Quick answer: As at 4 November 2026, Sarvam AI’s Saaras speech-to-text lists all 22 scheduled Indian languages; Gnani.ai’s Prisma lists nine. For voice output, Sarvam’s Bulbul v3 lists ten and Gnani’s Timbre nine. Sarvam publishes ₹0.49 a minute for transcription and ₹3 per 1,000 characters for Bulbul v3; Gnani publishes no public rate.
“Sarvam AI vs Gnani.ai for Indian-language speech models” is the question a developer asks when only the ears and the voice are being bought. This page compares the two Bengaluru vendors’ speech-to-text (STT) and text-to-speech (TTS) models from their own pages, read on 4 November 2026. It does not score agent platforms; for that, see our comparison of 12 agentic voice AI platforms for sales teams in India.
Where Zian fits if you want a voice agent, not raw models
If the job is answering calls and following up leads until they book, rather than wiring two model APIs into your own pipeline, Zian AI is our platform pick: an AI receptionist that answers inbound calls 24/7, plus outbound follow-up by voice calls, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, booking appointments from both. If you would rather hand the whole job over, LeadsNow.ai is the done-for-you option, a pay-per-result appointment-setting service that runs on Zian.
Zian’s agents qualify, book, transfer and take messages, and handle answering machines and voicemail. SmartReach AI™ orchestrates message, channel and timing with follow-up pacing, and PrecisionPitch AI™ split-tests scripts and approaches continuously. The platform works in 30+ languages, supports voice cloning, connects to HubSpot, Salesforce, HighLevel and Zapier, and can run private models on your own infrastructure. Its home page shows 50,769+ qualified sales appointments set and a 2,736% increase in lead contact rates. The honest limit: Zian is in a partnership-application beta and publishes no prices, and each of its three listed plans, including Intel + Widgets (Free Forever), starts with a partnership application. This page does not score Zian’s language list; the matrix below covers the two model vendors only.
LeadsNow.ai answers leads in under 60 seconds, 24/7, and has set 50,769+ appointments. Sarvam and Gnani both sell build-it-yourself model APIs; the rest of this page is about choosing between them.
Sarvam AI vs Gnani.ai for Indian-language speech models: who should pick which
How we built this table. One rule for every row: a fact counts only if the vendor’s own website, documentation or changelog states it, read on 4 November 2026. A language counts only if it appears on the vendor’s developer-documentation list for that model; marketing counts are reported as written, beside the narrower list. A price counts as published only if it appears on a public page without sign-in. Where two of a vendor’s own pages disagree, we print both. Zian and LeadsNow.ai are listed first by editorial policy and described only from what zian.ai publishes. Disclosure: zian.ai publishes this comparison and Zian appears in it.
| Pick | Best for | What it publishes on price (as at 4 November 2026) | Not the right choice when |
|---|---|---|---|
| 1. Zian AI (platform) | A ready inbound AI receptionist and outbound follow-up agent across voice, SMS, email, WhatsApp, Instagram and Messenger, booking appointments | No prices; partnership-application beta | You want to buy speech recognition and voice synthesis by the minute or character and build the agent yourself |
| 2. LeadsNow.ai (done for you) | Booked appointments without building agents; pay-per-result; runs on Zian | Pay-per-result; no rate quoted here | You want to own and edit every agent yourself |
| 3. Sarvam AI | Transcription in any of the 22 scheduled languages; published rupee rates; self-hosting on AWS | STT ₹0.49 a minute; Bulbul v3 ₹3.00 per 1,000 characters; Bulbul v4 Flash ₹2.00, or ₹1.00 under an offer running until 12 November | You need a refund on unused credits, or Malayalam and Odia on Sarvam’s newest voice model (its docs say not yet) |
| 4. Gnani.ai | The nine shared Indian languages with on-premise or air-gapped deployment, from a vendor that also sells agents, analytics, biometrics and agent assist | No public rate; its Prisma v3.0 page says “billed per hour of audio transcribed” | You must model unit cost before signing up or talking to sales, or you need Odia or any of the other 12 scheduled languages its docs do not list |
The quotable version: Sarvam is the vendor you can price and cover all 22 scheduled languages with for transcription; Gnani is the vendor you shortlist when your languages sit inside the nine it lists and deployment control matters more than a published rate.
Which speech models Sarvam and Gnani sell today
Both vendors shipped new versions in the last 90 days. Sarvam’s current recognition model is Saaras v4, which its Saaras model page calls “default, recommended, latest”; Saaras v3 remains available and Saaras v2.5 is marked for deprecation. Sarvam’s changelog shows Saaras v4 introduced on 18 August 2026 and made generally available on Voice Agents on 2 September. For voice output, Bulbul v4 Flash launched on 13 October 2026, while Bulbul v3 stays the API default when you omit the model. Sarvam’s changelog says Bulbul v3 reaches end of life on its Voice Agents product on 10 November 2026, and adds: “This change applies to Voice Agents only. Bulbul V3 stays available on the Text-to-Speech API.”
