Indian Languages: Models, Speech and Documents for Automation

Facts in this report are as of 12 October 2026. It draws on Hugging Face model cards and the Hugging Face API, vendor pricing pages and blogs, cloud documentation, arXiv papers and secondary coverage, all read on 11–12 October 2026. Many vendor pages carry no date, and several could not be read because they render in JavaScript or refused access: Bhashini, AIKosh, parts of the Azure and AWS price pages, and Meta's WhatsApp rate card. The report says where that leaves a gap. Dollar prices are converted at ₹96 per US dollar, the rate used across this set (the X-Rates October 2026 average was 95.95) (X-Rates). Prices that vendors quote in rupees are given as quoted, and the report says whether GST is included or the source does not say. Source labels: primary means the vendor's, regulator's or author's own page was read; vendor-reported marks a company's claim about its own product, including benchmark results it ran itself, even on public test sets; secondary, single-source and Wikipedia-only mean what they say; preprint marks an arXiv paper that has not been peer-reviewed. This is research, not legal advice. Related documents in this set: India, Production tool stack, Costs, Regulation and security, Other models, Glossary and Index.

Key takeaways

• No independent study compares current frontier and Indian models on Indian languages. The best independent evidence is one June 2026 preprint on romanised, code-mixed instructions such as Hinglish. Every model tested lost accuracy as code-mixing rose. Claude Opus 4.6 led the closed models (68.7% falling to 61.2%), Sarvam-30B was the strongest Indic model (64.2% to 56.1%) and Gemini 3.5 Flash came next (61.4% to 54.1%). Vendor benchmarks from Sarvam and Gnani cannot be compared with each other.
• Two Indian labs now ship Apache-2.0 models of 30B parameters or more: Sarvam (30B and 105B, March 2026) and Gnani (a 30B mixture-of-experts model, August 2026). In Gnani's own MILU table, a general open model, gemma-4-31B (85.29), beats both Gnani (78.74) and Sarvam-105B (75.71). Check every licence: Sarvam-Translate is GPL-3.0 and Krutrim's models are "other".
• Sarvam is the only vendor with a full rupee price list for Indian-language work. Speech-to-text costs ₹30 an hour (₹0.50 a minute) and text-to-speech ₹3 per 1,000 characters. The 105B LLM costs ₹29.28 per million input tokens and ₹73.20 per million output tokens, translation ₹0.005 a character, and document processing ₹0.50–1.00 a page. The page is undated, says nothing about GST, and its quality claims are vendor-reported.
• Tokenizers multiply cost. On the GPT-4-era cl100k tokenizer, the same text needs 4.1 times as many tokens in Hindi as in English, 9.9 times in Tamil and 13.0 times in Malayalam. The newer o200k tokenizer cuts the ten-language average from 8.0 to 2.1 times (preprint).
• Most speech-to-text APIs list ₹0.24–1.63 a minute, but only Sarvam publishes Indian-language error rates, and it publishes only its own (19.31% word error rate on IndicVoices' top ten languages for Saaras V3). Coverage differs sharply. Google's Chirp 3 is generally available only for Hindi and Indian English, and Deepgram's Nova-3 lists five Indian languages.
• A self-assembled Hindi voice-agent minute costs about ₹2.97 in this report's worked example. Telephony (including the number rental) and speech make up ₹1.87 of that, and orchestration ₹0.96. US hosted platforms list ₹6.72–29.76 a minute before Indian telephony, and Bolna's pilot plan is ₹5.52.
• Two vendors say WhatsApp service messages stopped being free beyond 1,000 per number per month on 1 Oct 2026 (vendor-reported; not confirmed on Meta's page, which, as read on 11 Oct 2026 through a summarising fetch, still said non-template messages are free). On the vendors' account each further message costs ₹0.1150 excluding 18% GST, or ₹0.1323 on a second vendor's rupee tab. Meta's own rate file was not read. In the worked example, translating every bot reply through a per-character API would cost nearly eight times the rest of the WhatsApp flow (more if service messages stay free).
• Document and search tools publish prices but no Indian test results. Sarvam Vision 2.1 scores 90.5 on Sarvam's own Indic OCR benchmark, and Google Document AI lists plain OCR at ₹0.144 a page. No one has published accuracy on Aadhaar, PAN, GST invoices or handwritten Indian forms, and no code-mixed retrieval benchmark was found.
• No primary Indian rule found yet requires an AI voice to identify itself. TRAI's May 2024 Direction moves commercial voice calling onto DLT-managed number series: 140 for promotional, transactional and service calls, and 160 for service and transactional calls only. A draft 2026 amendment appears to treat fully automated calls as application-to-person traffic that must be declared in advance (inference from TRAI-hosted comments on the draft). Final rules were reported as notified on 18 Sep 2026 (secondary); the official text was not read, and no explicit AI-voice rule is confirmed.

Four conditions decide whether an Indian-language bot works

Indian SME automation hits conditions that English-first stacks are rarely tested on. The first is script and code-mixing. WhatsApp and SMS users often type Hindi, Tamil or Bengali in Latin script mixed with English. The only independent 2026 test of this input finds that every model degrades as code-mixing gets denser, and that detection tasks degrade more than reasoning tasks (Indi-RomCoM, arXiv 2606.30790) (preprint). The second is breadth. India has 22 scheduled languages, and official vendor lists cover anywhere from one of them (Llama, Claude's interface) to all 22 (Sarvam, IndicTrans2). The third is audio. Phone calls carry narrowband audio, and a voice that wins a full-band listening test can lose on the telephone. Sarvam's commissioned listening study reports exactly that reversal at 8 kHz (Sarvam Bulbul V3) (vendor-reported). The fourth is documents. Scanned forms mix scripts and handwriting, and no independent benchmark covers the Indian documents SMEs actually process.

Two further problems cut across all four. Cost does not scale like English cost, because tokenizers split Indian scripts into more pieces and translation APIs bill per character. And almost every quality number is reported by the company that sells the product. The sections below take each layer in turn (LLMs, evaluation, tokenizers, translation, speech recognition, speech synthesis, voice platforms, hosting, WhatsApp, documents, search and regulation), then set out a selection and test plan with a costed example. Terms such as WER and LID are defined in the Glossary.

Two Indian labs ship 30B-plus Apache-2.0 models; a general open model still tops the vendor's own table

Sarvam open-sourced its 30B and 105B models under Apache-2.0 in a blog post dated 6 Mar 2026. The Hugging Face repositories were created on 3 Mar 2026. Sarvam says both models were trained from scratch in India on IndiaAI Mission compute, with 16 trillion tokens of pre-training for the 30B and 12 trillion for the 105B. Both are mixture-of-experts designs with 128 experts, and the ten most-spoken Indian languages received "a substantial share" of the multilingual pre-training budget. Sarvam says the models are "already in production", the 30B in its Samvaad product and the 105B in its Indus app (Sarvam blog, 30B and 105B). The model cards give the 105B 10.3B active parameters and a 128K context window, and the 30B 2.4B non-embedding active parameters (sarvam-105b card; sarvam-30b card). The dates correct the India report, which gave 20 Feb 2026 for the open-sourcing. Sarvam's blog index dates "Introducing Indus" to 20 Feb 2026 (not 15 Feb) and Sarvam Edge to 14 Feb 2026 (not 13 Feb) (Sarvam blog index).

Gnani.ai released gnani-evon-v3.3-30B-A3B on 10 Aug 2026, also under Apache-2.0. It has 30B total and about 3.5B active parameters in a hybrid Mamba2-Transformer mixture-of-experts design, a 128K context window and support for 11 languages. Its card states "Developed by Gnani.ai" (Gnani card). That settles a question the India report left open: whether the "gnani" Hugging Face namespace belongs to the company. With 26,282 downloads on 12 Oct 2026, it is the most-downloaded Indian LLM repository found, ahead of Sarvam's 105B (17,634) and 30B (13,370) (Hugging Face API, sarvamai; Hugging Face search, gnani). BharatGen published Param2-17B-A2.4B-Thinking, a reasoning model tagged for 22 languages, on 16 Feb 2026. It has 4,470 downloads and no licence tag (Param2 card). The rest of the field is older or smaller:

• Krutrim's Krutrim-2-instruct (February 2025, licence "other") has 355 downloads.
• CoRover's BharatGPT models have tens of downloads.
• Soket's public models are 1B-scale, and its 120B-plus Project EKA is "in training" (company claim).
• Fractal's only LLM-scale model is Fathom-R1-14B, a reasoning model under MIT from May 2025.

