AI & personal dataUpdated 22 November 20269 min read

Voice AI and Call Recordings Under the DPDP Act

Call recordings, transcripts, voice bots and voiceprints all process personal data. How to handle notice, consent, redaction, retention and vendors under DPDP.

By Suhas Bhairav · Reviewed against the DPDP Rules, 2025 (G.S.R. 846(E), 13 Nov 2025)

Is a call recording personal data?
Yes, if the speaker is identifiable
Notice
Rule 3 — itemised data, purposes, withdrawal link
Security baseline
Rule 6(1)(a)–(g)
Minimum retention
1 year for data and logs (Rule 8(3))
Children
Verifiable parental consent (Rule 10)
Applies from
13 May 2027

Quick answer

Under the DPDP Act, recording a call, transcribing it, running a voice bot or analysing speech is processing of personal data. From 13 May 2027 you need a lawful basis for each purpose, a clear notice at the start of the call, Rule 6 safeguards such as transcript redaction and access logs, at least one year of processing logs (Rule 8(3)), and processor contracts with speech vendors.

  • A voice recording, its transcript and any derived analytics are all personal data processing.
  • Tell callers at the start that the call is recorded and why; Rule 3 notices must itemise data and purposes.
  • Using recordings to train speech or language models is a separate purpose that needs its own basis.
  • Redact card numbers, ID numbers and similar identifiers from audio and transcripts before analytics.
  • Voiceprints used to identify callers are biometric processing; offer a non-biometric alternative.
  • Speech-to-text, voice bot and telephony vendors are your Data Processors.

Who this is for: Contact centres and BPOs, Voice AI and conversational AI startups, Banks, insurers and telecom customer care, Sales teams using call analytics, CX and quality leaders

Are call recordings personal data under the DPDP Act?

Yes. A recording of an identifiable person's voice is personal data, and so is everything derived from it: the transcript, the summary, sentiment and intent labels, quality scores, and any voiceprint. A voice bot that listens, transcribes and responds processes personal data continuously. Recording a call digitally, storing it and analysing it are all "processing" under the DPDP Act.

A single customer call can create several datasets:

ArtefactTypical contentsRisk
Audio fileVoice, background speech, spoken card or account numbersHard to search and redact
TranscriptFull text, including identifiers and health or financial detailsEasily copied into other systems
Summary / CRM noteCondensed facts about the personPersists longest
Analytics labelsSentiment, intent, churn risk, agent scoreInferences used for decisions
VoiceprintBiometric templateIdentifies the person across calls
Model training dataAudio and transcripts reused for ASR or LLM tuningPurpose creep

The DPDP Act has no "sensitive data" class, so health or financial details spoken on a call are protected like any other personal data. But sensitivity matters for risk assessment, penalties and, for large platforms, possible SDF designation.

Also remember the agent's voice: your employees and BPO staff are Data Principals too. Recording and scoring their calls is usually processing for employment purposes under section 7(i), and should be proportionate.

Notice and lawful basis for recorded calls

Notice at the start of the call

Rule 3 requires a notice that is understandable on its own, in clear and plain language, with an itemised description of personal data and the specified purposes, plus a link or other means to withdraw consent, exercise rights and complain to the Board. A spoken message cannot carry everything, so use layers:

  1. A short spoken disclosure at the start: the call is recorded, it may be handled or analysed with automated tools, the main purposes, and where to find full details.
  2. A full Rule 3 notice on your website or app, linked from the IVR menu, SMS or email follow-up.
  3. A way to object or withdraw — for example, "press 9 to speak without the call being used for training", where training relies on consent.

Lawful basis by purpose

  • Handling the caller's request — she called to get help; section 7(a) (voluntarily provided for a specified purpose) or consent for the service.
  • Quality and dispute evidence — often within the same service purpose if the notice says so; check any sector rules that require recordings.
  • Training ASR or LLM models — a separate purpose; generally requires specific consent.
  • Marketing analytics or upsell scoring — a separate purpose; requires consent and must respect withdrawal.