Gnani’s speech models are Prisma for recognition and Timbre for voice. Its models documentation says “The default model is currently Prisma v2.5”; Prisma v3.0, released on 22 October 2026 according to its speech API changelog, is opt-in and adds automatic language detection, a code-mixed output style and context hints. On voice, Gnani’s API introduction calls Timbre v2.5 “the recommended model”, and its voice catalogue also lists Timbre v3.0. The older names survive in the plumbing: the speech API base URL is api.vachana.ai and the Python SDK is published as gnani-vachana.
| Model | Job | Status as at 4 November 2026 | Ways to call it | Limits the vendor states |
|---|---|---|---|---|
| Sarvam Saaras v4 | Speech-to-text | Default, recommended | REST, Batch, WebSocket streaming; the product page labels its Realtime Streaming API beta | REST 30 seconds per request; Batch up to 2 hours per file; up to 50 keyterms |
| Sarvam Saaras v3 | Speech-to-text | Available; the self-hosted version | Same endpoints, plus SageMaker | Same as v4; no keyterm prompting |
| Sarvam Bulbul v4 Flash | Text-to-speech | Latest; 222 persona voice IDs | REST, HTTP streaming, WebSocket | 2,500 characters per REST request; no SSML |
| Sarvam Bulbul v3 | Text-to-speech | API default; 30+ named speakers | REST, HTTP streaming, WebSocket, SageMaker | 2,500 characters per REST request; no SSML |
| Gnani Prisma v2.5 | Speech-to-text | Default | REST, Realtime streaming, Batch | REST 30 seconds; Batch up to 4 hours per file; streaming sessions 15 minutes; 100 boost words |
| Gnani Prisma v3.0 | Speech-to-text | Opt-in, released 22 October 2026 | REST, Realtime streaming, Batch | Same as v2.5, with up to 200 keyterms and one to four language hints |
| Gnani Timbre v3.0 | Text-to-speech | 13 voices, each speaking all 10 listed languages | REST, Server-Sent Events streaming, WebSocket | Gnani labels mid-sentence code-switching beta on its Timbre v3.0 page |
| Gnani Timbre v2.5 | Text-to-speech | Recommended in the API introduction; 42 voices | REST, Server-Sent Events streaming, WebSocket | One preferred language per voice |
Sarvam documents streaming concurrency per plan (20 WebSocket streams on pay as you go, 100 on Pro and Growth); Gnani’s limits page says “Numbers will be listed on this page later”.
Which of India’s 22 scheduled languages each speech model lists
The Eighth Schedule to the Constitution of India lists 22 languages. Sarvam’s 22 Indian language codes for Saaras match that list exactly. The matrix applies one rule to every column: a tick means the language appears on the vendor’s developer-documentation list for that model, read on 4 November 2026. A dash means it is not on that list. A dagger (†) means the vendor’s documentation and its product page disagree. Hinglish-style voices (Sarvam’s English–Hindi “enhi” styles and Gnani’s “hi-en” voice) are not counted as separate languages.
| Language | Sarvam Saaras v4 and v3 (speech-to-text) | Sarvam Bulbul v3 (text-to-speech) | Sarvam Bulbul v4 Flash (text-to-speech) | Gnani Prisma v2.5 and v3.0 (speech-to-text) | Gnani Timbre v3.0 (text-to-speech) | Gnani Timbre v2.5 (text-to-speech) |
|---|---|---|---|---|---|---|
| Hindi | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Bengali | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Marathi | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Telugu | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Tamil | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Gujarati | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Kannada | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Malayalam | ✓ | ✓ | † | ✓ | ✓ | ✓ |
| Punjabi | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Odia | ✓ | ✓ | † | – | – | – |
| Assamese | ✓ | – | – | – | – | – |
| Urdu | ✓ | – | – | – | – | – |
| Nepali | ✓ | – | – | – | – | – |
| Konkani | ✓ | – | – | – | – | – |
| Kashmiri | ✓ | – | – | – | – | – |
| Sindhi | ✓ | – | – | – | – | – |
| Sanskrit | ✓ | – | – | – | – | – |
| Santali | ✓ | – | – | – | – | – |
| Manipuri | ✓ | – | – | – | – | – |
| Bodo | ✓ | – | – | – | – | – |
| Maithili | ✓ | – | – | – | – | – |
| Dogri | ✓ | – | – | – | – | – |
| English | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Of the 22 scheduled languages | 22 | 10 | 8 in the docs (10 on the product page) | 9 | 9 | 9 |
Four things in that grid change a buying decision.
- Sarvam’s 22 is a transcription number. Saaras covers all 22 scheduled languages plus English, but no Sarvam voice model speaks most of them. Bulbul v3 lists ten Indian languages and English; Sarvam’s Instant Voice Cloning, in beta at ₹2.00 per 1,000 characters, adds Assamese to those ten and says Urdu is not supported. An agent that must listen and reply in any of the other 11 scheduled languages, such as Urdu or Konkani, has a Sarvam recognition model and no Sarvam voice.
- Bulbul v4 Flash is narrower in the docs than on the product page. Sarvam’s Bulbul documentation says “Malayalam and Odia are not yet offered on v4 Flash”. Its text-to-speech product page answers “Which languages and voices are available?” with “Eleven languages”, Malayalam and Odia included, and its October launch blog says “83 voices across 11 languages”. Until the two agree, budget Malayalam and Odia at the Bulbul v3 rate.