(Hugging Face search, krutrim; Hugging Face search, corover; Hugging Face search, soket; Hugging Face, FractalAIResearch.) In July 2026 the minister told Parliament that the government had identified 20 sovereign model proposals for support, 12 LLMs and 8 small language models, naming Sarvam's 30B and 105B, Gnani's speech-to-speech model, BharatGen's multilingual models and Avataar's video model (Indian Infrastructure, 27 Jul 2026) (secondary). Download counts measure developer interest, not production use.

Sources: model cards and Hugging Face listings cited above; Sarvam API pricing; CXO Digitalpulse on Param-1.

Frontier vendors officially list few Indian languages, and Hindi is the only one they all list. Google's help page gives the Gemini app 13 of the 22 scheduled languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Tamil, Telugu and Urdu. That is an app-language list, not a statement of model capability (Gemini Apps Help). Among the scheduled languages, Claude's interface list includes only Hindi, though chat is not restricted to interface languages (Claude Help Center). Anthropic says separately that its Indic-language work covers ten languages (Anthropic, Bengaluru office) (vendor-reported). Anthropic's developer docs publish the only first-party per-language quality figures found. On a translated MMLU, scored relative to English, Hindi reaches 96.7% (Sonnet 4.5) and 92.4% (Haiku 4.5) and Bengali 95.4% and 90.4%. Tamil is not in the table (Claude multilingual support) (vendor-reported, older models). OpenAI's help article lists ten Indian languages, but the page returned HTTP 403 and only a search snippet was seen (OpenAI Help) (single-source). The Llama 3.1 and Llama 4 Scout cards name Hindi as the only scheduled Indian language (Llama 3.1 card; Llama 4 Scout card), and Mistral's 2024 Large 2 announcement does the same (Mistral Large 2).

Sources: help and documentation pages cited above; prices from Claude pricing, OpenAI pricing and Gemini API pricing; Gnani card.

Two practical readings follow. First, at equal token counts Sarvam 105B costs less than mid-tier frontier models but more than the smallest ones. Its ₹29.28 input and ₹73.20 output compare with ₹192 and ₹960 for Sonnet 5.5, ₹72 and ₹360 for Gemini 3.8 Flash, and ₹9.6 and ₹48 for Haiku 5.5 or GPT-6 Luna (our arithmetic at ₹96). Token counts are not equal across tokenizers, though, as the tokenizer section shows. Second, licences differ more than the "open" label suggests:

• Sarvam's and Gnani's LLMs, BharatGen's Param-1 2.9B and Soket's pragna-1b are Apache-2.0.
• AI4Bharat's IndicTrans2 and IndicConformer are MIT.
• Sarvam-Translate carries a GPL-3.0 tag, unusual for a model and material for any product or deliverable that includes it, whether you build it or a partner does.
• Krutrim's repositories say "other".
• BharatGen's Param2 has no licence tag.
• AI4Bharat's recommended successors, Indic-Transcribe and Indic-Speak, are gated, and their pages returned HTTP 401.

(sarvam-translate card; Hugging Face API, ai4bharat.) A licence review belongs in every statement of work, as the Production tool stack report argues for workflow tools.

The only independent 2026 test shows every model losing ground on Hinglish

Public Indic benchmarks exist, but none compares current flagship models side by side. MILU (AI4Bharat, preprint, version of November 2024) tests 11 Indic languages across 8 domains and 41 subjects. Of the more than 42 models it evaluated, GPT-4o had the highest average at 74%. Open multilingual models beat language-specific fine-tunes, and every model was weaker on arts, humanities, law and governance (MILU, arXiv 2411.02538). IndicGenBench (Google, ACL 2024, peer-reviewed) covers generation tasks in 29 Indic languages and 13 scripts. It found a significant gap against English for every language (IndicGenBench, arXiv 2404.16816). COMI-LINGUA (preprint, March 2025) has more than 125,000 Hindi-English code-mixed instances across five tasks. Closed models beat traditional tools and open models zero-shot, and one example in the prompt helped; model-level numbers were not read (COMI-LINGUA, arXiv 2503.21670).

Indi-RomCoM is the only 2026 study found that tests a frontier closed model and an Indian model on the same input. It covers romanised code-mixed instructions in Tamil, Bengali, Hindi and Gujarati, across seven tasks and three levels of code-mixing (preprint, June 2026). Zero-shot average accuracy, from the baseline to 75% code-mixing, was:

• Claude Opus 4.6: 68.7% → 61.2%
• Sarvam-30B: 64.2% → 56.1%
• Gemini 3.5 Flash: 61.4% → 54.1%
• GPT-3.5-Turbo: 52.3% → 44.3%
• LLaMA-3.2-1B: 24.3% → 11.6%

With three examples in the prompt, Claude rose to 74.6% → 67.6%, Sarvam-30B to 71.0% → 63.1%, Gemini to 67.2% → 60.5% and GPT-3.5 to 57.8% → 50.3%. The paper also measures a "register defection rate": how often a model abandons the user's code-mixed register and answers in English. The averages were Sarvam-30B 26.1%, Claude 35.9%, Gemini 44.5% and GPT-3.5 54.4% (Indi-RomCoM; PDF, tables 6, 7 and 13). Three cautions apply. The models are a generation old: Opus 4.6 and Gemini 3.5 Flash, not the Opus 5.5 or Gemini 3.8 Flash on sale now. The closed-model comparison point from OpenAI is GPT-3.5-Turbo. And the paper is not peer-reviewed. Even so, it shows two things that matter for WhatsApp bots. Heavy code-mixing costs the larger models roughly 7 to 8 points of accuracy, and the 1B Llama 12.7. And a model can be less accurate yet better at staying in the user's register, which affects how customers perceive a bot.

Vendor tables tell a different story because they measure different things. Gnani's card reports MILU scores from its own runs:

• Macro average over 11 languages: gemma-4-31B 85.29, gemma-4-26B-A4B 82.65, Gnani 78.74, Sarvam-105B 75.71, Sarvam-30B 67.15.
• Hindi: Gnani 81.99, Sarvam-105B 76.95, Sarvam-30B 67.08.
• Against hosted models on identical prompts: OpenAI's small gpt-5.4-nano edges Gnani on Hindi (82.29 against 81.99) and on a nine-language macro (79.69 against 79.46). DeepSeek-V4-Flash scores 76.02 on Hindi.

(Gnani card) (vendor-reported.) Sarvam's own benchmark uses 110 English prompts translated into 22 scheduled languages, in native and romanised script. Gemini 3 judges the answers, with Claude Opus 4 reference answers for usefulness. Sarvam 30B wins 89% of comparisons on average and 87% on STEM, maths and coding, but the fetched text does not name the opponents, so the result cannot be reproduced (Sarvam blog) (vendor-reported). Indi-RomCoM and Gnani's table do not contradict each other. One measures native-script knowledge questions, the other romanised instructions.

The evidence supports three conclusions. The sovereign models are not automatically better on standard knowledge tests. In Gnani's own table a general open model leads, which matches MILU's 2024 finding that multilingual open models beat language-specific fine-tunes. Hinglish needs its own evaluation, separate from native script, because it is the weak point for every model tested. A buyer cannot choose between Sarvam, Gnani and a frontier model from published numbers alone. The case for an Indian model rests on licence, hosting location, rupee pricing and register behaviour. Accuracy has to be measured on the buyer's own traffic. A 2026 study of Hinglish health questions was seen only as a search snippet with no retrievable URL, so it is not used here.

