Project definition
Locale, domain, volume, file and metadata formats, consent requirements, and the acceptance criteria that decide whether a batch passes.
Specialist capability
Language data collected under conditions you can reproduce
Alongside our AI systems and content work, Codelab runs language and speech data projects end to end: definition, collection, transcription, annotation, an independent review pass and structured delivery. Consistency is what makes this data usable — the same protocol, the same conventions, applied across every item in a batch.
Lifecycle
What a data project involves
Not every project needs every stage — but every stage that runs is written down first.
Sourcing & coordination
Identifying and coordinating participants or sources that match the specification — regional variety, age band, gender balance or domain familiarity, as the project defines.
Collection
Capture under controlled, repeatable conditions, following the agreed prompts, protocol and quality thresholds.
Transcription & annotation
Transcription, segmentation and labelling to your guidelines, with a documented house style so output stays consistent across the batch.
Validation & QA
A separate review pass against the acceptance criteria. Failures are corrected at source and re-checked rather than annotated as exceptions.
Structured delivery
Agreed directory layout, naming convention, metadata and a short document describing what was produced, how, and any known limitations.
Collection
What we can collect
Scripted speech
Prompted readings from a supplied or jointly written script — useful when you need controlled phonetic coverage, specific vocabulary or comparable utterances across speakers.
Spontaneous & conversational speech
Unscripted responses, prompted conversation or task-based dialogue, when the model needs natural pacing, hesitation and turn-taking rather than read prose.
Participant coordination
Recruiting and scheduling speakers to match the specification, and keeping the composition of a batch consistent with what was agreed.
Capture environments
Controlled collection environments, including partner facilities where a project requires them. Codelab does not operate its own recording studio; when a project needs studio conditions we arrange access and state that plainly in the plan.
Transcription & annotation
Consistent conventions, applied across the batch
Transcription quality is mostly a question of whether the same decision is made the same way every time. We agree the conventions up front and document them.
- Verbatim or clean-read transcription, as specified
- Segmentation and timestamping to your interval rules
- Speaker labelling and turn boundaries
- Disfluency, noise and inaudible-region tagging
- Number, casing and punctuation conventions fixed in writing
- Domain vocabulary lists agreed before production
Quality assurance
A separate review step, not a self-check
The reviewer is not the person who produced the item. Review is measured against the written acceptance criteria, and corrections go back to source.
-
01
Pilot batch
A small first run to expose ambiguity in the guidelines before volume production begins.
-
02
Guideline lock
Conventions and edge cases are settled and written down after the pilot is reviewed together.
-
03
Production
Batch production against the locked guidelines, with questions logged rather than improvised.
-
04
Independent review
A second pass checks audio quality, transcription accuracy, tagging and metadata completeness.
-
05
Correction & sign-off
Failures are corrected at source, re-checked, then delivered with a note of what changed.
Languages
Regional capability
Codelab is based in Brazil and works with international clients. Brazilian Portuguese is the locale we run as a matter of course. Project communication, documentation and delivery notes are in English or Brazilian Portuguese, whichever you prefer.
- Brazilian Portuguese
- Other locales per project
- English or Portuguese communication
Additional Latin American locales are assessed per project rather than offered as standing coverage. Tell us the locale you need and we will say plainly whether we can source it properly, instead of accepting work we cannot deliver to standard.
Start a project
Need speech data collected, transcribed or checked?
Tell us the locale, the number of hours or utterances, and the transcription conventions you work to. If you are still deciding, describe the model and we will propose a specification.