Turnaround
How long asr training data takes
Honest delivery schedules for asr training data, the stages that actually consume the calendar, and the three levers that safely compress a timeline.
- Standard
- Transcription adds roughly 1-2 weeks per 100 hours after recording, per language.
- Pilot
- 10-20 units in 1-2 weeks
- First batch
- Typically week 3
- Cadence
- Weekly batches after the first
Where the time actually goes
Recording is rarely the bottleneck. Specification agreement, speaker recruitment against quotas and native-reviewer QA are what set the calendar, and all three run in parallel once the spec is signed.
- Style guide authored per language, covering numerals, loanwords, script and disfluency rules
- Transcriber calibration round with inter-annotator agreement measurement
- First-pass transcription
- Second-pass native review
- Automated consistency checks against the style guide
- Split generation with speaker-disjoint partitions
Indicative schedules
| Volume | Spec + setup | Collection | QA + delivery | Total |
|---|---|---|---|---|
| Pilot (10-20 units) | 3-5 days | 5-7 days | 3-4 days | 1-2 weeks |
| 100 units | 1 week | 2-4 weeks | 1 week | 4-6 weeks |
| 500 units | 1-2 weeks | 4-7 weeks | 1-2 weeks | 6-10 weeks |
| 1,000 units | 2 weeks | 7-11 weeks | 2 weeks | 10-16 weeks |
Batches are delivered as they pass QA, so you start training long before the final unit lands.
What safely compresses a timeline
- Parallel cities: the same cohort recruited across several studios at once rather than sequentially
- A signed specification on day one — most slipped schedules start as an unsettled spec
- Accepting rolling batches instead of a single final delivery
- Widening one non-critical quota (usually education band or district) while holding the ones that matter
What cannot be rushed
Recruitment of narrow cohorts, native review depth and consent administration have floors. Compressing them is how vendors end up delivering data that fails acceptance, which costs more calendar time than it saved.
Agreement is measured, not assumed. We report word-level agreement on a held-out sample so you can judge label quality before training on it.
What lands with each batch
- Audio plus aligned transcripts (JSON/TSV, or your schema)
- Style guide as delivered documentation
- Inter-annotator agreement report
- Speaker-disjoint train/dev/test splits
Frequently asked
What is a realistic turnaround for asr datasets?
Transcription adds roughly 1-2 weeks per 100 hours after recording, per language. for a standard production run, with a pilot possible in one to two weeks and first batches usually arriving in week three.
Can you hit a hard deadline?
Often, by running more cities in parallel and delivering rolling batches. We will say no rather than accept a date that forces a quality compromise.
Do we wait for everything before we can train?
No. Batches are delivered as they pass QA so training can start early and the spec can be adjusted while collection is still running.
What causes delays in practice?
Late specification sign-off, mid-project quota changes and narrow cohorts in low-density districts. All three are visible in the plan before we start.
Related pages
Need asr datasets by a date?
Send the deadline with the volume. You get a schedule that says what is achievable and what is not.