- Gnani’s nine are the same nine on every model. Prisma v2.5, Prisma v3.0, Timbre v2.5 and Timbre v3.0 all list Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam and Punjabi, plus English. Odia is the gap: its language list omits it, while its speech-to-text marketing page names Odia and its API introduction says “10 Indian languages plus English”.
- Marketing counts are wider than either list. Gnani’s models page says “40+ Languages” for Prisma and “21+ Languages” for Timbre v2.5, which it labels beta there. Plan on the documented lists.
Sarvam’s Saaras and Gnani’s Prisma v3.0 each offer a “codemix” output that keeps English words in Latin script and the rest in native script. For how other speech-to-text APIs handle mid-sentence switching, see our comparison of nine speech-to-text APIs for multilingual and code-switched audio.
What Sarvam and Gnani charge for speech-to-text and text-to-speech
This is the sharpest difference. Sarvam prints a rupee rate card on its API pricing page and repeats it in its documentation. Gnani prints no rate on any public page we opened, as at 4 November 2026. We checked gnani.ai/pricing (which opens the Book a Demo page), its speech-to-text and text-to-speech API pages, its models, enterprise, Prisma v3.0 and Timbre v3.0 pages, the Prisma v3.0 launch post, its developer documentation index and its sitemap. Gnani’s Prisma v3.0 page answers “How is Prisma v3.0 priced?” with “Usage-based pricing, billed per hour of audio transcribed. Pricing is uniform across REST, Streaming, and Batch. Enterprise volume pricing is available.”
| Item | Sarvam AI (as at 4 November 2026) | Gnani.ai (as at 4 November 2026) |
|---|---|---|
| Speech-to-text, per minute | ₹0.49 pay as you go; ₹0.45 Pro; ₹0.41 Growth; ₹0.38 Business (real-time, streaming and batch) | Not published; “billed per hour of audio transcribed” |
| Speech-to-text with speaker labels | Batch with diarisation ₹0.59; ₹0.54; ₹0.50; ₹0.47 | Not published |
| Text-to-speech | Per 1,000 characters. Bulbul v3: ₹3.00; ₹2.85; ₹2.70; ₹2.60. Bulbul v4 Flash: pricing page ₹2.00 standard and ₹1.00 under “Limited time, 50% off until 12 Nov” (Pro ₹1.85 or ₹0.925); the documentation’s table shows ₹1.00 and ₹0.925 without the offer label | Not published |
| Billing increment | STT “priced per minute, billed per second”; TTS “rounded to the nearest character” | Speech-to-text “billed per hour of audio transcribed” (Prisma v3.0 page); text-to-speech not stated on the pages we read |
| Plans | Pay as you go ₹0 a month; Pro ₹9,999; Growth ₹49,999; Business ₹1,99,999; each monthly fee “comes back in full as credits”; Enterprise custom | Not published; “Enterprise volume pricing is available” (Prisma v3.0 page) |
| Free credits | ₹100 on sign-up | The Prisma v3.0 page says “1,000+ free credits” |
| Tax | “All prices are in Indian Rupees (₹) and exclude GST” | Not stated on the pages we read |
| Refunds | “All payments are final and non-refundable”; credits do not expire | Not stated on the Gnani pages we read |
| When credits run out | API calls fail with a 402 error until you top up | Gnani’s error reference lists a 402 “Insufficient credit” error |
Gnani’s rates sit in its developer console at app.gnani.ai, behind sign-in, so we have not seen them. Gnani’s own comparison with Ringg AI says, for its agent products, that “Gnani pricing is quoted per deployment through sales”. Note also the unit question: Sarvam bills voice output by characters of text, and Gnani does not publish its text-to-speech billing unit, so ask which unit a quote uses; the next section supplies a characters-per-minute figure for converting between the two.
What 10,000 minutes a month of speech-to-text and text-to-speech costs
Only Sarvam publishes numbers to multiply, so here is its month, worked end to end.
- Recognition: 10,000 minutes of caller audio transcribed in real time. Because Sarvam bills per second, 10,000 minutes is 600,000 billable seconds, plus at most a fraction of a second of rounding per request.
- Voice: 10,000 minutes of synthesised agent speech at 750 characters per spoken minute. That figure is our assumption, not a vendor number; it gives 7.5 million characters, or 7,500 thousand. Every 100 characters per minute you add or remove moves the Bulbul v3 line by ₹3,000.
- Prices: Sarvam’s published rates as at 4 November 2026, before GST, which Sarvam adds. The Bulbul v4 Flash columns use the ₹2.00 and ₹1.85 standard rates, not the offer that ends on 12 November.
