Skip to main content
June 16, 2026 · By Piyush Sahoo · Updated November 3, 2026 A call recording API lets your application start, stop, fetch, and delete call audio, and it announces finished files through webhooks (callbacks). The API is the easy part. A compliant pipeline also captures consent, keeps card data out, redacts transcripts, logs access, and erases on a schedule you can defend. This guide covers the engineering decisions in that order. It also maps the Indian rules that set how long you keep audio, including the SEBI and RBI changes that take effect by January 2027 and the DPDP duties due in May 2027.
Key takeaways
  • Record each party on its own channel. The channel then tells you who spoke, so you avoid guessing speakers from a mixed track. Check the default, because mono is often what you get unless you ask.
  • Treat consent, carve-outs, and erasure as pipeline stages. Each stage should leave evidence you can show an auditor later.
  • In India, retention has both a floor and a ceiling. RBI, SEBI, and DPDP Rule 8(3) can require you to keep records, while DPDP section 8(7) requires erasure once the purpose ends.
  • Card verification codes are the hard line. PCI guidance forbids keeping them after authorization, even encrypted, if the audio can be queried. Transcription makes audio queryable.
  • On Vobiz, four methods cover four call shapes. Recordings stay for 30 days, so any longer retention period needs a copy in your own storage.

What does a call recording API do?

A recording API has four jobs, and only the first happens during the call: The asynchronous step shapes the design. You never get the audio in the response that started the recording. You get an identifier, then a later event, so your pipeline should be a small state machine keyed on the recording ID rather than one blocking request. Provider terms differ, but the shape is common. Starting a recording and delivering its file are separate steps.

Should you record calls in mono or stereo?

Record in stereo unless storage cost truly outweighs everything downstream. A mono file mixes both voices into one track. A stereo file gives each party a channel, and that single choice decides what your analytics can measure. Illustrative waveforms of one call. In mono, the agent and customer share a single track and their overlapping speech merges. In stereo, each party has a separate track, so the overlap stays attributable and talk time is measurable. Illustrative waveforms of one call. In mono, the agent and customer share a single track and their overlapping speech merges. In stereo, each party has a separate track, so the overlap stays attributable and talk time is measurable.

Channels and diarization solve different problems

Deepgram’s comparison draws the line well. Multichannel transcription handles each audio channel on its own. Diarization segments audio by speaker identity, whatever the channel. With stereo you already know the speaker from the channel, so one source of error disappears. Stereo also gives you a compliance option. Because the voices sit in separate channels, you can apply different policies to each. You might keep the agent channel longer for coaching, and erase the customer channel sooner.

Three stereo traps

  • The default may be mono. Vobiz’s record_channel_type defaults to mono. Set this explicitly on any platform.
  • Conferences aren’t per-speaker. A room can hold more people than a file has channels, so some voices may share a channel. Don’t assume one channel per person on any platform.
  • Identical channels signal a file problem. If both channels transcribe to the same text, Deepgram points to joint stereo, where a space-saving conversion has mixed the channels. Test for this if you convert or compress files before transcribing.

MP3 or WAV?

Choose by what reads the file next. MP3 suits human review and long archives. WAV suits speech engines and forensic work, because no lossy step sits between the call and the model. Format also multiplies your storage, and retention multiplies it again. The chart assumes 100,000 recorded minutes a month. Illustrative bar chart of stored audio at 100,000 recorded minutes a month. One month holds 48 GB in MP3, 96 GB in mono WAV, and 192 GB in stereo WAV. Six months holds 288 GB, 576 GB, and 1.2 TB. Twelve months holds 576 GB, 1.2 TB, and 2.3 TB. Sixty months holds 2.9 TB, 5.8 TB, and 11.5 TB. Illustrative bar chart of stored audio at 100,000 recorded minutes a month. One month holds 48 GB in MP3, 96 GB in mono WAV, and 192 GB in stereo WAV. Six months holds 288 GB, 576 GB, and 1.2 TB. Twelve months holds 576 GB, 1.2 TB, and 2.3 TB. Sixty months holds 2.9 TB, 5.8 TB, and 11.5 TB. The arithmetic is bitrate times seconds. MP3 at 64 kbps gives 48 GB for 100,000 minutes. The Vobiz console guide gives 64 kbps for MP3 and 8 kHz mono PCM for WAV. We assumed 16-bit samples, which gives 128 kbps, and double that for stereo. A workable compromise is to transcribe from WAV, then keep a compressed copy for the long tail.

What makes a call recording pipeline compliant?