Do not make training consent a condition of getting support if it is not necessary for the support.

Engineering controls for voice pipelines

Redaction

  • Pause-and-resume recording or DTMF masking while callers speak or key card numbers, PINs and OTPs.
  • Transcript redaction — run PII detection on transcripts before they reach analytics, LLM summarisation or search; replace identifiers with tokens (Rule 6(1)(a)).
  • Audio redaction — bleep or silence the timestamps where identifiers were detected, if audio is kept for QA.

Access and logging

  • Restrict playback to roles that need it; log every playback and download with user, call ID and reason (Rule 6(1)(b)–(c)).
  • Watermark exported audio and block bulk downloads by default.

Storage

  • Encrypt audio and transcripts at rest, with separate keys per tenant for BPOs serving multiple clients.
  • Keep audio and transcripts in the region your contracts and sector rules require.

Voice bots

  • Disclose that the caller is speaking to an automated system, and offer escalation to a human.
  • Send only the minimum context to the language model; redact before prompts.
  • Keep conversation state separate from long-term customer profiles unless the notice covers it.

Children

If callers may be children — edtech helplines, gaming support — section 9 applies: verifiable parental consent for processing and no behavioural monitoring or targeted advertising.

How long to keep recordings and transcripts

Retention for voice data has three drivers:

  1. Purpose. Erase when the purpose is served or consent is withdrawn, unless law requires retention (section 8(7)).
  2. Minimum one year. Rule 8(3) requires every Data Fiduciary to retain personal data, associated traffic data and processing logs for at least one year from processing, for the Seventh Schedule purposes, then erase unless another law requires longer. Case 2 extends this to data held by your processor.
  3. Sector rules. Some regulators require call recordings for longer periods for specific transactions. Check the latest circular for your sector.

A practical tiering:

DataSuggested handling
Raw audioKeep for the longer of the service purpose, sector requirement or the Rule 8(3) minimum; restricted archive
Redacted transcriptSame period as audio, used for QA and disputes
Call metadata (CDR-like: time, duration, numbers, agent)At least one year as traffic data and logs
Analytics labelsOnly as long as the analytics purpose needs
Training copiesOnly with training basis; removed on withdrawal

Document the period for each artefact, automate deletion, and log deletions.

Voice biometrics and speaker identification

Voiceprint authentication ("my voice is my password") and speaker identification create biometric templates that identify a person across calls. The DPDP Act treats them as personal data without a special category, but they carry high risk: they cannot be changed if leaked.

Controls to apply:

  • Explicit, specific consent for enrolment, separate from consent to be recorded, with a clear alternative such as OTP or knowledge-based verification. Consent must not be a condition of service if the voiceprint is not necessary.
  • Template protection — store templates, not raw enrolment audio; encrypt them with keys held separately; prefer vendors that support template protection or on-premises matching.
  • Liveness and spoofing checks against replayed or synthetic voices.
  • Easy withdrawal — a menu option or app setting to delete the voiceprint, as easy as enrolment (Rule 3(c)(i)), followed by erasure at the vendor.
  • No covert identification — do not run speaker identification on callers who have not enrolled.

For facial and other biometrics, see facial recognition and biometrics. If you deploy emotion recognition on calls and serve EU users, Article 50 of the EU AI Act requires deployers to inform people exposed to it.

Vendors, BPOs and cross-border processing

A voice stack usually involves several processors: telephony or CCaaS platform, speech-to-text, text-to-speech, LLM provider, QA analytics and storage. Under section 8(2) and Rule 6(1)(f), each needs a contract covering security safeguards, breach notice to you, no use of recordings to train the vendor's own models without your authorisation, retention including the Rule 8(3) minimum, deletion and sub-processors.