- Currency: US dollars are our conversion at ₹94.75 per US$1, the cross rate from the European Central Bank’s 3 November 2026 reference rates (₹105.9155 and US$1.1179 per euro).
| Step | Sarvam: Pay as you go, Bulbul v3 | Sarvam: Pay as you go, Bulbul v4 Flash at the ₹2.00 standard rate | Sarvam: Pro plan, Bulbul v3 | Sarvam: Pro plan, Bulbul v4 Flash at the ₹1.85 standard rate | Gnani.ai |
|---|---|---|---|---|---|
| Speech-to-text: 10,000 minutes × rate | 10,000 × ₹0.49 = ₹4,900 | 10,000 × ₹0.49 = ₹4,900 | 10,000 × ₹0.45 = ₹4,500 | 10,000 × ₹0.45 = ₹4,500 | No public rate; ask for a quote on 10,000 minutes (Gnani says it bills per hour of audio) |
| Text-to-speech: 7,500 thousand characters × rate | 7,500 × ₹3.00 = ₹22,500 | 7,500 × ₹2.00 = ₹15,000 | 7,500 × ₹2.85 = ₹21,375 | 7,500 × ₹1.85 = ₹13,875 | No public rate; ask for a quote on 10,000 minutes (about 7.5 million characters) of synthesis |
| Usage before GST | ₹27,400 | ₹19,900 | ₹25,875 | ₹18,375 | – |
| Plan minimum (returned as credits) | None | None | ₹9,999 | ₹9,999 | – |
| Amount you pay before GST | ₹27,400 | ₹19,900 | ₹25,875 | ₹18,375 | Quote |
| Per minute of transcribed plus synthesised audio | ₹2.74 | ₹1.99 | ₹2.59 | ₹1.84 | – |
| In US dollars, our conversion at ₹94.75 per US$1 | US$289.18 | US$210.03 | US$273.09 | US$193.93 | – |
First, voice output is most of the bill: on pay as you go with Bulbul v3, transcription is ₹4,900 of ₹27,400, under a fifth. Second, Pro pays for itself here (its ₹9,999 fee returns as credits, against ₹25,875 of usage); Growth’s ₹49,999 minimum does not. Third, if the docs’ “not yet” on v4 Flash is right, Malayalam and Odia fall back to Bulbul v3 at ₹3.00 per 1,000 characters, 50% more than v4 Flash’s ₹2.00. While the offer lasts, pay as you go with v4 Flash at ₹1.00 comes to ₹12,400.
For Gnani, the only honest version of this table is a request: ask for a quote on 10,000 minutes of transcription and 10,000 minutes (about 7.5 million characters) of synthesis in your languages, with GST shown, and put it in the empty column. If you want the agent rather than the models, our Gnani vs Ringg AI voice-agent comparison covers what Gnani’s agent product publishes.
Rather not stitch speech models together yourself? Zian AI answers inbound calls 24/7 as an AI receptionist and follows up by voice, SMS, email, WhatsApp, Instagram and Messenger in 30+ languages, booking appointments into your calendar. Zian is in a partnership-application beta.
What accuracy evidence each vendor publishes, and who measured it
Both vendors publish word error rates (WER, the share of words a transcript gets wrong, so lower is better). Every figure here was published by a vendor. The first table is what each reports about its own models; the second lists the Sarvam accuracy and listening-test comparisons we found on Gnani’s pages: six speech-to-text benchmarks, then its main text-to-speech page and its per-language text-to-speech pages. Two Gnani blog posts, Cartesia vs ElevenLabs and Top 9 Voice AI Platforms, repeat the Prisma v2.5 explorer’s Kathbath Noisy result rather than adding a test. The only third-party study either vendor publishes about its own models is Sarvam’s text-to-speech listening test.
| Model | Headline figure | Test set | Who ran it | Compared with |
|---|---|---|---|---|
| Sarvam Saaras v3 (speech-to-text) | 19.31% WER | The ten most-spoken languages in IndicVoices | Sarvam | GPT-4o Transcribe, Gemini 3 Pro, Deepgram Nova3, Scribe v2 |
| Sarvam Saaras v4 (speech-to-text) | Language identification error 5.22% across 22 languages, 2.9% across the top ten; 16.03% WER on IndicContextEval with keyterms | Verified IndicVoices utterances; Vistaar across ten languages for WER | Sarvam | Per-dataset Vistaar results are in an interactive chart whose values are not in the page text, so we do not print them |
| Sarvam Bulbul v3 (text-to-speech) | Most preferred at 8 kHz telephony; at full band, “ElevenLabs v3 alpha leads on audio quality” | Blind A/B listening test in 11 languages, 50 to 70 annotators per language, over 20,000 votes | Josh Talks, which Sarvam calls “an independent third-party” | ElevenLabs v3 alpha and v2.5 flash, Cartesia Sonic-3 |
| Gnani Prisma v3.0 (speech-to-text) | Macro WER 16.93; Hindi 11.72 | Official Vistaar protocol, nine Indian languages, seven datasets | Gnani | ElevenLabs Scribe v2 (19.98), Microsoft Azure STT (20.51), Deepgram Nova 3 (26.57) |
| Gnani Prisma v3.0 (language identification) | 98% accuracy | Vistaar and Svarah clips longer than 3 seconds | Gnani | Not stated |
| Gnani Timbre v2.5 (text-to-speech) | MOS 4.23 on 8 kHz telephony audio | Not named on the page | Not named | Not stated for the MOS figure; Gnani’s separate listening tests, which include Sarvam, are in the next table |
Every row below is a Gnani vendor claim about a rival, read on 4 November 2026. Speech-to-text figures are WER (lower is better); the last two rows are blind listening tests, scored by ELO rating (higher is better) or, on two language pages, mean opinion score (MOS).