Tokenizers charge 2 to 13 times more for the same Indian-language text

A model's tokenizer decides how many billable tokens a sentence becomes, and Indian scripts split badly on English-first tokenizers. A July 2026 preprint measured the effect on the FLORES-200 development set (997 sentences). Under cl100k_base, the GPT-3.5 and GPT-4 tokenizer, the same content needs 4.1 times as many tokens in Hindi as in English, 6.5 in Bengali, 8.0 in Gujarati, 9.9 in Tamil, 10.7 in Telugu, 12.1 in Kannada and 13.0 in Malayalam, averaging 8.0 across ten Indian languages. The o200k_base tokenizer used from GPT-4o onwards cuts the average to 2.1 times, a 73% reduction. Under a fixed context budget, Indian-language documents keep much less of their original content (Tokenizer Tax, arXiv 2607.24276) (preprint).

Other studies point the same way. A Tokenization Equity Audit (IJCAI workshop, August 2026) found that Bengali needs 1.56 times as many o200k tokens as English for the same Python-debugging content. With the Qwen2.5 and Mistral tokenizers it needs up to 4.5 times (arXiv 2608.09046). Petrov and colleagues (NeurIPS 2023, peer-reviewed) found token-length disparities of up to 15 times across languages, and showed that cost and latency follow from the same mechanism (arXiv 2305.15425). Sarvam says its tokenizer is optimised for all 22 scheduled languages across 12 scripts, but its fertility chart is an image, and no numbers were extracted (Sarvam blog) (vendor-reported). No published study gives per-language token counts for the models on sale in October 2026 (GPT-5 and GPT-6 class, Claude 5.x, Gemini 3.x).

The cost effect is direct. A support exchange that takes 1,000 tokens in English takes about 4,100 in Hindi and 9,900 in Tamil on a cl100k-era tokenizer, and about 2,100 at the o200k average (our arithmetic). This interacts with per-token prices. At equal token counts Haiku 5.5 is cheaper than Sarvam 105B by about 3.05 times on input (₹9.6 against ₹29.28) and about 1.5 times on output (₹48 against ₹73.20). Sarvam's Indic tokenizer would need to produce 1.5 to 3 times fewer tokens for the same Hindi text to close that gap, and nobody has published that comparison (our inference). The Tokenizer Tax authors also warn against reading token efficiency as quality: the raw correlation between fertility and accuracy on Belebele is confounded by how much training data each language has. So measure tokens per message for each candidate model on your own transcripts, in each language and script, before budgeting a per-conversation cost. The worked example later does this with stated token counts.

IndicTrans2 is still the open reference, and per-character APIs can cost more than the chat

IndicTrans2 (AI4Bharat, MIT) was the first translation model family to cover all 22 scheduled languages. It came with the BPCC parallel corpus of 230 million sentence pairs, 126 million of them new (IndicTrans2, arXiv 2305.16307). It is still the most-downloaded Indic translation family on Hugging Face. The distilled 200M English-to-Indic model has 239,453 downloads, and the 1B English-to-Indic model 68,222. Both date from 2023 (Hugging Face API, ai4bharat). Its successor exists only as beta uploads: IndicTrans3-beta (April 2025, 303 downloads) and IndicTrans3-llama-beta (Hugging Face search, indictrans3). Sarvam-Translate (June 2025) is built on Gemma-3-4B and covers 23 languages, but carries a GPL-3.0 tag (sarvam-translate card). Sarvam's hosted translation and transliteration costs ₹0.005 a character on pay-as-you-go, ₹0.0045 on Growth and ₹0.004 on Business (Sarvam API pricing) (GST not stated).

The hyperscalers differ on the rarer languages. Google Cloud Translation lists 19 of the 22 scheduled languages; Bodo, Kashmiri and Santali are missing (Google Cloud Translation languages). Azure Translator's page includes Assamese, Bodo, Dogri, Kashmiri, Konkani, Maithili, Manipuri, Nepali, Odia and Sindhi. Sanskrit and Santali were not found, and the common languages were not checked one by one (Microsoft Learn, Translator languages). Neither vendor's per-character price was read. Bhashini, MeitY's translation platform launched in July 2022 (Wikipedia, Bhashini) (Wikipedia-only), cannot be costed. Its pages render in JavaScript and showed no pricing or quota text (Bhashini). One vendor article says Bhashini's documentation limits API use to proof-of-concept, with production access by arrangement, and that bhashini.ai and bhashiniservices.com belong to a separate private company (eCorpIT). That article was seen only as a search snippet (single-source).

For chat, per-character translation is often the most expensive line in the stack. At ₹0.005 a character, a 300-character bot reply costs ₹1.50 to translate. A WhatsApp service message beyond the free allowance costs ₹0.1150 on two vendors' account (not confirmed on Meta's page), and the LLM call that wrote the reply costs well under ₹0.40 in the worked example (our arithmetic). Letting the LLM answer directly in the user's language avoids that charge (our inference). Translation APIs earn their price on long documents, on languages a given LLM handles poorly, and on workflows that need a fixed, auditable translation. The test plan below includes both routes. Transliteration, converting romanised Hindi into Devanagari or back, is billed at the same per-character rate on Sarvam. Sarvam's speech model can also emit transliterated or code-mixed output directly, as the next section notes.

Speech-to-text costs ₹0.24 to ₹1.63 a minute, and only Sarvam publishes Indian error rates

Sarvam's Saaras models are the only speech recognisers in this research with published Indian-language error rates and rupee prices, and the rates are Sarvam's own.

• Saaras V4 (25 Sep 2026) claims "SOTA performance on all 22 Indian languages". It uses a 3B hybrid state-space LLM decoder and offers transcribe, translate, verbatim, transliterate and code-mix output modes. Sarvam reports language-identification error of 5.22% across 22 languages and 2.9% across the top ten, and a word error rate of 16.03% on AI4Bharat's IndicContextEval in the keyword-prompted "L5" setting (Sarvam Saaras V4). Its Vistaar charts cover Common Voice, Fleurs, Gramvaani, IndicTTS, Kathbath, Kathbath Noisy and MUCS in ten languages, but the values were not extracted.
• Saaras V3 (11 Feb 2026) reports 19.31% WER on IndicVoices' top ten languages, against about 22% for V2.5, from more than a million hours of training audio. It is compared with GPT-4o Transcribe, Gemini 3 Pro, Deepgram Nova-3 and ElevenLabs Scribe v2, but those numbers were not extracted (Sarvam Saaras V3) (vendor-reported).

Every Saaras mode costs ₹30 an hour (₹0.50 a minute), and batch transcription with speaker diarisation costs ₹45 an hour (₹0.75 a minute) (Sarvam API pricing) (page undated; GST not stated).

The global APIs publish languages and prices but no Indic accuracy:

• Google's Chirp 3 is generally available for Hindi and Indian English. Ten more Indian languages are in preview: Assamese, Bengali, Gujarati, Kannada, Malayalam, Marathi, Odia, Punjabi (Gurmukhi), Tamil and Telugu (Google Chirp 3). Its standard Speech-to-Text V2 tier, which includes "chirp", costs $0.016 a minute (₹1.54) for the first 500,000 minutes a month, falling to $0.004 (₹0.38) above 2 million (Google STT pricing).
• Deepgram Nova-3 multilingual lists Hindi, Bengali, Kannada, Tamil and Telugu (Deepgram models). It costs $0.0052 a minute for pre-recorded audio (₹0.50), and for streaming a promotional "current" $0.0058 (₹0.56) against a regular $0.0092 (₹0.88) (Deepgram pricing).
• AssemblyAI's Universal-2 costs $0.15 an hour across 99 languages (₹0.24 a minute). Universal-3.5 Pro costs $0.21 an hour (₹0.34) across 18 languages with code-switching, but whether each model supports Hindi is not confirmed (AssemblyAI pricing).
• OpenAI's transcription models cost $0.003 a minute for gpt-4o-mini-transcribe (₹0.29), $0.0045 for gpt-transcribe (₹0.43), $0.006 for gpt-4o-transcribe (₹0.58) and $0.017 for gpt-live-transcribe (₹1.63) (OpenAI pricing). Their language list was not read.
• ElevenLabs Scribe costs 330 credits a minute, which works out to about $0.055–0.066 (₹5.28–6.34) depending on plan (ElevenLabs pricing) (derived).
• Cartesia's Ink costs $0.39 an hour on its Scale plan (₹0.62 a minute) (Cartesia pricing).
• Azure lists hi-IN for speech-to-text, but its prices did not render (Azure speech languages). AWS Transcribe's India rates were not captured (AWS Transcribe pricing).

