Dataset Passport, illustrative specimen

NaraDatum multilingual conversational sample set

What is a NaraDatum Dataset Passport?

A Dataset Passport records what was reviewed about a dataset, what was established, and what was not. It states origin and collection context, technical specification, rights and consent, speakers and recording conditions, transcripts and QA, known limitations, and the verification scope, with each fact labelled by evidence class.

Evidence for your downstream obligations. A team building an AI system under the EU AI Act carries its own data-governance duties for that system. NaraDatum is a data supplier, not the provider of your model or system, and does not certify any model or system as compliant. What the Passport gives you is the dataset evidence those duties draw on: documented provenance and collection source, consent and rights evidence, dataset characteristics, representativeness and evaluation information where it was actually measured, traceability, and known limitations, in a form you can place into your own technical documentation.

This is an illustrative capability specimen. It shows the structure and evidence standard of a NaraDatum Dataset Passport, using material recorded internally as NaraDatum-recorded and NaraDatum-controlled. Those origin, ownership and control statements are internal records and were not independently established for this release.

It is not proof of current licensable inventory or availability. Availability is confirmed against each buyer requirement.
It is an illustrative capability specimen. Not offered as licensable production supply. This is demonstration-scale material.
It is limited to the evidence actually reviewed. Fields that were not established are shown as not established rather than omitted.

How to read the evidence states

Every state is written in words. The marker is a reading aid, carries no meaning on its own, and is not a certification.

Re-verified by NaraDatum: measured against the delivered files
Internally confirmed: checked against retained records
Internal record: not independently verified
Not established: shown rather than omitted

Origin and collection context

Origin category: Recorded internally as NaraDatum Direct Collection. Origin and ownership are internal records, not independently verified statements.
Collection route: Recorded in internal documentation as direct NaraDatum collection. Not independently verified for this website release.
Collection purpose: Demonstration and controlled sample evaluation.
Set contents: Six samples across three languages (Hindi, Indian English and Hinglish), each in a conversational and a single-speaker form.
Total duration: 934.64 seconds (15 minutes 35 seconds) across the six samples.
Speech type: Recorded internally as spontaneous conversational speech rather than scripted or read material. The reviewed audio retains interruptions, repetitions, fillers, self-corrections and code-switching where present. The spontaneous and unscripted origin description is an internal NaraDatum record and was not independently established for this website release.
Single-speaker form: Selected-speaker extracts from the same spontaneous conversations. These demonstrate single-speaker packaging and are not naturally recorded monologues.

Evidence classes differ across this table and are stated separately below. File integrity and technical format were re-verified by NaraDatum against the delivered files on 9 August 2026. Origin, collector, ownership and licensing authority were not independently verified: they are recorded in internal NaraDatum documentation. NaraDatum holds this specimen, so its own re-verification is reported as NaraDatum’s rather than as independent. Inspecting a delivered file can establish what the file is. It cannot establish who recorded it or who holds rights in it.

Technical specification

Measured directly from the delivered files, not transcribed from a supplier claim.

Container / codec: WAV, PCM signed 16-bit little-endian (pcm_s16le)
Sample rate: 16 000 Hz
Bit depth: 16-bit
Channels: 1 (mono)
Per-sample duration: 184.805 s · 119.300 s · 181.000 s · 119.690 s · 210.245 s · 119.600 s
Integrity: SHA-256 recorded per file. Source and output hashes match for all six files.
NaraDatum re-verification: NaraDatum re-hashed and re-probed all six files on 9 August 2026: 6 of 6 checksums matched the manifest, and codec, sample rate, bit depth, channel count and duration matched the recorded values in every case.
Processing: Audio was not altered, trimmed, resampled, denoised, normalised or re-encoded during packaging.

16 kHz, 16-bit, mono PCM WAV does not by itself establish telephony capture method, codec history, or suitability for a specific telephony model.

