How Speech Data Is Collected for Low-Resource Languages

High-resource languages have decades of broadcast, subtitled media and public corpora to draw on. Low-resource languages frequently have none of that in usable quantity. A low-resource language is one without enough existing digitised text, audio or annotated data for a model to learn it well from public sources alone, which means collection has to substitute for scraping entirely.

Orthography can also be unstable: some languages have competing spelling conventions, recent standardisation, or predominantly oral use, and transcription cannot begin until a convention is chosen and written down — a decision that affects every downstream quality metric. Dialect distance is often larger than in high-resource languages, so varieties a language-level plan treats as one thing may differ enough that a model trained on one performs poorly on another. Code-switching, where speakers mix languages within a single sentence, is normal rather than exceptional in many multilingual regions, and a corpus that excludes it excludes how people actually talk. Underlying all of this is a recruiting problem: finding, verifying, retaining and fairly paying speakers in a language a team does not currently cover is field work, and it is the single largest determinant of whether a programme delivers, a challenge covered in more depth when recruiting native contributors for African language data.

Coverage ratio is hours collected in a stratum divided by hours targeted in that stratum. The minimum across strata should be tracked and reported alongside the total, since a programme at 90% of total hours can still sit at 20% on a dialect that carries a third of the market. The comparison between high-resource and low-resource languages in AI training sets out why this planning step matters more the further a language sits from the high-resource end of that spectrum.

Read speech, where speakers read prepared sentences, is the cheapest and most controllable option and gives clean phonetic coverage, but it does not sound like conversation. Elicited spontaneous speech — a prompt or scenario spoken in the speaker's own words — is more expensive to transcribe and far closer to deployment reality. Conversational speech, with two or more speakers, overlaps, interruptions and repairs, is hardest to transcribe and essential for anything handling real dialogue. Prompts should be written in the target language culturally rather than translated from English, since translated prompts produce sentences native speakers would never say and are audible as awkward reading in the resulting data. The prompt set should also be checked explicitly for phonetic coverage — including sounds that are rare but distinctive — since in a language without existing corpora nothing else will surface that gap.

Verifying the variety rather than only the language matters because a speaker who says they speak the language may use a different regional variety than the one targeted, and skipping that short screening corrupts the stratification silently. Diaspora speakers are valuable contributors in general but their speech drifts over time — vocabulary, borrowings and register shift away from current in-market usage — so recruiting in-market is preferred where the deployment is also in-market. Retention across multiple sessions needs planning where a design calls for repeat recordings per speaker. Fair, transparent pay at rates appropriate to the local market, with terms the speaker can read in their own language, is both an ethical requirement and a practical one: underpaid collection produces rushed sessions, drop-out, and a speaker pool that will not return for the next programme, a point developed further in how data contributors should be consented and paid.

A model trained entirely on quiet-room recordings degrades in the environments where it will actually run, so a corpus should include the noise conditions of the use case — street, market, vehicle, home with background sound — and the devices people actually use in that market, which may differ from the devices a project team uses. Telephony-band audio should be included if the model will meet it, since bandwidth reduction is not something a wideband-trained model handles by default. Metadata should be recorded per session — device, environment, speaker stratum, session conditions — because missing metadata is the reason a corpus cannot be re-balanced later. Related considerations for speech captured outside studio conditions are covered in collecting accented and non-native speech for robust voice AI.

The convention needs to state which spelling standard applies and what happens to words with no standard spelling, whether numerals and dates are written as spoken or as digits, and how filled pauses, false starts and repetitions are handled — transcribed, tagged, or dropped. It should specify how mixed-language spans are marked, how laughter, noise and overlapping speech are recorded, and how speakers are labelled in multi-party audio. The convention also needs a defined way for a transcriber to mark unintelligible audio rather than guess, because a convention with no way to say "I could not hear this" produces confident wrong transcriptions — more damaging than gaps, since they train the model on invented text. Diarization and timestamping conventions for multi-speaker audio are covered in more detail in speech and audio annotation: transcription, diarization and timestamping.

WER is calculated as substitutions plus insertions plus deletions, divided by the number of reference words. Reporting it per dialect and per acoustic condition matters because an aggregate, language-level figure is dominated by the easiest stratum and hides weak performance elsewhere. Measuring inter-transcriber agreement on a subset that two transcribers each work independently gives the earliest and cheapest signal that a convention is ambiguous and needs revision. A gold set built natively in the language, refreshed periodically and injected into live work at a defined rate, keeps quality measured continuously rather than only at delivery, and native reviewers should listen to audio rather than only reading transcripts, since a wrong dialect, an unnatural reading, or a non-native speaker is inaudible on the page and obvious in the recording. The programme mechanics behind this verification loop are set out in how a speech data collection programme actually runs.

Consent needs to cover what the recording will be used for — including onward and commercial use — and how long it is retained, and a right to withdraw is only real if it can be traced to specific files and actually executed. Spontaneous speech routinely contains names, addresses and personal detail the speaker did not intend to make permanent, so detection and redaction need to sit inside the pipeline rather than as a later clean-up step. Per-item provenance — session, speaker stratum, consent record, transcriber and reviewer — should be retained for every file. 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 are part of doing the work responsibly.

The operating model depends on a physical footprint of 40+ delivery centres across 30+ countries, spanning China, the Philippines, Malaysia, India and Bangladesh alongside Africa, Europe and North America, supporting 100+ languages with 56,000+ registered contributors. That footprint is what makes in-market recruitment and verification possible rather than nominal, which is the step that decides whether dialect stratification in a plan is real once recording starts. Lifewood has run multilingual data operations since its founding in 2004, with speech and language engagements spanning voice-AI developers; buyers evaluating a partner for this specifically should also review multilingual data collection and broader AI data services scope alongside the speech specialism itself.