Sources: pages cited in the text; Hugging Face, IndicConformer; Hugging Face search, bharatgen (Shrutam-2); Plivo India voice pricing.

Three readings matter for builders. First, the cheap end of the range (₹0.24–0.50) includes models whose Indian-language support is unconfirmed or limited to a handful of languages. For anything beyond Hindi, the realistic options in this research are Sarvam, Google's preview locales, Deepgram's five languages and self-hosted IndicConformer. Second, the vendor error rates cannot be compared, because each vendor reports on different subsets of IndicVoices, Vistaar or Kathbath. The only defensible comparison is a run on a fixed test set that includes the buyer's own recordings. Third, telephony providers' built-in speech add-ons can cost far more than a direct API. Plivo's ASR at ₹1.70 per 15 seconds works out to ₹6.80 a minute, against Sarvam's ₹0.50 (our arithmetic), so the speech leg should be bought separately and wired in.

Text-to-speech prices differ fivefold, and the one Indian listening test was commissioned by a vendor

Sarvam's Bulbul V3 (7 Feb 2026) costs ₹3 per 1,000 characters for real-time or streaming synthesis (Sarvam API pricing) (GST not stated). It is the only Indian-language voice with a published listening study. Sarvam says Josh Talks ran a blind A/B test across 11 languages, with 50 to 70 annotators per language and about 20,000 votes. It compared Bulbul V3 with ElevenLabs v3 alpha, ElevenLabs v2.5 Flash and Cartesia Sonic-3. ElevenLabs v3 alpha led on full-band audio, Bulbul V3 beat Cartesia Sonic-3, and Bulbul V3 ranked first on 8 kHz telephone audio (Sarvam Bulbul V3). Sarvam describes the study, but the study itself was not read (vendor-reported). For call-centre use, the telephone result is the one that matters. It also shows why a voice should be judged at the bandwidth it will be heard at.

ElevenLabs charges 1 credit per character for Multilingual v2 and 0.5 to 1 credit for its Flash and Turbo v2.5 models. On the Pro plan ($99 for 600,000 credits) that is about $0.165 per 1,000 characters, or ₹15.84, about 5.3 times Sarvam's rate. ElevenLabs says its TTS covers 32 languages, but Hindi and Tamil were not confirmed in the fetched text (ElevenLabs pricing; ElevenLabs TTS docs). Deepgram lists Aura-2 at $0.030 per 1,000 characters (₹2.88) and Flux TTS at $0.045 (₹4.32), but names no Indian languages for them (Deepgram pricing). Cartesia charges about 1 credit per character for Sonic, which is about $0.037 per 1,000 characters (₹3.59) on its Scale plan ($299 for 8 million credits, derived). Its Indian-language list was not read (Cartesia pricing). Google prices Gemini-based voices per token ($0.50 per million input text tokens and $9.00 per million output audio tokens, with a second row at $1.00 and $18.00). How those units map to Chirp 3 HD voices, and which Indian voices exist, were not confirmed (Google TTS pricing). The open models are free but partly gated:

• AI4Bharat's Indic Parler-TTS (Apache-2.0) covers 20 Indic languages plus English, including Bodo, Dogri, Maithili, Manipuri, Santali, Sindhi and Sanskrit (Hugging Face, Indic Parler-TTS).
• IndicF5 (MIT, gated behind a contact form) covers 11 languages (Hugging Face, IndicF5).
• Both cards recommend the gated Indic-Speak (22 languages).
• BharatGen's sooktam2 has no licence tag (Hugging Face search, bharatgen).

No independent MOS study of Indian voices was found.

Converting TTS to a per-minute cost needs a speaking-rate assumption, and no Indian source gives one. Cartesia says Sonic uses roughly 750 to 800 credits per minute of audio at 1 credit per character, which implies 750 to 800 characters per minute of speech. That figure is for Cartesia's voices, with no language stated, so it is a placeholder. Hindi in Devanagari, or Tamil, may differ. At 800 characters a minute, Bulbul V3 costs about ₹2.40 per minute of synthesised speech, and ElevenLabs at the Pro-plan rate about ₹12.67 (our arithmetic). Builders should count characters per minute on their own recorded prompts. One rule applies whichever voice is chosen. Since 20 Feb 2026, India's IT Rules have required labels on distributed synthetic audio and video (Business Today, 10 Feb 2026; Hogan Lovells) (secondary). Whether that covers a synthetic voice on a live call is an open question, covered with the TRAI rules below.

A rupee voice-agent minute has to be assembled leg by leg

A voice agent's cost per minute is the sum of five legs: telephony, speech-to-text, the LLM, text-to-speech and the platform or orchestration layer. For Indian traffic the legs come from different vendors in different currencies. Indian telephony is the cheapest leg.

• Plivo's India page lists ₹0.38 and ₹0.25 a minute for calls within India and supported inbound routes, without saying which direction each figure covers. A domestic number costs ₹200 a month, and call recording is free (recordings stored free for 90 days, then $0.0004 a minute a month). The page does not state GST (Plivo India voice pricing).
• Exotel's own migration blog, as summarised by a search engine, puts Indian-number calls at about ₹0.80–1.00 a minute outbound and ₹0.30–0.50 inbound. It sets these against Twilio's ₹1.20–1.50 and ₹0.50–0.80 (Exotel blog) (vendor, via a search summary).
• Exotel's pricing page is sales-led, and Twilio's India page did not render readable rates (Exotel pricing; Twilio India voice).

Published platform prices for Indian voice agents are thin and partly contradictory. Bolna, an Indian platform, lists "6.00¢ (₹5.52) per minute" on its $15 pilot plan of 250 minutes, and 4.51¢ a minute at a higher volume tier (₹4.33 at ₹96). The page itself implies about ₹92 to the dollar, and what the per-minute price includes was not captured (Bolna pricing). No official Exotel voice-AI rate was found. A third-party integrator lists ₹6 a minute pay-as-you-go plus telephony, a ₹1,499 monthly plan at ₹5 a minute with 300 minutes, and a ₹4,999 plan at ₹4.50 a minute with 1,100 minutes. Other pages on the same site say "from ₹4 a minute" (edesy) (secondary, undated, conflicting). Gnani markets "Gnani Artha – India's sovereign AI stack" but publishes no price (Gnani). Smallest.ai claims speech recognition in 38+ languages with 64 ms latency (Smallest.ai) (vendor-reported). These are the only latency figures in this research, and they are claims, not measurements.

The US platforms price in dollars, and their default speech components are English-first:

• Vapi charges $0.05 a minute for hosting (₹4.80), plus pass-through components: a Deepgram transcriber at $0.0095–0.0099, an OpenAI model at $0.0077–0.0452 and an ElevenLabs voice at $0.0146–0.0238 (Vapi pricing).
• Retell lists $0.07–0.31 a minute all-in (₹6.72–29.76), with an example of $0.11 (Retell pricing).
• Bland charges $0.14 a minute on its Start plan (Bland pricing).
• Cartesia's managed agents cost $0.06 a minute (Cartesia pricing).
• OpenAI's gpt-live-1 voice sessions cost $0.05 a minute plus the backend model (OpenAI pricing).

For Indian languages, the Sarvam or AI4Bharat speech legs have to be plugged into these platforms as a custom integration (our inference from the components each platform lists). The open frameworks are the main self-assembly route: LiveKit Agents (Apache-2.0; cloud sessions $0.01 a minute beyond the included minutes) and Pipecat (BSD-2-Clause) (LiveKit pricing; LiveKit Agents repository; Pipecat repository). The Production tool stack report compares them in more detail.

