Is voice cloning legal in India? What the courts and the law say in 2026

Is AI voice cloning legal in India?

AI voice cloning can be legal in India if you have the speaker's consent, but no statute says so either way. No Indian law bans voice cloning as such, and none gives anyone a general right to clone another person's voice. The answer has to be pieced together from court orders on personality rights, from copyright law, and from the criminal and IT laws.

Put together, those sources give a fairly simple working rule. Cloning a person's voice with their clear, written consent, for a use they agreed to, is low risk. Cloning a recognisable person's voice for commercial use without consent is exactly what the courts have stopped, again and again, since 2022. Using any cloned voice to deceive someone into handing over money or information can be a criminal offence.

A sound-alike is not a safe way round the rule. The courts have restrained tools that imitate a celebrity's voice, vocal style and singing technique, not only exact copies, so using a famous voice in any recognisable form needs a written licence. The rest of this guide explains where each part of the rule comes from, and it is a starting point, not legal advice.

How Indian courts came to protect the voice

Most of that working rule comes from judges, not from Parliament. Their tool is personality rights, which protect the parts of a person's identity that others could exploit, such as their name, image and likeness. No Indian statute sets them out. Judges have built them from privacy, from passing off (the rule that stops one business presenting its goods as someone else's) and from publicity principles.

Since 2022 the High Courts have repeatedly named the voice as one of the attributes these rights protect, so a voice is covered in the same way as a face or a name. Read in date order, the orders show how quickly the issue has moved. The 2022 Amitabh Bachchan order dealt with fake lottery and WhatsApp scams. By 2024 courts were restraining AI chatbots and AI voice models, and the 2026 Sonakshi Sinha order, as reported, expressly covers "voice-cloned audio".

The Arijit Singh order and AI voice models

One order in that list deals most directly with the tools businesses now use. On 26 July 2024 the Bombay High Court stopped AI platforms from letting users turn any audio into singer Arijit Singh's voice. The order records that one platform ran 'AI Voice Models' using his name and photograph, where users could paste in any YouTube link or audio file and get it back in his AI voice.

The court treated his voice, vocal style and singing technique as part of his personality, alongside his name and image. It said that "making AI tools available that enable the conversion of any voice into that of a celebrity without his/her permission constitutes a violation of the celebrity's personality rights" (para 18). The restraint covers voice models, voice conversion tools, synthesised voices, deepfakes and similar technology.

This was an ex parte ad-interim order, a temporary order made before the platforms were heard. It was also the earliest order we found aimed squarely at AI voice-model and voice-conversion tools, and trade-mark and IP commentators have described it as the first AI voice-cloning decision in India. The 2023 Anil Kapoor and May 2024 Jackie Shroff orders had already covered AI misuse of a voice, though, so "first" depends on how narrowly you define the tool. The Asha Bhosle order of 29 September 2025 followed the same path, against an AI platform offering songs 'in her voice'.

Copyright, performer's rights and the training data

Personality rights protect a voice as part of someone's identity. Copyright law works differently, and it does not protect a voice from cloning directly, because it protects works and performances, not the sound of a voice itself. But a singer or voice artist holds a performer's right in each performance they give, and that right lasts 50 years (Copyright Act, 1957, section 38).

The performer has the exclusive right to record, reproduce, copy and communicate the performance (section 38A). They also have moral rights. These are the right to be named as the performer, and the right to stop any "distortion, mutilation or other modification" of the performance that would harm their reputation (section 38B).

This is why the training data matters as much as the voice. Training a clone on someone's recorded performances, or remixing those recordings, can engage these rights even where the voice itself is not protected. The Asha Bhosle suit relied on her section 38B moral rights, and the interim order, as reported, restrains breaches of them. The practical rule is to train only on recordings the speaker made for you, never on commercial songs, films or adverts, which carry other people's performer's rights.

The 2026 deepfake rules and the criminal law

There is no standalone deepfake act in India. Two sets of statutes sit alongside the court orders instead: a 2026 amendment to the IT Rules that reaches online platforms, and criminal offences that reach anyone.

What a voice cloning consent form should cover

All of this comes back to consent. If you clone an employee's, a founder's or a voice artist's voice, a written release is your main protection. Every court order above turns on the lack of consent, so the release should leave no doubt about what the person agreed to, and it should be signed before any training audio is recorded.

Data protection law points the same way. On our reading, a voice recording that identifies a person is personal data under the Digital Personal Data Protection Act, 2023 (DPDP Act), which expects consent to be free, specific, informed, unconditional and unambiguous (section 6(1)). For an employee that matters most at the exit: the release should say what happens to the model when they leave. The guide to what the DPDP Act means for voice AI covers the notice and erasure duties that start on 13 May 2027.

Some details depend on the tool. On Pāṇini TTS, for example, a voice can be cloned from a 10-second recording, and a cloned voice carries across languages, so a release for a business clone should name the languages as well as the purposes. A release that covers the points below meets both the courts' focus on consent and the DPDP standard:

• Who and what: the speaker's full name, the legal entity that may use the voice, the recordings used for training, and that a synthetic voice model will be built from them.
• Use: the purposes allowed (for example, outbound service calls in named languages) and those excluded (adverts, political content, endorsements).
• Disclosure: whether listeners will be told the voice is AI-generated, and the exact wording.
• Term, territory and pay: how long and where the model may be used, and any fees or royalties, especially for professional voice artists with performer's rights.
• Withdrawal and exit: how the person can withdraw consent, and what happens to the model and data when they do or when they leave the company.
• Security and deletion: who can access the model, a promise not to share or sell it, and deletion of the training audio and model at the end of the term.
• Moral rights: no use that distorts the person's performance or harms their reputation.

Running a cloned voice once it is live

Consent is the start, not the end. Telling listeners at the start of each call that they are hearing an AI-generated voice is the habit that keeps a lawful clone from becoming a deceptive one. A line such as "This message uses an AI-generated version of [Name]'s voice, created with their permission, on behalf of [Company]" does the job, and the AI disclosure scripts have versions in several languages.

The model itself needs protecting, because a leaked clone can be used to impersonate the speaker in scam calls. Limit who can use it, log every use, and never share or sell it. Keep the release, the training recordings and each model version on file while the voice is in use. When the speaker withdraws consent, leaves the company or the term ends, delete the model and training audio and record the date.

A published contact for reports of a misused cloned voice, with quick takedowns and a record of each complaint, rounds this off. The chart shows the whole sequence, from agreeing the use to deleting the model.

Key dates