Rights and consent

Ownership: Recorded internally as NaraDatum-owned direct collection. This is an internal assertion by the accountable NaraDatum representative, not independently established for public licensing, and the underlying rights instrument was not reviewed for this website release.
Licensing control: Recorded internally as NaraDatum. Not independently established for public licensing.
Consent status: Internally confirmed by NaraDatum against retained records.
Uses contributors authorised NaraDatum to make: Commercial AI data use; commercialisation and licensing of the data; technical and commercial evaluation; controlled sample sharing with prospective customers.
Contributor identities: Not included in the package.
Consent documents: Not included in the package. Supporting records are maintained internally by NaraDatum.
Consent version and date: NOT RECORDED as a versioned consent artefact for this specimen. Supporting records are maintained internally by NaraDatum.
Prohibited or restricted uses: This specimen conveys no evaluation, training, fine-tuning, redistribution or production right. It is not licensed. Any such use requires a separate written agreement.
Territories or jurisdictions: NOT SPECIFIED for this specimen. Speaker origin is recorded as Rajasthan, India; no permitted-territory restriction is asserted or granted here.
Withdrawal process: NOT RECORDED for this specimen. For NaraDatum-managed production collection, contributor withdrawal and deletion are handled under the NaraDatum Collect consent framework, which suspends further processing on a non-rejected request.
Retention constraints: NOT RECORDED as a fixed retention term for this specimen.
Onward disclosure or licensing rights: None granted here. This specimen is published for illustration only and confers no onward disclosure or licensing right on any reader.
Biometric or sensitive-data classification: No voiceprint or biometric identifier was derived from this material; it is raw speech audio with transcripts. Whether a recording is treated as biometric or special-category data depends on the downstream purpose and the applicable jurisdiction: identification or speaker-verification use can bring it within those categories, plain content collection generally does not.
Minors status: No minors. All speakers are recorded as approximately 22–25 years of age. Exact individual ages are not recorded.
External legal review: Not carried out and not claimed.
Licence offered here: None. Commercial status: illustrative capability specimen, not offered as licensable production supply. This specimen is published for illustration. It conveys no evaluation, training, fine-tuning, redistribution or production right.

Underlying contributor consent documents are not included in this public specimen and were not independently re-reviewed for this website release. The permitted uses above describe what contributors authorised NaraDatum to do. They are not a grant of rights to any reader of this page. Production or model-training use of any NaraDatum material requires a separate written agreement.

Speakers and recording conditions

Speaker gender: Male. All speakers.
Speaker age: Approximately 22–25 years. Exact individual ages are not recorded.
Speaker origin: Rajasthan, India.
Broad regional label: North India, resting on confirmed speaker origin only.
Accent classification: No independent phonetic or accent-classification study has been completed.
Speaker count: Two per conversational sample; the demographic description applies to both.
Conversation structure: Two-speaker for the conversational samples. The single-speaker form is a selected-speaker extract from the same two-speaker recording, not a natively single-speaker capture. No three-or-more-speaker material is present in this set.
Channel structure: Single mixed channel. Both speakers share one mono channel. This was measured: see the technical specification above. The reviewed files are not channel-separated.
Speaker attribution: Present in the transcript as manually assigned speaker labels with timestamps, reviewed against the audio. No automated diarisation result is claimed.
Overlap marking: Not established. The transcript carries no overlap annotation.
Environment: Quiet indoor environment with low background noise.
Acoustic measurement: No measured ambient-noise level, signal-to-noise ratio or room acoustic measurement is available.
Recording device: Not recorded.
Studio / soundproofing: No claim is made in either direction.
Domain: General conversational speech. No specific industry or operational domain is assigned.

Transcripts, QA and PII

Transcript format: Timestamped speaker transcript, plain text. Hindi in Devanagari with spoken English terms preserved; Indian English in Latin script; Hinglish in Devanagari plus Latin-script English code-switching.
Transcript review: Manually reviewed against the matching audio for controlled sample sharing. Recorded complete by the accountable owner on 23 July 2026.
Transcript qualification: This does not constitute production acceptance or confirmation against any buyer specification.
Manual audio review: End-to-end listening completed on all six files. Playability confirmed.
Language labels: Manually confirmed.
Speaker structure: Manually reviewed. Single-speaker extracts were checked for residual opposite-speaker speech; none was identified in the final shared excerpts.
Manual PII listening: No obvious personally identifiable information was heard. This is not an absolute or forensic guarantee.
Automated PII scan: NaraDatum review dated 9 August 2026. All six transcripts scanned for phone numbers, email addresses, URLs, long digit strings, government-ID patterns, monetary amounts and dates: zero matches.
Automated scan limitation: The scan covers Latin-script and numeric patterns only. It would not detect a personal name written in Devanagari. It supplements manual listening and does not replace it.
QA basis: This is a current NaraDatum review, dated as above. It is not collection-time QA evidence, and no collection-time QA record is claimed for this material.

