Short answer. Speech data for low-resource languages is collected rather than found. There is no large public corpus to scrape, so the work is field operations: recruit and verify native speakers stratified by dialect, age, gender and region; design prompts that elicit both read and spontaneous speech; record across the device and acoustic conditions the model will actually meet; transcribe using a written convention agreed in advance for a language that may have unstable orthography; and verify with native reviewers measuring word error rate and agreement per dialect rather than per language. Consent and fair compensation are part of the method, not a compliance afterthought — a corpus with weak consent provenance is unusable at exactly the moment it becomes valuable.
Speech models fail on the languages with the least data, and those are the languages spoken by the people least served by the technology. Closing that gap is a collection problem with a specific methodology, and most of the difficulty is operational rather than technical.
This guide sets out how the work is actually done and what an enterprise buyer should require.
Why low-resource speech is different
There is nothing to scrape. High-resource languages have decades of broadcast, subtitled media and public corpora. Low-resource languages frequently have none of that in usable quantity, so every hour is produced deliberately.
Orthography may be unstable. Some languages have competing spelling conventions, recent standardisation, or predominantly oral use. Transcription cannot begin until the convention is chosen and written down — and that decision affects every downstream metric.
Dialect distance is often larger than in high-resource languages. Varieties that a language-level plan treats as one thing may differ enough that a model trained on one performs poorly on another.
Code-switching is normal, not exceptional. In many multilingual regions, speakers mix languages within a sentence as standard practice. A corpus that excludes code-switched speech excludes how people actually talk.
The speaker pool is a recruiting problem. Finding, verifying, retaining and fairly paying speakers in a language you do not currently cover is field work — and it is the single largest determinant of whether a programme delivers.
Designing coverage before recording
Coverage is designed, not discovered. Five stratification axes, each with a target:
| Axis | Specify | Why |
|---|---|---|
| Dialect and regional variety | Named varieties with a target share each | The most common gap; invisible in a language-level plan |
| Speaker demographics | Age bands, gender, urban/rural, education mix | Determines who the model works badly for |
| Speech type | Read, scripted, spontaneous, conversational | Read-only corpora produce models that fail on natural speech |
| Acoustic condition | Quiet, ambient noise, outdoor, in-vehicle, telephony | Must match deployment conditions |
| Device | Phone models common in the market, headset, far-field | Microphone characteristics shape the signal materially |
Coverage ratio = Hours collected in stratum ÷ Hours targeted in stratum
Track and report the minimum across strata alongside the total. A programme at 90% of total hours can still be at 20% on a dialect that carries a third of the market.
Prompt and script design
Three material types, and a corpus generally needs all three:
- Read speech — speakers read prepared sentences. Cheapest, most controllable, gives clean phonetic coverage. Produces speech that does not sound like conversation.
- Elicited spontaneous speech — a prompt or scenario, spoken in the speaker's own words. More expensive to transcribe, far closer to deployment reality.
- Conversational speech — two or more speakers, with overlaps, interruptions and repairs. Hardest to transcribe, essential for anything handling real dialogue.
Two design rules that matter more in low-resource work:
Write prompts in the target language, culturally. Prompts translated from English produce sentences native speakers would never say, and speakers read them awkwardly, which is audible in the data.
Design for phonetic coverage explicitly. Check that the prompt set exercises the language's phoneme inventory, including sounds that are rare but distinctive. In a language without existing corpora, nothing will do this for you.
Recruiting and verifying speakers
The operational core of the work.
- Verify the variety, not just the language. A speaker who says they speak the language may speak a different regional variety than the one you need. Verification is a short native-speaker screening, and skipping it corrupts the stratification silently.
- Recruit in-market. Diaspora speakers are valuable and their speech drifts — vocabulary, borrowings and register shift away from current in-market usage.
- Plan for retention across sessions where the design needs multiple recordings per speaker.
- Pay fairly and transparently, at rates appropriate to the local market, with terms the speaker can read in their own language. This is an ethical requirement and also a practical one: underpaid collection produces rushed sessions, drop-out, and a speaker pool that will not return for the next programme.
Recording conditions
Match the deployment. A model trained entirely on quiet-room recordings degrades in the environments where it will actually run.
- Include the noise conditions of the use case — street, market, vehicle, home with background sound.
- Include the devices people actually use in that market, which may be different from the devices your team uses.
- Include telephony-band audio if the model will meet it; bandwidth reduction is not something a wideband-trained model handles by default.
- Record metadata per session — device, environment, speaker stratum, session conditions. Missing metadata is why a corpus cannot be re-balanced later.
Transcription conventions
The decision that governs every quality figure afterwards. Agree in writing, before the first transcription:
- Orthography. Which spelling convention, and what happens to words with no standard spelling.
- Numerals, dates and abbreviations — written as spoken or as digits.
- Disfluencies — filled pauses, false starts, repetitions: transcribed, tagged, or dropped.
- Code-switching — how mixed-language spans are marked.
- Non-speech events — laughter, noise, overlapping speech.
- Speaker labels for multi-party audio.
- Uncertainty — how a transcriber marks unintelligible audio rather than guessing.
That last one is worth insisting on. A convention with no way to say "I could not hear this" produces confident wrong transcriptions, which are more damaging than gaps because they train the model on invented text.
Quality verification
WER = (Substitutions + Insertions + Deletions) ÷ Reference words
Four requirements:
- Report WER per dialect and per acoustic condition, not per language. An aggregate figure is dominated by the easiest stratum.
- Measure inter-transcriber agreement on a double-transcribed subset. Low agreement means the convention is ambiguous — the earliest and cheapest signal that the guideline needs work.
- Use a gold set built natively in the language, refreshed periodically and injected into live work at a defined rate, so quality is measured continuously rather than at delivery.
- Have native reviewers listen, not only read. Some errors — wrong dialect, an unnatural reading, a speaker who is not a native speaker — are inaudible in the transcript and obvious in the audio.
Consent, ethics and provenance
For collected speech this is the difference between an asset and a liability.
- Informed consent in the speaker's own language, covering what the recording will be used for, including onward and commercial use, and how long it is retained.
- Withdrawal path that can actually be executed — a right to withdraw that cannot be traced to specific files is not a right.
- PII handling. Spontaneous speech contains names, addresses and personal detail that the speaker did not intend to be permanent. Detection and redaction should be part of the pipeline.
- Per-item provenance — session, speaker stratum, consent record, transcriber and reviewer.
- Community respect where a language belongs to a small or marginalised community. Fair compensation, transparency about use, and — where appropriate — the community's own access to the resulting resource.
How Lifewood approaches this
Low-resource language speech is a specialism rather than an extension of Lifewood's general speech work, and it is delivered as field collection with in-market operations: bespoke data collection including image, video and audio recordings across diverse geographic and demographic segments, with multilingual transcription and phonetic labelling behind it.
The reason the model works is structural. 40+ delivery centres across 30+ countries, spanning China, the Philippines, Malaysia, India and Bangladesh alongside Africa, Europe and North America, with 50+ languages and 56,788 contributors, means speakers are recruited and verified in-market rather than sourced remotely — which is the step that decides whether dialect stratification is real or nominal. Lifewood has run multilingual data operations since 2004, with the current AI-data company established in 2018; speech and language engagements span voice-AI developers.
See low-resource speech data, multilingual data collection, AI data services and global AI data.
Sources and further reading
- Companion guides: Multilingual LLM Training Data and 9 Criteria for Choosing AI Annotation Services.
- Lifewood low-resource speech scope is published at lifewood.com/low-resource-speech-data.