A pipeline is compliant when you can prove what happened to one specific call. That means every stage needs an output you could hand to an auditor, not just a setting you believe was on. Table-style diagram of seven pipeline stages and the proof each leaves: consent gate keeps notice version, response, and time; capture keeps recording ID and call UUID; carve-out keeps stop and restart timestamps; ingest keeps file hash and copy time; transcribe and redact keeps entity list and redaction version; control access keeps who opened it, when, and why; erase keeps a deletion record per recording. Table-style diagram of seven pipeline stages and the proof each leaves: consent gate keeps notice version, response, and time; capture keeps recording ID and call UUID; carve-out keeps stop and restart timestamps; ingest keeps file hash and copy time; transcribe and redact keeps entity list and redaction version; control access keeps who opened it, when, and why; erase keeps a deletion record per recording. Run this before capture, and let its result decide whether recording starts. Store the notice version with the response, because your script will change and you must know what each caller heard. Consent to call and consent to record are different questions. TRAI’s rules govern placing the call, and data protection law governs the recording. Vobiz’s UCC guide covers the calling side.

Ingest

Treat the webhook as a hint rather than as your source of truth. Verify the sender, respond quickly, and queue the work. Then fetch the file, hash it, copy it to your store, and record the result under the recording ID so a repeated event does nothing. Run a reconciliation job against the list API as well. It catches any recording whose webhook never arrived, and it matters most when the platform’s retention window is short.

Transcribe and redact

Redact before a transcript leaves your boundary. Once raw text has reached a third-party model, redaction can’t recall it. Decide scope using the identifiers your callers really speak. In India, that list should include Aadhaar, PAN, and card numbers. Measure misses on your own languages and accents, including calls that switch between languages. Our ASR guide explains how telephony audio affects accuracy. Derived data counts too. Treat sentiment tags, voiceprints, and lead scores as personal data with purpose limits of their own. So your erasure job has to reach transcripts, embeddings, and analytics tables as well as the audio.

Control access and erase

Index recordings by person, not only by call. An erasure request names a person, and your system has to find every file, transcript, and backup connected to that person. Numbers alone won’t always get you there, since people share and change phones. Log each retrieval with a reason. DPDP Rule 6 specifically asks for visibility into access through logs and monitoring, alongside encryption or masking and access control. Recording with clear notice, consent, and a lawful purpose is the defensible position for a business, but no single rule says so. We found no Indian statute that sets a one-party or all-party standard for business call recording. Instead, the position comes from privacy case law, the DPDP Act, and sector rules. Take legal advice for your own case.
  • Privacy. In PUCL v. Union of India, the Supreme Court called telephone tapping a serious invasion of privacy and protected private telephone conversation under Article 21. That case concerned state tapping, not business recording.
  • Evidence. In Vibhor Garg v. Neha (2025), the Supreme Court admitted a secretly recorded spousal call in a matrimonial dispute. Don’t read this as a licence, because it’s about admissibility and marital privilege.
  • Employees. BNB Legal reports that the Delhi High Court, in Sanjay Pandey v. Directorate of Enforcement, tied recording employees’ calls without consent to Article 21 privacy. So your notice should cover agents as well as customers. We read this summary, not the judgment.
  • Data protection. Under the DPDP Act, you process personal data on consent or a listed legitimate use (section 4). Consent has to be free, specific, informed, and shown by a clear affirmative action (section 6).
That last phrase has a practical consequence. Staying on the line is a weak sign of a clear affirmative action. A keypress or spoken “yes” gives you stronger evidence than an announcement alone. Whether section 7’s legitimate use for voluntarily provided data covers your flow is a question for counsel.

What the DPDP timeline means for you

The DPDP Rules, 2025 were notified in November 2025 and start in stages. Most operational rules begin 18 months after publication, which Shardul Amarchand Mangaldas reads as 14 May 2027. Purpose limits bite hardest on AI work. Consent for “quality and training” may not stretch to training a voice model, so name that purpose separately if you intend it.

How long should you keep call recordings?

Keep each recording for the longest period a rule actually requires of you, and no longer. That sounds simple until you see that Indian rules push from both sides. Three points need care:
  • SEBI’s rule isn’t a blanket audio mandate. Analysts may keep other verifiable records, and the circular says call recording isn’t needed where an interaction runs over a channel like email. It also makes recordings optional for institutional investors.
  • Rule 8(3) and section 8(7) pull against each other. Shardul Amarchand Mangaldas flags the one-year retention as a new duty that wasn’t in the draft. The rule’s text covers “personal data” and ties the duty to the purposes in its Seventh Schedule, but how it applies to audio after a consent withdrawal is unsettled, so get advice before you fix a schedule.
  • Check the RBI text yourself. The six-month detail above comes from published summaries, including Corplaw Updates. We confirmed the date from RBI’s press release but couldn’t open the directions themselves.
We also saw conflicting vendor claims about five- and seven-year retention across Indian financial regulators. We could not trace them to a regulator’s text, so we left them out. Check with your compliance team for your own licence. In practice, set a retention rule per purpose, and let each recording carry its purpose as metadata. A single global timer will be wrong for someone.