| Gnani page | Date on the page | Dataset or test | Sarvam model named | Gnani’s figure for itself | Gnani’s figure for Sarvam |
|---|---|---|---|---|---|
| Speech-to-text API page (chart values set by a script on the page) | Undated | Kathbath Noisy; Gramvaani; “All datasets except Kathbath Noisy” | Not named (“Sarvam”) | Prisma v2.5: 6.2%; 8.4%; 3.8% | 16.3%; 21.7%; 9.1% |
| Rural India ASR research page, Table 3, “Gnani AI Internal Benchmark” | 19 May 2026 (table dated May 2026) | GramVaani Hindi | Saarika v3.0 and Saarika v2.5 | “Gnani AI ASR Model”: 24.94% | 26.12% (v3.0); 26.75% (v2.5) |
| How to benchmark an STT API on Indian languages | June 2026 | Kathbath Clean; Gramvaani | Not named (“Sarvam”) | “Gnani STT API”: 8.38%; 26.24% (degradation +17.9 pp) | 9.41%; 26.71% (degradation +17.3 pp) |
| The real cost of low-accuracy STT (blog) | 11 August 2026 | Gramvaani | Not named (“Sarvam”) | “Gnani STT API”: 22.0% | 23.4% |
| Training on 14 million hours of telephonic speech (research) | 15 May 2026 | Kathbath Noisy, averaged across supported languages | “Sarvam 3.0” | 17.5% | 19.9% |
| Prisma v2.5 benchmarks (embedded explorer) | June 2026 | Kathbath and Kathbath Noisy among public datasets; 9 languages, 48 evaluations | Saaras v3 | Prisma v2.5 ranked #1 in “27 / 48” evaluations and in “8 / 9” languages on real-world and noisy audio; clean-to-noisy degradation +3.6% | The heatmap legend includes “Sarvam leads”; degradation +2.6% |
| Text-to-speech page (listening test) | Undated | “Ten-language Indic panel”; the Hindi and English pages print the same figures as “126 Hinglish comparisons” and “126 Indian English comparisons” | Not named (“Sarvam”) | Timbre v2.5: 1621 ELO, 75.0% win rate | 1417 ELO, 32.8% win rate |
| Per-language text-to-speech pages: Bengali, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu | Undated | A blind test per language, for example “100 Kannada comparisons” and “102 Marathi comparisons”; the Gujarati and Tamil pages show MOS | Not named (“Sarvam”) | Timbre v2.5 ELO: Bengali 1578.7; Kannada 1529; Malayalam 1697; Marathi 1515; Punjabi 1579; Telugu 1616. MOS: Gujarati 7.70; Tamil 7.32 | ELO: Bengali 1513.3; Kannada 1492; Malayalam 1419; Marathi 1551, ahead of Timbre v2.5; Punjabi 1486; Telugu 1490. MOS: Gujarati 6.20; Tamil 6.43 |
Read together, Gnani’s own figures disagree with each other. Its Gramvaani WER for itself is 8.4% in the speech-to-text page chart but 22.0% to 26.24% on its blog and research pages, and its Sarvam figure on that dataset runs from 21.7% to 26.71%; on Kathbath Noisy it gives itself 6.2% on one page and 17.5% on another. The published gaps run from under 1 point to more than 13 points depending on page and dataset: 0.47 points on Gramvaani in the benchmarking guide, 1.18 in the May 2026 internal benchmark naming Saarika v3.0, 13.3 in the undated chart naming no Sarvam model. In the Prisma v2.5 explorer, the fullest of them, Prisma v2.5 ranks first in 27 of 48 evaluations and in 8 of 9 languages on real-world and noisy audio, while Sarvam leads some cells and degrades less from clean to noisy audio (+2.6% against +3.6%); the explorer calls its results “independent evaluations” without naming who ran them. We found no independent test of either vendor against the other. The text-to-speech figures disagree too. The main page puts Timbre v2.5 204 ELO points ahead of Sarvam on a “ten-language Indic panel”, and the Hindi and English pages print the same figures as 126 Hinglish and 126 Indian English comparisons. On the other language pages Gnani’s lead over Sarvam runs from 37 points (Kannada) to 278 (Malayalam), and on its Marathi page Sarvam is ahead, 1551 to 1515.
We found no comparison with Gnani on Sarvam’s website or in its documentation, as at 4 November 2026: its Saaras v4 post, its Saaras v3 post and its speech-to-text page compare against other vendors instead. Gnani’s newest Prisma v3.0 benchmark post compares against ElevenLabs, Microsoft and Deepgram, not Sarvam. The 19.31 and 16.93 headline figures measure different things: an IndicVoices subset against Vistaar’s seven datasets. Do not subtract one from the other.
Run 200 of your own recorded calls, at 8 kHz, in your own languages, through both APIs, and score them the same way.