The constructed floor is about ₹1.9 a minute for speech and telephony. Plivo's higher listed telephony rate (₹0.38), Sarvam speech-to-text (₹0.50) and Bulbul speech for 40% of each minute at the assumed 800 characters a minute (₹0.96) add up to ₹1.84 before the LLM and orchestration (our arithmetic). That is a constructed estimate, not a quoted price. It sits well below the hosted US platforms and below the published Indian platform prices, which suggests most of a platform's per-minute price pays for orchestration, margin and the LLM rather than for Indian speech (our inference). The worked example below completes the arithmetic. The only Indian outcome claim found comes from Meesho, which said its generative-AI support voice bot cut some customer call costs by 75%. TechCrunch reported the claim without a base, volume or period (TechCrunch, 26 Nov 2024) (company claim).

Processing inside India is possible for some layers, not yet for all

Where call audio and chat text are processed differs by layer. No vendor's own API processes requests in India. OpenAI, Anthropic and xAI process outside India, and xAI offers a US endpoint only. OpenAI's India endpoint (in.api.openai.com) offers regional storage but not regional processing, and needs approval and Modified Abuse Monitoring or ZDR (OpenAI data controls). Anthropic's inference_geo setting accepts only "global" and "us" (US at 1.1x); it does not apply on Bedrock, Google Cloud or Foundry (Claude data residency). The cloud platforms do better. AWS publishes India-only geo profiles (in.) for Claude Opus 5, Claude Sonnet 5, Claude Haiku 4.5, Kimi K3, GPT-5.6 Terra and GPT-5.6 Luna, routed only to Mumbai (ap-south-1) and Hyderabad (ap-south-2). AWS's regional-availability page also lists Claude Opus 5 and Sonnet 5 as In-Region in Mumbai on the bedrock-mantle endpoint (the model cards omit it; confirm in the console). The current generation is global-only from the India regions: Claude Haiku 5.5, Sonnet 5.5, Opus 5.5, Opus 4.7/4.8, Fable 5/5.1, Mythos 5.1, GPT-5.6 Sol, the GPT-6 family and Grok 4.6/4.7. AWS warns "Don't rely on the geographic prefix ... alone": check each profile with GetInferenceProfile (AWS Bedrock, Claude Sonnet 5 model card; Bedrock regional availability). AWS notes that prompts and outputs may be stored in opt-in regions for abuse detection. Hyderabad is an opt-in region, so that storage would stay in India.

The speech and open-model picture is mixed:

• Bedrock In-Region in Mumbai also serves many open-weight models: DeepSeek V3.1/V3.2, gpt-oss-120b/20b, Qwen3 235B/Coder 480B/32B/Next 80B/VL, Kimi K2.5, GLM 4.7/5, MiniMax M2.x, Mistral Large 3, Gemma 3 and Llama 3 8B/70B. Hyderabad has no In-Region generative model. Nova's "Geo" profile from Mumbai is the APAC profile, so it is not India-bound. Mumbai prices were not captured (Bedrock regional availability).
• Google's Mumbai region (asia-south1) processes Gemini 3.5 Flash, Gemini 2.5 Flash and text-embedding-005. Gemini 3.6, 3.7 and 3.8 Flash (the model in stack B and the worked example) are US and EU multi-region only. Delhi (asia-south2) is not listed, and no Claude model has India processing on Google Cloud. Chirp 3 is not processed in India; its ML processing is listed in Singapore, Taiwan and Tokyo (Google data residency).
• Oracle OCI Hyderabad (ap-hyderabad-1), the only OCI India generative-AI region, serves Cohere Command A (on demand and dedicated), Embed and Rerank. Llama 4 Maverick/Scout, Llama 3.3 70B and gpt-oss-120b/20b run there on dedicated AI clusters only. A Gemini 2.5 Flash call there "results in a call to a Google location", so it is not India processing (OCI generative AI regions).
• Azure's South India region, under its India-geography deployment type, offers gpt-4.1-mini, gpt-4o (2024-11-20), embeddings and Whisper, but no GPT-5.x or GPT-6 (Azure Foundry region availability).
• E2E Networks' TIR platform says it serves an "OpenAI-style endpoint served from Indian data centers". Its catalogue includes Sarvam-2B, and Sarvam models in its shared GenAI API (E2E inference endpoints) (vendor-reported).
• Yotta's Shakti Cloud serves Llama, DeepSeek, Qwen3 and Mistral Large 2 from "Navi Mumbai and Greater Noida", with no Indian-language model listed (Yotta Shakti Cloud endpoints) (vendor-reported).
• Sarvam does not say where its own API processes data on any page read. Ask in writing.

For self-hosting the Apache-2.0 Indian models, the IndiaAI compute portal lists an H100 SXM at ₹153.0 per GPU-hour on demand before subsidy (reserved: ₹134.1 for 1 month, ₹117.0 for 12 months; L40S ₹67.5, L4 ₹44.86; GST extra), with a maximum subsidy of 40% subject to approval (₹91.80 after the full subsidy). Vendor guides' "₹92" is this post-subsidy level, not a list price (India). Startups and MSMEs need DPIIT recognition plus ₹50 lakh in annual revenue or ₹100 lakh in funding. Other businesses are not a named category. Requests under 5,000 GPU-hours are auto-approved; larger ones go to the PMEC committee (IndiaAI price calculator; IndiaAI Ready Reckoner; IndiaAI eligibility). Commercially, E2E lists an H100 at ₹255.55 an hour on demand, as low as ₹155.90 on a commitment of three months or more, and ₹70 on spot (E2E H100). Both price lists exclude GST. How many concurrent calls or chats a Sarvam-30B or Gnani model serves on one such GPU was not measured in any source read.

Vendor-reported WhatsApp rates set the per-conversation cost; Meta has not confirmed the new service-message charge

Two vendors say that from 1 Oct 2026 each WhatsApp business phone number gets 1,000 free service messages per calendar month, and that from the 1,001st each service message is charged at the utility rate (vendor-reported; not confirmed on Meta's page). On their account the allowance belongs to the number, not the account, one-to-one and group sends share it, and unused messages do not roll over; before 1 October, service messages were free. Meta's own pricing page, as read on 11 Oct 2026 through a summarising fetch, still said non-template messages are free, and Meta's rate-card file was not obtained, because Meta's developer and business pages returned HTTP 400 to the fetch tool (Gallabox overview; ChatMaxima). ChatMaxima, titled "Updated July 2026" and "effective 1 Oct 2026", lists:

• marketing ₹0.8631
• utility ₹0.1150
• authentication ₹0.1150
• authentication-international ₹2.4971
• service ₹0.1150

All are per message and exclude 18% GST. Utility and authentication carry volume discounts of up to 30%. Gallabox's rupee tab shows ₹0.9926 for marketing and ₹0.1323 for utility, authentication and service. That is exactly 1.15 times ChatMaxima's figures, for reasons Gallabox does not state (Gallabox rate card). Both are single vendors, and neither is Meta.

Two vendors give ₹0.1150 excluding GST (or ₹0.1323) per service message after a 1,000-message allowance (vendor-reported). Meta's own page, as read on 11 Oct 2026 through a summarising fetch, still said non-template messages are free, so the charge itself is unconfirmed and Meta's file is still the only authority. A second conflict remains. ChatMaxima says utility templates delivered inside an open service window "remain free", while Gallabox's overview says that after 1 October they are charged "from the very first one". The INR billing deadline stands as the earlier reports gave it, from Meta's page as read on 11 Oct 2026. Eligible customers whose Sold-To country is India must migrate every WhatsApp Business account to rupee billing by 31 Dec 2026, or their messages stop being delivered on 1 Jan 2027 (Meta WhatsApp pricing). A re-read on 12 Oct 2026 was refused (HTTP 400). Business-solution-provider markups (Gupshup, Interakt, AiSensy, Wati) were not researched, and for small accounts they may matter more than Meta's rates.

The language design matters as much as the rate. Users type romanised, code-mixed text on WhatsApp, which is the input Indi-RomCoM found weakest for every model (the paper's notes draw this link to WhatsApp and SMS; it is an inference, not a usage survey). Two behaviours should be tested before launch. The first is accuracy on Hinglish against native script. The second is whether the bot keeps replying in the user's register or drifts into English, where Sarvam-30B drifted least (26.1%) and GPT-3.5 most (54.4%) in that study. Voice notes arriving on WhatsApp need the speech-to-text leg too: at Sarvam's batch rate, a 30-second note costs ₹0.25 to transcribe (our arithmetic).

