Hindi voice agents and IVR on AWS with Amazon Polly's Generative engine: what launched, where it runs, and the data questions for Indian firms

The launch added 14 new voices across nine locales. Hindi and Finnish became generally available on the Generative engine for the first time, with Kiara and Arjun as the new Hindi and Indian English voices.

The AWS Machine Learning blog, published on 22 October 2026, describes it as a drop-in upgrade. AWS says voice names and API calls are unchanged, so existing applications continue to work as they do today. AWS also says the upgrade gives "improved voice quality and accuracy at the same price", with "no new SKU". Standard and Neural voices are unaffected.

AWS lists Kiara and Arjun as fully bilingual voices, alongside Aditi and Kajal. A bilingual voice can switch between Hindi and Indian English within one sentence. Hindi can be written in Devanagari or in Romanagari, which is Hindi in the Latin alphabet, and mixed with English. By default, Arabic numerals are read in Indian English. To have them read in Hindi, pass the hi-IN language code.

The table below reflects the Amazon Polly Generative voices documentation, Neural voices and Standard voices pages as of 29 October 2026. AWS updates these lists often. In 2026, the Amazon Polly document history records Generative voices added in Europe (Zurich) on 28 May, Asia Pacific (Sydney) on 12 August and AWS GovCloud (US-West) on 22 October. AWS recommends calling DescribeVoices, or checking the console, to confirm a voice in your Region at deploy time.

Of the Asia Pacific Regions that offer Generative voices, only Singapore and Sydney are listed for bidirectional streaming as of 29 October 2026. We also found no AWS announcement of a date for Generative voices in Mumbai or Hyderabad by that date.

When your application calls Polly in Singapore, the text you send is processed there. The Amazon Polly FAQ says content processed by Polly is encrypted and stored at rest in the AWS Region where you use the service.

There is a second, separate flow. The Amazon Polly FAQ says Polly "may store and use text inputs processed by the service" to provide and maintain the service and to improve Polly and other Amazon AI technologies. The AWS Organizations documentation adds that AWS AI services "might store such content in an AWS Region outside of the AWS Region where you are using the service" for service improvement. You can stop that use with an AI services opt-out policy, set for Polly alone or for all supported AI services.

Indian law adds two layers.

• Section 16(1) of the Digital Personal Data Protection Act, 2023 lets the Central Government, by notification, restrict the transfer of personal data by a Data Fiduciary for processing to a country or territory outside India. Section 16(2) keeps in force any other Indian law that gives a higher degree of protection for, or restriction on, transfer of personal data outside India. Before you rely on a Region outside India, check with your Data Protection Officer whether a notification under section 16 covers it.
• Sector rules can be stricter. For example, the Reserve Bank of India's circular of 6 April 2018, Storage of Payment System Data, states that "All system providers shall ensure that the entire data relating to payment systems operated by them are stored in a system only in India."

Whether a voice prompt falls under either rule depends on what it contains and on your business, so ask your compliance team. For the wider picture, read our guides on AWS data residency under DPDP and on Bedrock and in-country inference in India.

1. Keep everything in Mumbai. Use the Neural voice Kajal, for Hindi and Indian English, or the Standard voices in ap-south-1 when the text must not leave India. You give up the Generative voices, but the text you send to Polly is processed in Mumbai. With the opt-out applied, the Amazon Polly FAQ says your content is not stored in another Region.
2. Pre-render static prompts. Generate fixed IVR menus and greetings with Kiara or Arjun once, store the audio in your own Region and replay it. AWS says you can cache and replay synthesized static prompts at no additional cost. The text sent to the Generative Region is only the fixed script. Dynamic content, such as names or amounts, uses a Neural voice in Mumbai, or contains no personal data.
3. Use Generative voices from Singapore for live voice agents. For a voice agent built on a large language model, use the Generative engine with bidirectional streaming, StartSpeechSynthesisStream, in Singapore. Keep personal data in the spoken text to a minimum, apply the opt-out policy and record the cross-border flow for your Data Protection Officer.

Pattern 2 has one trade-off. AWS notes that updates to the Generative model can slightly change how a voice sounds, so prompts rendered months apart may not match exactly. Plan to re-render the full set together.

Amazon Connect Customer is AWS's current name for the product previously called Amazon Connect. In Amazon Connect Customer, the Set voice flow block can override the speaking style to Neural or Generative voices. AWS says you are charged for Generative voices under Amazon Polly pricing, while Polly default voices, such as the Neural and Standard voices, are free in Amazon Connect Customer. AWS also says the Generative upgrade reaches Amazon Connect flows automatically.

The AWS documentation we reviewed on 29 October 2026 does not state which Generative voices an Amazon Connect Customer instance in Asia Pacific (Mumbai) can select. Test it in your own instance before you plan an IVR around Kiara or Arjun.

• Speech marks. Generative voices do not support speech marks. If you need word highlighting or lip-sync, use Neural or Standard voices, which support them.
• Newscaster style. The Newscaster speaking style is not supported on the Generative engine.
• SSML. Generative voices support many, but not all, SSML tags. Test the tags you rely on.
• Hallucination. AWS says the Generative model can, in rare cases, hallucinate. A built-in stop mechanism reduces this risk but does not remove it. Listen to output from your real prompts.

Upvantix is an AWS Select Tier Services Partner. Our AI/ML and agentic AI services follow the 5/30 AI Delivery Commitment. For a Hindi voice use case, a 5-Day AI POC tests one agreed business workflow on a customer-provided data sample. It ends with a production architecture recommendation. If the POC is approved, a 30-Day Production Use Case builds it within the approved scope: identity, access and approval controls; guardrails, evaluation and observability; AWS deployment and documentation; and operational handover.

The commitment needs one prioritized use case, agreed success and acceptance criteria, the required data and system access, and a customer decision-maker. Customer, vendor or third-party delays pause the schedule, and regulated validation needs separate scope. We do not give legal or regulatory advice. Read more about our AWS data and AI delivery.

Talk to us about a Hindi voice use case on AWS.