A business that calls customers across India needs a voice that speaks Hindi, Tamil, or Bengali, and often one that slips into English halfway through a sentence. Sarvam built its text to speech model, Bulbul, for that job, and Sarvam describes its speech, text, and document models as "trained for 22 Indian languages and English."
As of November 2026, Sarvam prices Bulbul in rupees and builds it around Indian languages. Bland Speech bills in dollars per character, and Bland Speech v3 ranked first of 13 AI voices in the blind test Design Arena published on 4 August 2026.
What is Sarvam and how does it work?
Sarvam sells Indian-language models by API, plus Sarvam Voice Agents (Samvaad), a platform for building AI calling agents that Sarvam prices from ₹3 a minute with the models included. Bulbul is the text to speech line inside that catalog, and Sarvam names voice agents, IVR prompts, narration, and audio content as its best fit. Sarvam publishes a customer story about Tata Capital's voice AI.
Which Bulbul models does Sarvam offer?
Sarvam offers three Bulbul models: bulbul:v4-flash, the latest and the one Sarvam recommends for new integrations, bulbul:v3, the stable model that the API uses when a request names no model, and bulbul:v2, which is legacy. Sarvam says V4 Flash takes the same request shape as V3, so an existing integration switches in one line.
Bulbul V3 reached end of life on Sarvam's Voice Agents on 10 November 2026, and stays available on the text to speech API. Voices are styled personas, with names such as simran_en_customer and ishita_hi_devotional that mark a language and a register like customer care, banking, or devotional.
Which languages does Bulbul cover?
Sarvam says Bulbul V3 covers 11 languages, 10 Indian languages plus English. V4 Flash voices cover 10 language tags, and Sarvam states that Malayalam and Odia are not yet offered on V4 Flash. A team that needs one of those two has to stay on V3 for now.
Bulbul accepts code-mixed text, meaning a single passage in English and an Indian language together. Bulbul also wants native script. Sarvam says transliterated input, such as "Aapka order confirm ho gaya hai" typed in Latin letters, significantly reduces output quality.
How does Bulbul stream, and how fast is it?
Sarvam offers three ways to call Bulbul: a REST request, HTTP streaming, and a WebSocket. A WebSocket is a connection that stays open so text can go in and audio can come back without a new request each time. Sarvam's WebSocket has configure, send, flush, and ping messages, and a setting for the minimum number of characters that triggers a flush.
Request sizes are capped at 2,500 characters for REST, 3,500 for HTTP streaming, and 2,500 per WebSocket message, with Sarvam recommending under 500 characters per message for the lowest latency. On speed, Sarvam says "Under 250 ms to first audio in production." Time to first audio is the gap between sending text and hearing the first sound. Bland publishes 199ms to first audio, each figure as that company publishes it and neither a race Bland ran.
What phone audio does Bulbul return, and what does it plug into?
Sarvam offers mu-law and A-law output codecs for streaming, the two 8-bit encodings that phone networks carry, and sample rates from 8,000 Hz up to 48,000 Hz on REST and up to 24,000 Hz on streaming. The default is 22,050 Hz, so a phone line has to ask for 8,000 Hz on purpose.
Sarvam integrates with LiveKit, Pipecat, Vapi, Twilio, and Exotel, and ships SDKs for Python and Node.js. For calling agents, Sarvam says a customer can bring its own Exotel, Twilio, Plivo, or other provider, or rent a number from Sarvam.
How much control does a developer get over the voice?
Sarvam says Bulbul does not support SSML, the markup standard that lets a developer wrap text in tags to set pauses and emphasis. Bulbul V3 and V4 Flash instead take a pace from 0.5 to 2.0, a temperature from 0.01 to 1 with a default of 0.6, and a pronunciation dictionary, and Sarvam says pitch and loudness are not supported on those two models.
What does voice cloning look like?
Sarvam's Instant Voice Cloning is in beta, including streaming over HTTP and WebSocket. Sarvam recommends a reference clip of 10 to 15 seconds, caps an inline clip at 15 seconds and 10 MB, and says the reference language does not have to match the output language. Cloning costs ₹2.00 per 1,000 characters on every pricing tier.
On consent, Sarvam says "Voice cloning requires consent," and its dashboard makes a customer tick a box confirming the rights to the voice before saving it. On 9 November 2026 Sarvam added a detector that checks whether audio carries the Sarvam watermark embedded by Instant Voice Cloning. Custom voices on Sarvam's voice agents are enterprise-only.