Document AI has rupee page prices but no independent test on Indian documents

Sarvam Vision 2.1 (24 Sep 2026) is a general vision-language model with a layout harness. Sarvam claims OCR in 22 Indian languages and new recognition of Indic handwriting. On Sarvam's own "Indic OCR Bench", Vision 2.1 scores 90.5 against 82.0 for Bodhan Indic-OCR (the metric label was not extracted). On OmniDocBench v1.6 it scores 94.97 overall against a comparison row of 92.17 (Sarvam Vision 2.1) (vendor-reported). Sarvam says the model API now costs less than it announced at launch. Its price page lists digitisation at ₹0.50 a page and extraction at ₹1.00 a page (Sarvam API pricing) (GST not stated). The hyperscalers are cheaper for plain OCR, but none states its Indian-script accuracy:

• Google Document AI's Enterprise Document OCR lists $1.50 per 1,000 pages (₹0.144 a page) from 1,000 to 5 million pages a month, with the first 1,000 free and $0.60 per 1,000 above 5 million. Its Layout Parser costs $10 per 1,000 pages (₹0.96 a page) and its custom extractor $30 per 1,000 (₹2.88 a page). Some column labels were not fully captured (Google Document AI pricing).
• Azure Document Intelligence, read through Azure's retail-price API, shows two unreconciled Central India "Read" rows at $0.60 and $1.50 per 1,000 pages. Prebuilt models cost $10 per 1,000 and custom models $30 per 1,000 (Azure retail prices API).
• AWS Textract lists text detection at $0.0015 a page (₹0.144) for the first million pages in US West, expense analysis at $0.01 (₹0.96) and ID analysis at $0.025 (₹2.40) (AWS Textract pricing).
• Mistral OCR's per-page price was not captured (Mistral pricing).

Indian-language coverage was not read for Google, Azure or AWS.

Sources: pages cited in the text; Surya; olmOCR; Tesseract data files; Docling repository; DeepSeek-OCR repository; Hugging Face search, krutrim.

The price comparison changes with the task. For plain text detection, Sarvam's ₹0.50 page is about 3.5 times Google's or Textract's ₹0.144. For field extraction, Sarvam's ₹1.00 is close to Azure's prebuilt models and Textract's expense analysis (₹0.96), and about a third of Google's or Azure's custom extractors (₹2.88) (our arithmetic). Sarvam's "digitisation" may well be layout-aware conversion rather than plain OCR, which would make Google's Layout Parser the fairer comparison, but the sources read do not say. That price gap only matters if accuracy on the target documents is at least equal, and no one has published independent accuracy on Aadhaar or PAN cards, GST invoices, bank statements or handwritten Indian forms. Sarvam's handwriting claim and its synthetic-form training data are vendor statements. The practical consequence is the same as in the Production tool stack report: pilot on your own scans before relying on any accuracy figure. Use masked or consented samples, because identity documents carry personal data that falls under the DPDP duties applying from 13 May 2027 (DPDP Rules, G.S.R. 846(E)). The vision abilities of Gemini, GPT and Claude on Indian scripts were not researched for this report.

Multilingual embeddings claim 100 languages; nobody has tested Hinglish retrieval

Retrieval-augmented bots for Indian SMEs search FAQs, catalogues and policies that mix English, native script and romanised text. The open embedding models claim broad coverage, but none publishes Indian-language scores on its card. BGE-M3 (MIT) supports "more than 100 working languages" (BGE-M3). multilingual-e5-large-instruct covers 94 languages (multilingual-e5-large-instruct). Jina embeddings v4 covers "30+" (Jina v4), and Jina says buying API tokens does not license its weights for self-hosting (Jina embeddings). Qwen3-Embedding (Apache-2.0) ranked first on the MTEB multilingual leaderboard on 5 Jun 2025, a dated vendor claim (Qwen3-Embedding-8B card). Krutrim's Vyakyarth embedding model has 29,232 downloads under an "other" licence and no published score (Hugging Face search, krutrim).

The Indian benchmarks exist but have not yet produced a readable ranking. Indic MTEB covers 25 languages and seven task types, including retrieval and reranking. It appears in an August 2026 arXiv paper (2608.05785), seen only through a search summary, and extends the MMTEB framework (MMTEB, arXiv 2502.13595) (secondary). The leaderboard itself was not reached. An IIT Gandhinagar study benchmarked eight multilingual embedding models on low-resource Indian languages (arXiv 2601.10205). According to a summary that was not read in full, E5-Large-Instruct was strongest on monolingual retrieval and LaBSE strongest on classification (secondary). Qwen-Indic-v1 self-reports an MTEB (Indic) average of 73.80 across 20 tasks (Qwen-Indic-v1) (single-source; card not fetched). The general MTEB leaderboard renders in JavaScript and was not read.

Sources: cards and pages cited in the text; Cohere pricing; Voyage pricing; Google data residency.

No code-mixed retrieval benchmark was found, and the IIT Gandhinagar abstract does not confirm code-mixing as a test condition. That gap matters, because a customer may type "refund kab milega" against an FAQ written in English or Devanagari. Until a benchmark exists, a builder has to construct the test from real queries. Run each query in its native script, in romanised form and after transliteration, and compare recall. Sarvam's transliteration API (₹0.005 a character) or Saaras's transliteration output mode can produce the transliterated variant (our recommendation; no source tests it). Self-hosted BGE-M3 or multilingual-e5 keeps embedding cost at zero and data in-house. Cloud embeddings processed in India exist on Google (text-embedding-005 in Mumbai), on Bedrock (Cohere Embed Multilingual in-region in Mumbai) and on Azure's India geography (Bedrock regional availability).

TRAI's draft rules appear to treat fully automated calls as A2P traffic; no AI-voice rule is confirmed

The only primary TRAI text on voice calling read for this report is the Direction of 4 May 2024. It ordered access providers to put the 140-level number series (promotional, transactional and service voice calls) on a DLT-based voice platform within 60 days. It also ordered them to plan within 30 days, and implement within 90 days, the 160-level series reserved for service and transactional calls. The Direction records that 140-series traffic had been running outside the DLT platform, and that banks and others had been making promotional calls from ordinary 10-digit numbers (TRAI Direction, 4 May 2024) (primary). Its implementation status in 2026 was not found. For a business deploying voice agents, the practical reading is that outbound calls go out on number series the operator manages through DLT, so the calling business has to complete its operator's registration before a campaign starts. Promotional and service calls sit in different series (our inference from the Direction, which is addressed to operators).

A draft TCCCPR Third Amendment Regulations, 2026 exists, and its final status is unclear. Comments that TRAI hosts from April 2026, submitted by Exotel and Veeno, confirm the draft. They discuss:

• a new definition of application-to-person (A2P) calls
• termination charges and advance declaration for A2P calls
• a lower threshold of three complaints in ten days for suspending a sender's calling-line identity
• AI-based flagging of high-volume calls by the originating and terminating operators

The commenters ask that API-initiated calls with a live human on the line be excluded from A2P. That implies a fully automated voice agent falls inside A2P as drafted (TRAI-hosted comments on the draft) (primary for the draft's existence; draft text not read). MediaNama reported a termination charge of up to 5 paise a minute for such calls (MediaNama, March 2026) (secondary). Several commentaries say the amendment was published in the Gazette on 18 Sep 2026. On that account, entities making automated or pre-recorded calls must declare that use, and their calling-line identifiers, to their telecom provider, undeclared traffic is treated as unsolicited, and most provisions apply within 30 days, some within 60 and the annexed amendments within 90 (CareerIndia; indialaw; Policy Edge) (secondary). But TRAI's unsolicited-commercial-communication page, read on 12 Oct 2026, lists no final 2026 amendment (TRAI, unsolicited commercial communication). So the final text, its date and the charge are unverified. A Delhi High Court challenge listed in March 2026 has no known outcome.

No primary text read contains an explicit rule for AI voice agents. Secondary summaries of TRAI's February 2025 amendment disagree. S.S. Rana says explicit consent for a commercial transaction is valid for seven days from when it is obtained (S.S. Rana). Other summaries describe a spam-reporting window extended from three days to seven, and inferred consent limited to the life of a contract (secondary, not re-read). This replaces the earlier reports' statement that TRAI rules for AI calls were "not found" or "not researched". The binding position is still unknown.