Where the models run: India processing, private cloud and self-hosting
| Question | Sarvam AI (own words) | Gnani.ai (own words) |
|---|---|---|
| Managed API | Runs in “Sarvam’s cloud”; your data is “sent to Sarvam over HTTPS” | Cloud API at api.vachana.ai; hosting region not stated on the speech API pages we read |
| Where processing happens | Speech-to-text page: “All processing happens within India. No cross-border transfers.” | Agent Builder regions: Asia, Americas, Europe (this list is for agents, not the speech API) |
| Retention | “Zero data retention after processing unless you explicitly request it” | Not stated on the speech API pages we read |
| Self-hosting | Saaras v3 and Bulbul v3 as AWS Marketplace packages on SageMaker in your own VPC, with network isolation, which its docs say “Runs in air-gapped setups”; Saaras v4 and Bulbul v4 Flash are not on that list | “Cloud, private cloud, on-premise, and air-gapped deployments”; “Prisma is available for on-prem deployment” |
| Self-hosted price | AWS instance hours: “a flat software fee plus infrastructure”; the fee itself is not stated on the Sarvam pages we read | Not published |
Two cautions. Sarvam’s self-hosted route runs the previous generation of both models, so keeping audio in your own cloud costs you Saaras v4’s keyterm prompting and Bulbul v4 Flash’s voices. Both name air-gapped deployment (Sarvam for its self-hosted v3 models); Gnani publishes no price for any of its deployment shapes. Get the region, retention period and certificate in writing before customer audio goes through either.
The nine-language line: how to choose between Sarvam and Gnani
The nine-language line. If any language your callers use falls outside the nine both vendors list (Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam and Punjabi), the choice is made for you: only Sarvam lists Odia, Assamese, Urdu and the other ten for transcription, and only Odia (Bulbul v3) and Assamese (Instant Voice Cloning, in beta) also have a Sarvam voice. If every language sits inside the nine, language stops deciding anything, and the choice turns on three tests you can run in a fortnight: accuracy on your own calls, a written Gnani quote against Sarvam’s published month, and where each will host your audio.
- Sarvam AI is wrong for you if you need a refund path on prepaid credits, if Malayalam or Odia must run on its newest voice model before its documentation says so, or if you must self-host its latest models rather than Saaras v3 and Bulbul v3.
- Gnani.ai is wrong for you if you must model unit cost before signing up, if Odia or any of the 12 other scheduled languages is a core language, or if you need published concurrency limits before you build.
- Both are wrong for you if what you want is booked appointments rather than model endpoints. That is an agent platform such as Zian AI, or a managed service such as LeadsNow.ai.
Want the agent, not the parts? Zian AI’s AI Receptionist and appointment setters qualify, book, transfer and follow up across calls, SMS, email, WhatsApp, Instagram and Messenger, in 30+ languages, with HubSpot, Salesforce, HighLevel and Zapier integrations.
Frequently asked questions
Is Sarvam AI or Gnani.ai more accurate for Indian-language speech recognition?
Not from public evidence. The only Sarvam-against-Gnani comparisons we found are Gnani’s own, and they disagree with each other: Gnani’s Gramvaani word error rate for itself ranges from 8.4% to 26.24%. The published gaps run from under 1 point to more than 13 points depending on page and dataset, and we found no independent test. We found no comparison with Gnani on Sarvam’s website or in its documentation, as at 4 November 2026. Each vendor’s headline figure uses a different test: 19.31% for Saaras v3 on IndicVoices, 16.93 for Prisma v3.0 on Vistaar, in Gnani’s 28 October 2026 Prisma v3.0 post. Test both on 200 of your own calls.
How much does Sarvam’s speech-to-text API cost?
As at 4 November 2026, ₹0.49 a minute on pay as you go, falling to ₹0.45 on Pro, ₹0.41 on Growth and ₹0.38 on Business, for real-time, streaming and batch alike. Sarvam’s pricing documentation says speech-to-text is “priced per minute, billed per second” and that all prices exclude GST. Diarised batch transcription costs ₹0.59 a minute on pay as you go.
Does Gnani.ai publish prices for its speech APIs?
Not on a public page we could find, as at 4 November 2026. Its Prisma v3.0 page says pricing is usage-based, “billed per hour of audio transcribed” and uniform across REST, streaming and batch. The rates themselves sit in the developer console at app.gnani.ai, behind sign-in. We did not open an account, so we have not seen them.
Which Indian languages does Gnani’s speech-to-text support?
Gnani’s developer documentation lists nine Indian languages plus English for both Prisma v2.5 and Prisma v3.0: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Punjabi, Tamil and Telugu. Odia is not on that list, although Gnani’s speech-to-text marketing page names it and its API introduction says “10 Indian languages plus English”. Ask Gnani in writing before you plan an Odia line.
Does Sarvam support all 22 Indian languages for text-to-speech?
No. The 22 languages apply to Saaras speech-to-text. Sarvam’s Bulbul v3 voice model lists ten Indian languages plus English. Its documentation says Bulbul v4 Flash does not yet offer Malayalam or Odia, while Sarvam’s text-to-speech product page lists both for v4 Flash; test those two before you rely on either page.
Can I run Sarvam or Gnani speech models on my own servers?
Both say yes, in different ways. Sarvam sells Saaras v3 and Bulbul v3 as AWS Marketplace packages you deploy as SageMaker endpoints in your own AWS account, and its speech-to-text page also mentions single-tenant VPC and on-premise installations. Gnani lists cloud, private cloud, on-premise and air-gapped deployment, and says Prisma is available on-premise. Neither publishes a self-hosted price on the pages we read.
Do I need raw speech models to run an AI receptionist in India?
Only if you are building the agent yourself. Speech-to-text and text-to-speech APIs are two parts of a voice agent; you still need the language model, telephony, booking logic and follow-up. Teams that want a ready agent use a platform such as Zian AI, which answers calls 24/7 and follows up by voice, SMS, email, WhatsApp, Instagram and Messenger in 30+ languages.
