent policy have not been started. Whether the study requires formal ethics review is an open question to be settled before freeze.

Who

Kaddu is a project of Baseline Institute (Baseline Lab), The Gambia. Mario Gomez leads it. The study needs native Wolof speakers as validators, evaluators and reviewers of the protocol and prompt set, and those roles are not yet filled.

Contributing

Phase 1 will need volunteers in several roles: speakers who record responses to prompts, validators who check recordings against prompts, optional transcribers, and reviewers of the prompt set for cultural fit. Contributors must be 18 or over in Phase 1.

All contribution is voluntary. Terms, credit, data use and withdrawal follow Baseline Institute policy. The consent text and data protection policy for this study will be published before the campaign opens, and no recording will be collected before a contributor has seen them.