Three other rules bind, or will bind, a voice or chat deployment:

• The IT Rules amendment on "synthetically generated information" (G.S.R. 120(E), notified 10 Feb 2026, in force 20 Feb 2026) covers audio, visual or audio-visual content created or altered by computer so that it appears real. It requires a prominent label. Sources disagree on whether provenance metadata is mandatory or required only "where technically feasible" (iPleaders; Gala Law) (secondary; Gazette not read). Whether a live synthetic voice on a call counts as distributed content is an open question for counsel.
• The DPDP Rules, read from the Gazette text, require from 13 May 2027 breach notice "without delay", a detailed report to the Board within 72 hours (Rule 7(1) and 7(2)), logs needed to detect unauthorised access for one year (Rule 6(1)(e)) and processing logs for at least one year (Rule 8(3)). On cross-border transfer, Act s.16(1) lets the government restrict transfers to notified countries, and Rule 15 makes transfers subject to conditions set by order. Both apply from 13 May 2027, and no s.16 notification or Rule 15 order was found (MeitY's index was unreadable). That matters because every frontier API processes outside India (DPDP Rules, G.S.R. 846(E)) (primary).
• CERT-In's binding directions under section 70B of the IT Act (28 Apr 2022) require service providers and body corporates to report cyber incidents within 6 hours of noticing them, and to keep logs for 180 days within India. "Attacks or malicious/suspicious activities affecting systems/servers/software/applications related to Artificial Intelligence and Machine Learning" are on the reportable list (CERT-In s.70B directions) (primary).

Until the final TRAI text is read, a safe design has six parts (our inference, not a legal requirement shown in primary text):

• Complete the operator's DLT registration and declare automated calling before any outbound campaign, keeping promotional and service traffic on their separate series.
• Open every AI call with a short spoken disclosure.
• Keep timestamped consent records.
• Label synthetic audio wherever it is distributed.
• Treat recordings and transcripts as personal data under DPDP.
• Put a six-hour incident-reporting step into the runbook.

Inbound support calls carry less of the TRAI burden than outbound campaigns, though DPDP and CERT-In apply to both. If the reported A2P termination charge of up to 5 paise a minute is confirmed, it would add at most ₹0.05 a minute to outbound automated calls, small against the ₹2.97 voice minute in the worked example below (our arithmetic). Regulation and security covers DPDP and CERT-In in depth.

Selection and test plan: pilot three stacks on your own calls, chats and scans

The evidence above leaves no shortcut. The published benchmarks are vendor-run, old or on the wrong input, so you have to run your own comparison. The plan below is designed to be cheap enough for an SME pilot and repeatable for each new workflow or language.

What to pilot

Run all three on the same inputs, with the same prompts and tools, for each target language. Within stack B, also compare the LLM writing directly in the user's language with the LLM writing in English followed by IndicTrans2 or an API translation. The translation route only makes sense for chat if it wins on quality, given its cost.

Test sets

Metrics

Two rules keep the pilot honest. Choose the cheapest stack whose task success is not measurably worse than the best stack's, rather than the stack with the best single benchmark. And report results per language and per script, because an average across languages hides the failures that customers notice.

Worked cost example: a Hindi/English WhatsApp and voice support flow

The example prices one month of a support flow for an Indian SME using stack A's list prices, then varies one leg at a time. Every volume and token count below is an assumption, marked A1 to A12. Prices come from the sources above.

On tax, the WhatsApp rates (vendor-reported) exclude 18% GST, adding ₹642 on the ₹3,565 of WhatsApp charges. The service-message line (₹3,335) rests on the vendor-reported 1,000-message allowance. If Meta's page, which on 11 Oct 2026 still said non-template messages are free, is right, that line is ₹0, the WhatsApp subtotal is ₹2,456 and the total before tax is ₹20,305 (our arithmetic). The Sarvam and Plivo pages do not state GST, and how GST applies to dollar-billed services such as LiveKit was not verified. Adding 18% to every line gives about ₹27,900. Whether that applies, and whether it can be offset, depends on each supplier's invoicing.

The structure of the result matters more than the total. On the voice side, text-to-speech and orchestration are the largest lines (₹5,760 each), and the LLM is the smallest (₹849). On the WhatsApp side, the per-message charges outweigh the LLM if the vendor-reported service charge applies; if it does not, the LLM (₹1,976) is the largest WhatsApp line. Changing one leg at a time shows where the decisions really are:

The LLM rows use equal token counts for every model, which the tokenizer section shows is not realistic for Hindi. The Claude rows price cached tokens at the cache-read rate only. Anthropic charges cache writes at 1.25x input (5-minute cache) or 2x (1-hour): Haiku 5.5 $0.125 / $0.20 and Sonnet 5.5 $2.50 / $4 (Claude pricing). Treat those rows as lower bounds. Replace A5 and A10 with token counts measured on the pilot transcripts before budgeting.

Four lessons follow, all from the arithmetic rather than from any vendor's claim. The voice choice moves the bill more than the LLM choice. Moving from Bulbul to ElevenLabs adds more than moving from Sarvam 105B to Sonnet 5.5. Orchestration costs as much as speech synthesis, so self-hosting Pipecat or LiveKit is worth costing once volume is steady. Per-character translation in the chat loop is the single most expensive design mistake available, at nearly eight times the rest of the WhatsApp side (about 18 times if service messages stay free). Price changes already on the calendar are material. Gemini 3.8 Flash's January 2027 doubling would add ₹6,653 a month here on its own (₹13,306 against ₹6,653). Costs covers how such price changes flow into budgets and contracts.

What this means for your business

Run your own evaluation before you choose. No published evidence settles which LLM, voice or OCR engine works best for your traffic, and most of what exists is vendor-run. Build a small test set from consented call recordings, romanised WhatsApp transcripts, masked document scans and FAQ queries in three scripts, and score it per language and per script, with costs in rupees. Re-run it whenever a vendor ships a new model, which in 2026 has meant every few weeks. Choose the cheapest stack whose task success is not measurably worse than the best.

Keep each leg swappable. In the worked example the voice and orchestration legs dominate cost, and the quality evidence suggests different winners for different tasks: a frontier model may do best on code-mixed reasoning, an Indian model on staying in register, and an Indian voice on telephone audio. Routing each leg separately lets you switch one on evidence. Apache-2.0 Indian models hosted in India give a tested fallback if a frontier API becomes unavailable or you need in-country processing. The India report sets out the model-access risks.

Budget in rupees, with usage lines kept separate. The worked support flow costs about ₹23,640 a month before tax (₹1.16 a chat, ₹2.97 a voice minute), on assumptions to replace with pilot measurements. Sarvam bills speech, documents and its LLM in rupees, WhatsApp bills in rupees from 2027 for eligible accounts, and Indian telephony is in rupees; the dollar exposure sits mostly in orchestration and frontier LLMs. Confirm GST for every supplier; the price pages rarely state it. Translating every chat reply through a per-character API would add ₹45,000 a month in the example.

Mark the dates. India-billed WhatsApp accounts must move to rupee billing by 31 Dec 2026, or their messages stop being delivered on 1 Jan 2027. Gemini 3.8 Flash prices double on 1 Jan 2027. The DPDP duties, including a 72-hour breach report and one-year logs, apply from 13 May 2027.

Make disclosure and consent the default. No primary Indian text read so far requires an AI voice to identify itself, but the draft TRAI amendment points towards declared, flagged and charged automated calling, and final rules were reported as notified on 18 Sep 2026 (secondary; the official text was not read). Open every AI voice call with a spoken disclosure, and register outbound automated calling with your operator. Log consent and recordings on a DPDP-ready retention schedule, with a six-hour CERT-In reporting step in the incident runbook. Building this in from the first pilot costs little; retrofitting it after a complaint threshold triggers costs far more.