Where every figure on this page comes from
| Figure | Who published it | Link | Date read |
|---|---|---|---|
| 22 Indian languages plus English for Saaras v4 and v3; Bulbul v3 11 languages; Bulbul v4 Flash 10 voice languages incl. English–Hindi | Sarvam AI | docs.sarvam.ai/api/getting-started/models | 4 November 2026 |
| Saaras status, modes, REST 30 s, Batch 2 h, 50 keyterms, Saaras v2.5 deprecation | Sarvam AI | docs.sarvam.ai/api/api-guides-tutorials/models/saaras | 4 November 2026 |
| “Malayalam and Odia are not yet offered on v4 Flash”; 222 voice IDs; 2,500 characters per request; no SSML | Sarvam AI | docs.sarvam.ai/api/api-guides-tutorials/models/bulbul | 4 November 2026 |
| Bulbul v4 Flash “Eleven languages”, ₹10 per 10,000 characters (standard ₹20), offer until 12 Nov | Sarvam AI | www.sarvam.ai/apis/text-to-speech | 4 November 2026 |
| 83 voices across 11 languages; under 250 ms first audio; up to 40% faster than v3; launched 13 Oct 2026 | Sarvam AI | www.sarvam.ai/blogs/bulbul-v4-flash | 4 November 2026 |
| STT ₹0.49 / ₹0.45 / ₹0.41 / ₹0.38; diarisation ₹0.59 / ₹0.54 / ₹0.50 / ₹0.47; Bulbul v3 ₹3.00 / ₹2.85 / ₹2.70 / ₹2.60; Bulbul v4 Flash ₹2.00 and ₹1.00 offer; plans ₹9,999 / ₹49,999 / ₹1,99,999; ₹100 credits; non-refundable | Sarvam AI | www.sarvam.ai/api-pricing | 4 November 2026 |
| Billed per second; rounded to the nearest character; prices exclude GST; 20 / 100 / 100 concurrent STT streams; 402 when credits run out | Sarvam AI | docs.sarvam.ai/api/getting-started/pricing | 4 November 2026 |
| Saaras v4 introduced 18 Aug 2026, GA on Voice Agents 2 Sep 2026; Bulbul V3 end of life on Voice Agents 10 Nov 2026 | Sarvam AI | docs.sarvam.ai/changelog | 4 November 2026 |
| 19.31% WER on ten-language IndicVoices subset (Saaras v3); ~22% for v2.5 | Sarvam AI | www.sarvam.ai/blogs/asr | 4 November 2026 |
| 5.22% and 2.9% language identification error; 16.03% WER on IndicContextEval; Vistaar across ten languages; no mention of Gnani (nor in the Saaras v3 post or speech-to-text page); under 150 ms first token; 25 Sep 2026 | Sarvam AI | www.sarvam.ai/blogs/introducing-saaras-v4 | 4 November 2026 |
| Josh Talks blind A/B test, 11 languages, 50 to 70 annotators per language, over 20,000 votes, over 500 annotators | Sarvam AI | www.sarvam.ai/blogs/bulbul-v3 | 4 November 2026 |
| “All processing happens within India”; zero data retention by default; VPC and on-premise options (Realtime Streaming API beta label: www.sarvam.ai/apis/speech-to-text) | Sarvam AI | www.sarvam.ai/speech-to-text | 4 November 2026 |
| Self-hosted Saaras v3, Bulbul v3 on SageMaker; network isolation | Sarvam AI | docs.sarvam.ai/api/self-hosted/introduction | 4 November 2026 |
| Managed API data “sent to Sarvam over HTTPS”; self-hosted billed by instance hours | Sarvam AI | docs.sarvam.ai/api/self-hosted/hosted-vs-self-hosted | 4 November 2026 |
| Prisma v2.5 and v3.0: nine Indian languages plus English, no Odia | Gnani.ai | docs.gnani.ai/api/STT/languages | 4 November 2026 |
| Prisma v2.5 is the default model | Gnani.ai | docs.gnani.ai/api/STT/models/overview | 4 November 2026 |
| Prisma v3.0 released 22 Oct 2026, opt-in | Gnani.ai | docs.gnani.ai/api/introduction/changelog | 4 November 2026 |
| “10 Indian languages plus English”; Timbre v2.5 recommended; base URL api.vachana.ai | Gnani.ai | docs.gnani.ai/api/introduction/introduction | 4 November 2026 |
| REST 30 s; Batch 4 h; streaming 15 min; 200 keyterms; 100 boost words; limits “will be listed on this page later” | Gnani.ai | docs.gnani.ai/api/STT/limits | 4 November 2026 |
| Timbre v3.0 13 voices, 10 languages; Timbre v2.5 42 voices | Gnani.ai | docs.gnani.ai/api/TTS/available-voices | 4 November 2026 |
| Macro WER 16.93 vs 19.98, 20.51, 26.57; Hindi 11.72 vs 14.47; 98% language identification; seven datasets; 28 Oct 2026 | Gnani.ai | www.gnani.ai/resources/blogs/prisma-v3-0-speech-recognition | 4 November 2026 |