How do you keep card numbers out of call recordings?

Stop capture before card entry, so the data never reaches a file. The PCI Security Standards Council’s telephone payment supplement gives the reasoning. It’s from March 2011 and cites PCI DSS 2.0 numbering, so map it to the version you’re assessed against. The supplement separates two kinds of data:
  • Card verification codes (CVV, CVC, CID). You must not store these after authorization, even when encrypted, in audio that can be queried. Where technology can prevent recording them, it says to use it.
  • The card number (PAN). Storing it is not banned outright, but it has to be unreadable and justified by a business need.
The supplement’s list of query methods includes data mining and analysis tools. So if you transcribe recordings, assume they’re queryable. It also says encryption alone doesn’t make data non-queryable, and that vaulting audio is a last resort. Timeline of one call. File A records the conversation, then recording stops and is confirmed. The caller enters card details while nothing records. Recording starts again as file B. Both files share one call UUID, and the digits should stay out of webhook logs. Timeline of one call. File A records the conversation, then recording stops and is confirmed. The caller enters card details while nothing records. Recording starts again as file B. Both files share one call UUID, and the digits should stay out of webhook logs. Order matters here. Prompt for the card only after the stop call has succeeded, or a slow request will capture the first digits. Then listen to a sample of real gaps. Stopping the recording only protects the recording. If digits reach your server, your application and its logs now hold them. A validated payment service, or a separate flow that never touches your recording stack, shrinks that exposure. Ask your assessor which fits your scope. The strictest party on the call usually sets your default. If you can’t tell where everyone is, announce and capture consent for all of them.

How Vobiz handles call recording

Vobiz gives you the capture, webhook, retrieval, and deletion layer, and you own the policy around it. The right method depends on how the call reaches Vobiz. See the call recording overview for common use cases. Decision map. If the call runs over your own SIP trunk, use the trunk recording setting. If it is a conference room, use the Conference Record API. If you need stereo, a mid-call stop, or transcription, use the Record Calls API. If you want the whole call, use Record with recordSession. Otherwise use standalone Record for a single message. Decision map. If the call runs over your own SIP trunk, use the trunk recording setting. If it is a conference room, use the Conference Record API. If you need stereo, a mid-call stop, or transcription, use the Record Calls API. If you want the whole call, use Record with recordSession. Otherwise use standalone Record for a single message. Some details matter when you build on these: Calls, recording storage, and transcription can all carry charges, so check the Vobiz pricing page first. For India deployments, read the calling regulations as well.

How do you test a recording pipeline before launch?

Run these against your own account, with your own Auth ID and Auth Token and with test participants who have agreed to take part. Each call is real and billed, so confirm the numbers first.
  1. Decline consent. Confirm that no recording exists for the call.
  2. Talk for ten minutes. Confirm the file is ten minutes long, not one.
  3. Take your webhook down. Confirm that reconciliation still finds the file.
  4. Check the channels. Confirm that each party is alone on its channel, with little bleed.
  5. Enter a test card number. Confirm the digits appear in no file, transcript, or log.
  6. Seed known identifiers. Measure what redaction missed.
  7. File an erasure request. Confirm it removes the audio, transcript, derived data, and bucket copy.
The output of these tests is also your audit evidence, so keep it. For next steps, feed clean transcripts into post-call analytics, or use them to pass context during a call transfer.

Frequently asked questions

Use stereo for almost every business call. Each party gets a separate channel, so a transcript’s speaker labels come from the channel rather than from diarization. It also makes talk time and overlap measurable. Mono mainly wins on storage, so reserve it for cases where nobody will analyse the audio.
It depends on who you are. RBI’s recovery rules are reported as at least six months from January 2027, SEBI asks research analysts to keep interaction records for five years, and DPDP Rule 8(3) sets a one-year minimum from May 2027. Set a period per purpose, then erase on schedule. Confirm yours with counsel.
Card verification codes must not be kept after authorization in audio that can be queried, even when encrypted. A stored card number must be unreadable and justified. The practical answer is to stop recording before card entry, or use a validated payment service, then verify by listening to samples.
Use the Record Calls API and set record_channel_type to stereo, because the default is mono. Vobiz then places the caller and callee on the left and right channels. Set time_limit too, since it defaults to 60 seconds, and handle the completed file at your callback_url.

Further reading on Vobiz

Sources

Indian law and regulators Law firm and practitioner summaries Payment, US, and EU rules Vendor documentation and guides, read November 3, 2026

Record calls on Vobiz

Provision a number and start capturing stereo recordings you can transcribe and analyse.