Ask any vendor or implementation partner for:

• results on your own test set, per language and per script, rather than vendor benchmarks;
• written confirmation of where call audio and chat text are processed (Sarvam's pages do not say);
• the licence of every model in the statement of work, since Apache-2.0, MIT, GPL-3.0, "other", untagged and gated licences all appear;
• WhatsApp and telephony as separate lines in every quote, because both changed in 2026.

Open questions and unverified items

Sources

All sources were read or queried on 11–12 October 2026 unless the text gives another date. Labels follow the scope note.

1. X-Rates: USD to INR monthly averages, 2026
2. Sarvam blog: open-sourcing Sarvam 30B and 105B (6 Mar 2026) (vendor)
3. Sarvam blog index (vendor)
4. Hugging Face: sarvam-105b model card
5. Hugging Face: sarvam-30b model card
6. Hugging Face: sarvam-translate model card
7. Hugging Face API: sarvamai models
8. Sarvam API pricing (undated) (vendor)
9. Hugging Face: gnani-evon-v3.3-30B-A3B model card (vendor-reported benchmarks)
10. Hugging Face search: gnani
11. Hugging Face: BharatGen Param2-17B-A2.4B-Thinking
12. Hugging Face search: bharatgen
13. Hugging Face search: krutrim
14. Hugging Face search: corover
15. Hugging Face search: soket
16. Hugging Face: FractalAIResearch models
17. Hugging Face API: ai4bharat models
18. Hugging Face search: indictrans3
19. Indian Infrastructure, 27 Jul 2026: government identifies 20 indigenous AI model proposals (secondary)
20. CXO Digitalpulse: BharatGen among eight firms selected for phase two (secondary)
21. Google: Gemini Apps supported languages
22. Claude Help Center: using Claude in your preferred language
23. Claude Platform docs: multilingual support (vendor-reported scores)
24. Anthropic: Bengaluru office and partnerships across India (16 Feb 2026) (vendor)
25. OpenAI Help: ChatGPT language support (HTTP 403; search snippet only) (single-source)
26. Hugging Face: Llama 3.1 8B Instruct model card
27. Hugging Face: Llama 4 Scout model card
28. Mistral: Mistral Large 2 announcement (July 2024)
29. Claude Platform docs: pricing
30. OpenAI API pricing
31. Gemini API pricing
32. MILU, arXiv 2411.02538 (preprint)
33. IndicGenBench, arXiv 2404.16816 (ACL 2024)
34. COMI-LINGUA, arXiv 2503.21670 (preprint)
35. Indi-RomCoM, arXiv 2606.30790 and PDF v1 (preprint)
36. Tokenizer Tax, arXiv 2607.24276 (preprint)
37. Tokenization Equity Audit, arXiv 2608.09046 (IJCAI workshop)
38. Petrov et al., language model tokenizers introduce unfairness between languages, arXiv 2305.15425 (NeurIPS 2023)
39. IndicTrans2, arXiv 2305.16307
40. Google Cloud Translation: language support
41. Microsoft Learn: Translator language support
42. Bhashini (JavaScript-rendered; no pricing read)
43. Wikipedia: Bhashini (Wikipedia-only)
44. eCorpIT: Indic language app localisation, 2026 (search snippet only) (single-source)
45. Sarvam: introducing Saaras V4 (vendor-reported)
46. Sarvam: Saaras V3 (vendor-reported)
47. Sarvam: Bulbul V3 (vendor-reported)
48. Hugging Face: AI4Bharat indic-conformer-600m-multilingual
49. Hugging Face: AI4Bharat IndicF5
50. Hugging Face: AI4Bharat indic-parler-tts
51. Google Cloud: Chirp 3 model and languages
52. Google Cloud: Speech-to-Text pricing
53. Deepgram: models and languages
54. Deepgram pricing
55. AssemblyAI pricing
56. ElevenLabs pricing
57. ElevenLabs docs: text to speech
58. Cartesia pricing
59. Microsoft Learn: Speech service language support
60. AWS Transcribe pricing (rates not captured)
61. Google Cloud: Text-to-Speech pricing
62. Smallest.ai homepage (vendor-reported)
63. Plivo: India voice pricing
64. Exotel: migrating from Twilio to Exotel voice AI (via search summary) (vendor)
65. Exotel pricing (sales-led)
66. edesy: Exotel AI voice agent pricing (via search summary) (secondary)
67. Twilio: India voice pricing (not readable)
68. Bolna pricing (vendor)
69. Gnani homepage (vendor)
70. Vapi pricing
71. Retell pricing
72. Bland pricing
73. LiveKit pricing
74. GitHub API: livekit/agents
75. GitHub API: pipecat-ai/pipecat
76. TechCrunch, 26 Nov 2024: Meesho cuts customer call costs by 75% (company claim)
77. OpenAI: data controls and residency
78. Claude Platform docs: data residency
79. AWS Bedrock: Claude Sonnet 5 model card (India geo profile)
80. AWS Bedrock: model regional availability
81. Google Cloud: Gemini Enterprise Agent Platform data residency
82. Microsoft Learn: Foundry models sold by Azure, region availability
83. Oracle OCI: generative AI model endpoint regions
84. E2E Networks: inference endpoints (vendor)
85. Yotta Shakti Cloud: AI endpoints (vendor)
86. IndiaAI compute portal: price calculator
87. IndiaAI compute portal: Ready Reckoner (rates as of 13 Mar 2026)
88. IndiaAI compute portal: eligibility criteria
89. E2E Networks: NVIDIA H100 pricing (vendor)
90. ChatMaxima: WhatsApp API pricing, India (single-source vendor)
91. Gallabox: WhatsApp rate card (single-source vendor)
92. Gallabox: WhatsApp pricing overview (vendor)
93. Meta: WhatsApp Business Platform pricing
94. Sarvam: Sarvam Vision 2.1 (vendor-reported)
95. Google Cloud: Document AI pricing
96. Azure Retail Prices API: Document Intelligence
97. AWS Textract pricing
98. Mistral pricing
99. GitHub: datalab-to/surya (vendor-reported benchmark)
100. GitHub: allenai/olmocr (self-reported cost)
101. Tesseract documentation: data files
102. GitHub API: docling-project/docling
103. GitHub API: deepseek-ai/DeepSeek-OCR
104. Hugging Face: BAAI/bge-m3
105. Hugging Face: intfloat/multilingual-e5-large-instruct
106. Hugging Face: jinaai/jina-embeddings-v4
107. Jina AI: embeddings
108. Hugging Face: Qwen/Qwen3-Embedding-8B (vendor-reported ranking)
109. Hugging Face: LingoIITGN/qwen-indic-v1 (card not fetched) (single-source)
110. MMTEB, arXiv 2502.13595 (secondary for Indic MTEB)
111. IIT Gandhinagar: multilingual embeddings for low-resource Indian languages, arXiv 2601.10205 (abstract read; results secondary)
112. Cohere pricing
113. Voyage AI pricing
114. CareerIndia: TRAI anti-spam rules 2026 compliance guide (secondary)
115. indialaw: TRAI anti-spam rules summary (secondary)
116. Policy Edge: TRAI finalises new spam rules (secondary)
117. S.S. Rana: TRAI telemarketing consent (secondary)
118. Business Today, 10 Feb 2026: rules on AI-generated content (secondary)
119. Hogan Lovells, 13 Feb 2026: India introduces mandatory labelling for AI (secondary)
120. MeitY: DPDP Rules, G.S.R. 846(E), 13 Nov 2025 (primary)
121. Gazette of India: DPDP Rules 2025, G.S.R. 846(E) (read 12 Oct 2026) (primary)
122. TRAI: Direction on voice DLT for 140- and 160-series calls, 4 May 2024 (primary)
123. TRAI-hosted comments (Exotel and Veeno) on the draft TCCCPR Third Amendment Regulations, 2026, April 2026 (primary for the draft's existence)
124. TRAI: unsolicited commercial communication page
125. MediaNama, March 2026: TRAI proposes a charge on automated calls (secondary)
126. iPleaders: IT Rules 2026 (secondary)
127. Gala Law: India's 2026 IT Rules amendment on AI-generated content (secondary)
128. CERT-In: Directions under section 70B of the IT Act, 28 Apr 2022 (primary)