Known limitations

Sixteen. Published with the specimen rather than disclosed on request.

Short evaluation excerpts. Not confirmed production inventory.
Demonstration scale. 15 minutes 35 seconds in total, across six samples and three languages.
Not assessed against any buyer’s acceptance criteria.
Exact individual speaker ages are not recorded.
No independent phonetic or accent-classification study has been completed.
Recording-device details are not recorded.
No measured ambient-noise, signal-to-noise or room acoustic values are available.
No studio or soundproofing claim is made in either direction.
No specific industry or operational domain is assigned.
Any broad topic guidance used before recording is not recorded in the metadata.
Manual listening heard no obvious PII; this is not an absolute or forensic guarantee.
All speakers are male and from one state. This set is not demographically representative.
16 kHz mono does not establish telephony capture method or codec history.
No production volume, delivery timeline or commercial term is committed.
No external legal review of the rights position has been carried out.
Underlying contributor consent documents are not included in this public specimen and were not independently re-reviewed for this website release.

Verification scope

The distinction between what was confirmed, what NaraDatum re-verified against the files, and what was never established.

Ownership: Internal record, not independently verified
Consent: Internally confirmed by NaraDatum against retained records, not independently re-reviewed for this release
Direct collection: Internal record, not independently verified
Speaker demographics: Internal record, not independently verified
Speaker region: Internal record, not independently verified
Recording environment: Internal record, not independently verified
Audio technical properties: Re-measured by NaraDatum on 9 August 2026. Establishes format only
Audio integrity: Re-verified by NaraDatum on 9 August 2026, 6/6. Establishes file integrity only
Transcript vs audio: Manually reviewed by accountable owner, 23 July 2026
Transcript PII (automated): Scanned by NaraDatum on 9 August 2026, Latin/numeric patterns only
Accent classification: Not independently classified
Acoustic measurements: Not recorded
Recording device details: Not recorded
Collection-time QA record: Not established
External legal review: Not carried out

Delivery structure

sample_pack/
├── 00_README_FIRST.txt
├── 01_SAMPLE_MANIFEST.csv          one row per sample, with SHA-256
├── 02_KNOWN_LIMITATIONS.txt
├── 03_RIGHTS_AND_USAGE_NOTICE.txt
├── 01_Hindi/
│   ├── conversational/  audio/ transcript/ metadata/ qa_summary/
│   └── single_speaker/  audio/ transcript/ metadata/ qa_summary/
├── 02_Indian_English/  (same structure)
└── 03_Hinglish/        (same structure)

Every sample carries its own metadata file, QA summary and transcript. The manifest records the SHA-256 of each audio file so a recipient can verify integrity independently rather than trusting the packaging.

If this is the standard you need

Tell us the requirement. We review it against reviewed existing data first, then managed sourcing, and scope funded collection only where existing data cannot meet the brief.

Submit a requirement How verification works

This is the standard NaraDatum documents a dataset against, including the parts that stay unverified. If you hold human data and want it reviewed to this standard, see how supplying works.

Common questions

What does the published specimen cover?: Six samples across Hindi, Indian English and Hinglish, totalling 934.64 seconds. File integrity and technical format were re-verified by NaraDatum against the delivered files on 9 August 2026, with 6 of 6 checksums matching. Origin, collector, ownership and licensing authority are internal NaraDatum records and were not independently verified. Sixteen limitations are published with the specimen.
Can the specimen be licensed?: No. It is an illustrative capability specimen, published to show the structure and evidence standard of a Passport. It is demonstration-scale material, it is not offered as licensable production supply, and it conveys no evaluation, training, fine-tuning, redistribution or production right.