aidataservices.inAI data collection · India

Quality

Quality standards and acceptance criteria for speech data

The measurable acceptance criteria we contract against: audio thresholds, transcription accuracy, quota compliance and what happens when a batch fails.

Request a dataset quoteReply within one working day
Audio QC engineer inspecting waveforms and spectrograms — Quality standards and acceptance criteria for speech data
Audio
SNR and clipping thresholds
Transcripts
Accuracy measured, not asserted
Quotas
Enforced at intake
Failures
Re-collected, not edited
01

The short version

  • Acceptance criteria are written into the statement of work before recording, so a batch either passes or it does not
  • Audio is checked for signal-to-noise ratio, clipping, DC offset, dropouts and channel integrity
  • Transcription accuracy is measured by blind re-transcription of a random sample and reported as a word-level agreement figure
  • Quota compliance is reported per batch against the agreed cohort design
  • Failed material is re-collected at our cost; delivered files are never repaired by editing, because repairs leave artefacts that survive into your model
Provenance survives an external audit of your training dataSpeaker screenedIdentity + demographicsWritten consentAI training use, namedSpeaker ID issuedConsent bound to IDManifest + auditEvery file traceableNo scraped, resold or re-licensed third-party audio enters a delivery.DPDP Act 2023 handling, mapped to GDPR obligations for EU buyers.
02

Why it matters in practice

The point of a written acceptance standard is that quality stops being a negotiation. Either the batch meets the numbers in the statement of work or it is re-collected, and both parties can tell which from the QA report.

The same report is what lets you diagnose a model problem later: if word error rate is high on one dialect, the quota and agreement figures tell you whether the data or the model is at fault.

Two speakers recording natural conversational speech data — supporting quality standards and acceptance criteria for speech data
Two speakers recording natural conversational speech data
03

How it is operated

CriterionTypical threshold
Signal-to-noise ratio≥ 25 dB studio, ≥ 15 dB field unless the spec says otherwise
ClippingZero clipped samples in accepted files
Transcription agreement≥ 98% word-level on blind re-transcription
Quota complianceWithin 5% of every agreed cohort cell
Metadata completeness100% of required fields populated
Consent coverage100% of files mapped to a signed consent record
04

What you receive

  • The relevant policy or template as a document, not a claim on a web page
  • Programme-specific artefacts delivered with the corpus manifest
  • Named contact for audit questions during and after the programme
  • Written confirmation at close-out

If your legal or procurement team has a questionnaire, send it with the specification. It is faster to answer it once, up front, than to unblock a signed programme later.

Frequently asked

What happens if a batch fails acceptance?

It is re-collected at our cost against the same specification. You are invoiced only for delivered units that pass.

How is transcription accuracy measured?

A random sample is blind re-transcribed by an independent reviewer and compared at word level. The figure is published in the QA report per batch.

Can we set our own thresholds?

Yes, and you should. The thresholds above are defaults; whatever is in the statement of work is what is enforced.

Do you allow noisy data?

Where deployment is noisy, deliberately noisy capture is better training data. The specification sets the target noise profile rather than assuming studio silence.

Related pages

Send your compliance questionnaire with the spec

We answer procurement, legal and security questionnaires alongside the technical scope, in the same working day where we can.

Request a dataset quote