How much does Sarvam cost?
Sarvam bills Bulbul V3 at ₹3.00 per 1,000 characters on pay as you go, with the rate falling to ₹2.85 on the Pro plan, ₹2.70 on Growth, and ₹2.60 on Business. The plans cost ₹0, ₹9,999, ₹49,999, and ₹1,99,999 a month, and Enterprise is custom, with custom concurrency and throughput and dedicated support and SLAs. Sarvam's rupee prices exclude GST, and credits are universal across Sarvam's APIs.
Every verified user gets ₹100 in free credits, usable on any Sarvam API and not limited to evaluation. When credits run out, API requests fail with a 402 error.
How is Bland Speech priced?
Bland Speech is $0.04 per 1,000 characters, which is $40 per million, pay as you go with no monthly plan, and a new account gets 50,000 free characters once. Credits do not expire. Enterprise pricing is custom, set with Bland's sales team.
A million characters of Bulbul V3 on pay as you go works out to ₹3,000, against $40 for a million characters of Bland Speech. Sarvam prices in rupees before GST, so a team paying in dollars converts the rupee total before comparing.
Which sounds more human, Sarvam or Bland Speech?
Sarvam wasn't in Design Arena's Audio Realism Bench results of 4 August 2026, so no score places Bulbul next to Bland Speech v3, which ranked first of 13 AI voices at 1365 while the human recording scored 1500. In Design Arena's published results, vetted native speakers heard two recordings of one script and chose the voice that sounded more like a real person, and each model was represented by its most conversational American-English voices.
- #1BlandBland Speech v31365
- 2ndMicrosoftMAI-Voice-21224
- 3rdxAIGrok TTS1156
- 4thOpenAIGPT Realtime 21119
- 5thGoogleGemini 2.5 Pro TTS1109
- Not rankedSarvam's own voicesNot ranked
Ranks are among 13 AI voices, and Sarvam's own voices weren't ranked. Real human recordings scored 1500, at the right edge of this scale. Bars start at 600, not zero. Source: Design Arena's published results.
A team with callers in an Indian language should hear both voices read its own scripts before deciding, since that test scored American English.
The controls differ too. Bland Speech takes performance tags such as [laughs], [sighs], and [say warmly] written right before the words they shape, two numeric controls for expressiveness and stability from 0.0 to 1.0, and timed pauses from 0.1 to 10 seconds. Sarvam offers pace, temperature, and a pronunciation dictionary.
On cloning, Bland Speech makes an instant clone from one clean sample of about 10 seconds and fine-tunes a professional clone on 30 minutes or more of audio after a consent and rights review. Every clone is private to the customer's organization, and the customer confirms the voice is theirs to use.
Is Sarvam secure enough for enterprise calls?
Sarvam says it is ISO 27001:2022 certified and holds a SOC 2 Type II report, and says it provides "Complete data residency in India." Sarvam says ISO 42001, a standard for managing AI systems, is in progress. Retention is configurable, and Sarvam supports an ephemeral mode where "data passes through our systems only for processing and is purged immediately afterwards."
Sarvam offers deployment as managed SaaS, single-tenant on-premises, or fully air-gapped with zero external network dependency.
Bland carries ISO 27001:2022 as well, is SOC 2 Type II audited, and is HIPAA compliant with a signed BAA. Bland was granted FedRAMP 20x Class C at Moderate on 17 November 2026, and Bland states that speech synthesis runs inside the FedRAMP boundary. Bland says zero-retention processing is available for sensitive workflows and that there are zero AI sub-processors in the call path. Enterprise customers get dedicated orchestration and dedicated GPUs run by Bland.
Is Sarvam worth it for enterprises?
For an enterprise phone line, Bland Speech is the better buy: Bland Speech v3 scored 1365 in the blind tests Design Arena published on 4 August 2026, a test Sarvam wasn't in, and Bland bills a published dollar rate per character with no monthly plan to size.
What these rows mean
- Price: Sarvam prices in rupees and excludes GST. Bland Speech prices in dollars.
- Latency: Time to first audio is the gap between a request and the first sound. Each figure is the company's own, measured its own way.
- Phone-ready output: Mu-law is the 8-bit encoding used on phone lines in North America and Japan, and A-law is its counterpart elsewhere. Sarvam's default sample rate is 22,050 Hz.
