Collecting low-resource dialects at scale

Most low-resource dialects were never written down at scale, let alone recorded. There is no corpus to scrape — the only honest path is to collect from native speakers, in the field, with consent and a quality process. Here is how we do it.

Start with a sampling matrix, not a target hour count

Before recording anything we define the matrix: which dialects, which age and gender strata, which regional accents and recording environments. Speaker recruitment is sampled against that matrix so the corpus is balanced and auditable — not skewed toward whoever was easiest to record.

Capture scripted and spontaneous speech

Dialect surfaces differently in read prompts and in spontaneous talk. We collect both — scripted prompts, spontaneous speech, scenario-based dialogues and prompt-based responses — and tag each by content type.

Metadata is the whole game

Every record carries language and dialect, speaker demographics, device type, environment category and location. Without that metadata you can't balance, audit or debug the dataset later.

Validate in stages

Automated quality checks, manual review, script-adherence and noise/clarity assessment, and metadata verification. Failed recordings are flagged for correction or replacement before anything ships.