| “Billed per hour of audio transcribed”; “1,000+ free credits”; Prisma on-premise | Gnani.ai | www.gnani.ai/resources/prisma-v3 | 4 November 2026 |
| Timbre v2.5 MOS 4.23 on 8 kHz | Gnani.ai | www.gnani.ai/text-to-speech-api | 4 November 2026 |
| Odia named on the speech-to-text page; chart (page script): Prisma v2.5 6.2% / 8.4% / 3.8% vs “Sarvam” 16.3% / 21.7% / 9.1% on Kathbath Noisy / Gramvaani / all datasets except Kathbath Noisy; undated; no Sarvam model named | Gnani.ai | www.gnani.ai/speech-to-text-api | 4 November 2026 |
| Table 3, May 2026 internal benchmark, GramVaani Hindi: 24.94% vs Saarika v3.0 26.12%, Saarika v2.5 26.75%; page dated 19 May 2026 | Gnani.ai | www.gnani.ai/resources/research/the-rural-india-asr-problem-engineering-speech-recognition-for-background-noise | 4 November 2026 |
| Kathbath Clean / Gramvaani: Gnani STT API 8.38% / 26.24% (+17.9 pp) vs Sarvam 9.41% / 26.71% (+17.3 pp); June 2026 | Gnani.ai | www.gnani.ai/resources/research/how-to-benchmark-stt-api-indian-languages | 4 November 2026 |
| Gramvaani: Gnani STT API 22.0% vs Sarvam 23.4%; 11 August 2026 | Gnani.ai | www.gnani.ai/resources/blogs/cost-low-accuracy-stt-contact-center-compliance-churn | 4 November 2026 |
| Kathbath Noisy: 17.5% vs “Sarvam 3.0” 19.9%; 15 May 2026 | Gnani.ai | www.gnani.ai/resources/research/how-we-train-acoustic-models-on-14-million-hours-of-telephonic-speech | 4 November 2026 |
| Prisma v2.5 explorer vs Saaras v3: “27 / 48” evaluations ranked #1; “Sarvam leads” heatmap legend; degradation +3.6% Prisma vs +2.6% Sarvam; “independent evaluations”, evaluator not named; June 2026 | Gnani.ai | www.gnani.ai/resources/prisma-v2-5-stt-benchmarks | 4 November 2026 |
| Blind ELO: Timbre v2.5 1621 / 75.0% vs Sarvam 1417 / 32.8%, “ten-language Indic panel”; undated | Gnani.ai | www.gnani.ai/text-to-speech | 4 November 2026 |
| Same ELO figures captioned “126 Hinglish comparisons” (Hindi) and “126 Indian English comparisons” (English) | Gnani.ai | www.gnani.ai/text-to-speech/hindi and www.gnani.ai/text-to-speech/english | 4 November 2026 |
| Per-language blind tests, Timbre v2.5 vs Sarvam: ELO Bengali 1578.7 / 1513.3, Kannada 1529 / 1492, Malayalam 1697 / 1419, Marathi 1515 / 1551, Punjabi 1579 / 1486, Telugu 1616 / 1490; MOS Gujarati 7.70 / 6.20, Tamil 7.32 / 6.43; undated | Gnani.ai | Bengali, Gujarati, Kannada, Malayalam, Marathi, Punjabi, Tamil, Telugu (www.gnani.ai/text-to-speech/ plus the language) | 4 November 2026 |
| Timbre v3.0 code-switching labelled beta | Gnani.ai | www.gnani.ai/resources/timbre-v3 | 4 November 2026 |
| “40+ Languages” (Prisma), “21+ Languages” (Timbre v2.5, beta) | Gnani.ai | www.gnani.ai/models | 4 November 2026 |
| Cloud, private cloud, on-premise, air-gapped deployment | Gnani.ai | www.gnani.ai/enterprise | 4 November 2026 |
| Agent regions Asia, Americas, Europe | Gnani.ai | docs.gnani.ai/Platform/Get_Regions | 4 November 2026 |
| 402 “Insufficient credit” error | Gnani.ai | docs.gnani.ai/api/introduction/errors | 4 November 2026 |
| “Gnani pricing is quoted per deployment through sales” | Gnani.ai | www.gnani.ai/resources/blogs/gnani-ai-vs-ringg-ai | 4 November 2026 |
| ₹105.9155 and US$1.1179 per euro on 3 Nov 2026, giving our ₹94.75 per US$1 | European Central Bank | www.ecb.europa.eu/stats/policy_and_exchange_rates/euro_reference_exchange_rates/html/index.en.html | 4 November 2026 |
| 22 languages in the Eighth Schedule | Government of India, Ministry of Law and Justice, Legislative Department (The Constitution of India, as on 1 May 2024, pp. 325–326) | The Constitution of India (PDF) | 4 November 2026 |
| 750 characters per spoken minute; 10,000-minute worked month; 200-call test | Zian AI editorial assumption | This page | 4 November 2026 |
| 50,769+ appointments; 2,736% increase in lead contact rates; 30+ languages | Zian AI | zian.ai home page and features page (first-party) | 4 November 2026 |
| Pay-per-result; under 60 seconds, 24/7; 50,769+ appointments | LeadsNow.ai | leadsnow.ai/ | 4 November 2026 |