- Compliance: Bland's statements are company-level, and Bland places speech synthesis inside the FedRAMP boundary.
Sarvam pros and cons
Pros
- Models trained for 22 Indian languages and English, with code-mixed text supported
- India data residency, plus single-tenant on-premises deployment and a fully isolated option with zero external network dependency
- ISO 27001:2022 and a SOC 2 Type II report
- ₹100 in free credits for every verified user, with commercial use allowed
- Mu-law and A-law output, and integrations with LiveKit, Pipecat, Vapi, Twilio, and Exotel
- A watermark and a detector on cloned audio
Cons
- No SSML, and pitch and loudness are not supported on V3 or V4 Flash
- Prices in rupees, before GST
- Malayalam and Odia are not yet on V4 Flash
- Transliterated input in Latin letters lowers output quality
- Instant Voice Cloning is in beta
Why Bland Speech is the Sarvam alternative for enterprise phone calls
Bland Speech carries a dated blind result: v3 ranked first of 13 AI voices in Design Arena's results of 4 August 2026. Bland Speech streams over HTTP chunks and a WebSocket that takes text as a language model writes it, supports barge-in so a caller can interrupt, and returns mu-law at 8 kHz. An OpenAI-compatible endpoint lets a team switch a client by changing a base URL, a key, and a model, Pipecat has a first-party Bland service, and LiveKit Agents has a Python plugin.
Pros
- First of 13 AI voices in Design Arena's results of 4 August 2026, at 1365
- 199ms to first audio, as Bland publishes it
- One standard dollar rate per character, with no monthly plan
- Performance tags, expressiveness and stability controls, and timed pauses
- Mu-law at 8 kHz and a WebSocket with barge-in
- HIPAA compliant with a signed BAA, and speech synthesis inside the FedRAMP boundary
Cons
- More than a hobby project needs, because Bland Speech is built for production calls
Verdict: Sarvam or Bland Speech for enterprise calls?
Bland Speech is the text to speech API to choose for enterprise phone calls, because a buyer gets a dated blind result for the voice and a dollar bill with no monthly plan to size. Sarvam suits a team whose callers speak an Indian language and whose data must stay in India. A Bland Speech vs Sarvam comparison sets the two side by side, Sarvam alternatives covers the wider field, and best text to speech for phone calls ranks the options built for a live line. Bland Speech's text to speech API has the endpoints, the text to speech comparison hub and review index list more reviews, and a team can try Bland Speech with 50,000 free characters.
Frequently asked questions
How much does Sarvam cost?
Sarvam bills Bulbul V3 at ₹3.00 per 1,000 characters on pay as you go, falling to ₹2.60 on the Business plan at ₹1,99,999 a month. Sarvam's rupee prices exclude GST. Bland Speech is $0.04 per 1,000 characters with no monthly plan.
What languages does Sarvam text to speech support?
Sarvam says Bulbul V3 covers 11 languages, 10 Indian languages plus English, and V4 Flash voices cover 10 language tags, without Malayalam and Odia. Bulbul accepts code-mixed English and Indian-language text but loses quality on transliterated text typed in Latin letters.
How fast is Sarvam text to speech?
Sarvam says "Under 250 ms to first audio in production." Bland publishes 199ms to first audio, and each figure is the company's own, measured its own way.
Does Sarvam work for phone calls?
Yes, Sarvam offers mu-law and A-law output for streaming and integrates with Twilio, Exotel, LiveKit, Pipecat, and Vapi. Bland Speech returns mu-law at 8 kHz and has a first-party Pipecat service and a LiveKit Agents plugin.
Does Sarvam clone voices?
Yes, Instant Voice Cloning is in beta, takes a reference clip of 10 to 15 seconds, and costs ₹2.00 per 1,000 characters. Sarvam requires consent and added a watermark detector on 9 November 2026. Bland Speech clones an instant voice from about 10 seconds of audio and fine-tunes a professional clone on 30 minutes or more.
Where does Sarvam keep data?
Sarvam says it provides complete data residency in India and offers managed SaaS, single-tenant on-premises, and air-gapped deployment.
What is the best Sarvam alternative?
Bland Speech is the alternative for enterprise phone calls, with a first-place blind result dated 4 August 2026, one standard dollar rate, and mu-law output. Other options are laid out in Sarvam alternatives.
