Speech data you can put in production — and defend

At a glance

Consent	Every contributor accepts a versioned Contributor Agreement in plain language before their first recording. Acceptance and version are stored against the account.
Rights	Contributors assign the rights in their recordings to LootBeez, so we can grant customers a clear licence under our Data Licence Terms.
Privacy	Customers receive a pseudonymous speaker ID and broad demographics only. Names, emails, phone numbers, UPI IDs and PAN never leave LootBeez.
Quality	Training gate → automatic signal checks → speech-to-text check → human review by a native speaker who cannot review their own work.
Vetted speakers	Recording is invitation-only: every contributor account is individually approved by our team before it can submit any audio.
Fraud controls	Exact-duplicate audio detection, per-speaker caps, review sampling and account-level enforcement.
Infrastructure	Cloudflare global network; data encrypted in transit (TLS) and at rest; role-based access to admin tools.
Law	Designed around India's Digital Personal Data Protection Act, 2023 and the IT Rules, 2021, with a named Grievance Officer.

1. Informed consent

Consent is the foundation of every dataset we deliver. Before anyone can record:

• They confirm they are 18 or older. We do not knowingly accept minors as contributors.
• They read and accept the Contributor Agreement & Consent. It explains, in simple language, what is collected, who receives it, what customers may and may not do with it, how payment works, and how to withdraw. The app is available in Hindi and English.
• Acceptance is versioned. We record which version each contributor accepted. If the agreement changes materially, contributors must accept the new version before recording again.
• Withdrawal is real. Contributors can withdraw consent and request erasure at any time. Recordings not yet delivered are removed from future deliveries; the effect on data already licensed is explained honestly in the agreement.

2. Privacy by design

We collect only what we need to pay people and to describe the data. When a dataset is exported for a customer:

• Each speaker is represented by a pseudonymous ID generated with a salted one-way hash. The same person keeps the same ID within a dataset, so you can split train/test by speaker, but the ID cannot be reversed to an account.
• Demographics are broad: age band (not date of birth), gender, state, district and dialect.
• Never shared: name, email, phone number, UPI ID, PAN, exact age, IP address or device identifiers.
• Prompts are written to avoid collecting personal information. Contributors are told not to say their name, address, phone number or anyone else's personal details while recording, and reviewers reject recordings that do.

Full details are in our Privacy Policy.

3. Security

• Encryption. All traffic is served over HTTPS with HSTS. Audio and database storage are encrypted at rest by our infrastructure provider.
• Passwordless sign-in. Users log in with a one-time code sent to their email. Codes expire quickly, are stored only as hashes, are limited in number of attempts, and are rate-limited per email and per network address.
• Least-privilege access. Contributors see only their own work. Reviewers hear recordings only through the review queue, without contact details of the speaker. Admin tools are restricted to named administrators and are excluded from search engines and from being embedded in other sites.
• Hardened web app. A strict Content Security Policy, no third-party trackers or advertising scripts, and no audio is ever served from a public link — every playback request is checked.
• Customer delivery. Datasets are delivered through a private project link or a secure transfer method agreed in the Order. Customers can track progress on a private dashboard without seeing any contributor identity.

4. Quality you can measure

1. Training gate. New contributors must pass a short training quiz on the recording rules before they can work on paid projects.
2. Automatic signal checks on every clip: loudness, silence ratio, clipping and duration limits. Clips that fail are flagged before a human ever hears them.
3. Speech-to-text check. An automatic transcript is compared with the prompt to catch wrong, missing or mismatched speech.
4. Human review. A trained native-speaker reviewer approves or rejects each item against written guidelines. No one can review their own recordings.
5. Duplicate and fraud detection. Identical audio uploaded more than once is caught automatically. Per-speaker caps keep datasets diverse.
6. Acceptance window. Customers can report items that fail the agreed specification, and we replace or credit them — see the Refund & Cancellation Policy.

Every delivery ships with a metadata.csv and a README describing speaker counts, gender and age mix, districts covered, recording method and QC process.

5. Fair treatment of contributors

Ethically sourced data means the people behind it are treated fairly:

• Joining is free. We never charge contributors any fee.
• The rate for each project is shown before work begins, and only approved work is paid — so expectations are clear.
• Payouts are made weekly by UPI, with the UTR reference shown in the app and sent by email.
• Contributors can dispute a review decision, and every complaint has a named owner and a response timeline under our Grievance Redressal process.

See the Payment & Payout Policy and Community & Quality Guidelines.

6. Responsible use

Customers may use LootBeez data to build, train and evaluate speech and language technology. Under our licence they may not try to identify contributors, clone or imitate an individual contributor's voice, or use the data for surveillance or biometric identification. We may decline projects that conflict with these principles.

7. For procurement and legal teams

We are happy to support vendor onboarding. On request we can provide:

• Company registration details and tax information
• A mutual NDA before sharing samples or specifications
• Our Contributor Agreement and consent records for a delivered dataset
• A completed vendor or security questionnaire
• A free sample of 2–5 hours so your team can check quality first-hand

Write to business@lootbeez.shop or request a sample. To report a security or privacy concern, email privacy@lootbeez.shop.