Indian BPOs handling calls of customers outside India under a foreign client's contract may fall within section 17(1)(d), which disapplies most of the DPDP Act but keeps section 8(5) security safeguards. Calls involving Indian customers are fully in scope. CERT-In Directions also apply to Indian service providers, including 6-hour incident reporting and 180-day log retention in India. See IT services and BPOs.

Cross-border speech processing is allowed except to countries restricted by notification under section 16, subject to Rule 15 and stricter sector rules.

Further reading on suhasbhairav.com

Before 13 May 2027

Compliance checklist

  1. Inventory every voice artefact

    List audio, transcripts, summaries, labels, voiceprints and training copies with system, vendor and region.

    Section 8(1)

  2. Add a layered call-start disclosure

    Announce recording and automated analysis, and link to a full Rule 3 notice via IVR, SMS or app.

    Rule 3

  3. Separate training consent from service

    Ask specifically before using recordings to train speech or language models, and honour opt-outs.

    Section 6

  4. Mask card numbers, PINs and OTPs

    Use pause-and-resume or DTMF masking and redact identifiers from transcripts.

    Rule 6(1)(a)

  5. Restrict and log playback

    Limit recording access by role and log every playback and download.

    Rule 6(1)(b)–(c)

  6. Set retention per artefact

    Apply purpose, sector and Rule 8(3) one-year minimums, then automate deletion.

    Rule 8(3); Section 8(7)

  7. Get specific consent for voiceprints

    Offer a non-biometric alternative and make deletion as easy as enrolment.

    Section 6; Rule 3(c)(i)

  8. Contract all speech vendors

    Cover safeguards, breach notice, training restrictions, retention, deletion and sub-processors.

    Section 8(2); Rule 6(1)(f)

Key dates

  1. DPDP Rules notified

  2. Notice, safeguards, retention and breach duties apply to call recordings and voice AI

Answers

Frequently asked questions

Is it legal to record customer calls in India under DPDP?

Yes, with a lawful basis and a clear notice. Tell callers at the start that the call is recorded and why, link to a full Rule 3 notice, and apply security safeguards. Using recordings for model training or marketing analytics is a separate purpose needing its own basis, usually consent.

How long can I keep call recordings in India?

Keep them only as long as the purpose needs, but Rule 8(3) requires personal data, traffic data and processing logs to be retained for at least one year from processing. Sector rules may require longer. Document the period for each artefact and automate deletion.

Can I use call recordings to train AI models?

Only with a lawful basis for that purpose, which is generally specific consent, since training differs from handling the caller's request. Offer a way to opt out, redact identifiers before training, and exclude recordings of callers who withdraw consent from future training runs.

Do voice bots need to disclose they are AI?

The DPDP Act has no explicit AI disclosure rule, but your notice must describe automated processing clearly. Disclosing an automated system and offering a human is good practice, and the EU AI Act's Article 50 requires it for systems interacting with people in the EU, unless it is obvious.

Are voiceprints biometric data under DPDP?

The DPDP Act has no separate biometric or sensitive category, but voiceprints are personal data that identify a person and cannot be changed if leaked. Use specific consent, a non-biometric alternative, encrypted templates and easy deletion, and contract strictly with voice-biometric vendors.

Does DPDP apply to Indian BPOs recording foreign customers' calls?

Mostly not, if the callers are outside India and the processing is under a contract with a foreign client: section 17(1)(d) disapplies most obligations. Security safeguards under section 8(5) still apply, as do CERT-In Directions and the client's own law, such as GDPR.

Legal references and sources

This guide explains the law for engineering and business teams. It is not legal advice; confirm how it applies to your organisation with qualified counsel, and check the Gazette for later notifications or amendments.

About the author

Suhas Bhairav

Suhas Bhairav

Systems architect and applied AI practitioner who designs production AI systems for Indian and global companies, with a focus on privacy-by-design under the DPDP Act and the EU AI Act.