Sarvam AI Ships Bulbul V4 to Bring Emotional Voice to 11 Indian Languages

Sarvam AI just shipped Bulbul V4, the latest version of its text-to-speech model built specifically for Indian languages. The improvements center on richer emotion, more natural expression, and a wider vocal range, building directly on a V3 architecture that already beat ElevenLabs and Cartesia in independent blind listening tests.

Why Indian TTS is a different problem

Building TTS for India means handling speech that switches languages mid-sentence, varies by region, and mixes scripts. Bulbul handles code-mixed text like Hinglish, number normalization, and natural prosody out of the box. The global model providers do English well and a handful of European languages adequately. For a farmer in Maharashtra asking a government AI about crop subsidies in Marathi, or a patient in Tamil Nadu trying to understand a prescription in Tamil, those models fall short. Bulbul V4 is Sarvam's answer to that gap.

What V3 established

Bulbul V3 uses an LLM to analyze text and infer prosodic elements: emphasis, pauses, tone, and pacing. Prosody is the rhythm, stress, and intonation that makes a question sound like a question rather than a statement. By understanding context and intent rather than treating words as a flat sequence, V3 generates speech that aligns with the emotional content of what is being said.

In an independent blind listening study by Josh Talks AI covering 11 Indian languages and more than 500 human annotators, V3 ranked highest in listener preference for 8 kHz telephony-grade audio and showed lower error rates for skipped words and mispronunciations than ElevenLabs and Cartesia Sonic-3. V4 pushes that baseline further.

What V4 adds

Sarvam describes V4 as delivering richer emotion, more natural expression, and greater vocal range. The key changes:

• More expressive rendering across the existing voice catalog of 30-plus professional voices recorded by trained voice artists across 11 Indian languages.
• Sub-250ms streaming latency and emotion control, now with a wider dynamic range across vocal styles.
• Consent-based voice cloning for branded voices: provide a 30 to 60 second speech sample, give explicit consent, and the system creates a custom voice.

Languages and voices

The model supports 11 languages with BCP-47 codes:

• Hindi (hi-IN)
• Bengali (bn-IN)
• Tamil (ta-IN)
• Telugu (te-IN)
• Gujarati (gu-IN)
• Kannada (kn-IN)
• Malayalam (ml-IN)
• Marathi (mr-IN)
• Punjabi (pa-IN)
• Odia (od-IN)
• English (en-IN)

The voice catalog includes 39 named voices: Shubh, Aditya, Ritu, Priya, Neha, Rahul, Pooja, Rohan, Simran, Kavya, Amit, Dev, Ishita, Shreya, and more. Speech speed is adjustable from 0.5x to 2.0x.

Where it performs best

The use cases that benefit most from better emotion and vocal range are the ones India's voice-first economy runs on:

• Voice agents and call centers: A banking agent can read back an account balance in Hinglish without pre-recording specific sentences.
• EdTech: AI tutors explain concepts in a student's native language. Emotional expressiveness keeps learners engaged over long sessions.
• BFSI collections: Bulbul handles loan collection calls with financial terminology (EMI, credit records, late charges) across Hindi, Kannada, and other languages.
• Audiobooks and long-form content: Richer vocal range sustains attention across extended narration, not just short demos.
• Gaming: Characters speak back to players in their own language.

How to use it

The API accepts REST or WebSocket calls. Python quickstart:

from sarvamai import SarvamAI
from sarvamai.play import save
client = SarvamAI(api_subscription_key="YOUR_API_KEY")
response = client.text_to_speech.convert(
    text="नमस्ते, कैसे हैं आप?",
    target_language_code="hi-IN",
    model="bulbul:v4",
    speaker="anushka",
)
save(response, "output.wav")

Two modes are available: a REST API for conversions up to 2,500 characters, and a Streaming API for real-time audio via HTTP stream or WebSocket. The streaming path is the right choice for live voice agents where latency is user-facing. Supported output formats include WAV, MP3, Linear16, Mulaw, Alaw, Opus, FLAC, and AAC.

Pricing

All plans include ₹1,000 in free credits. Bulbul V3 is priced at ₹30 per 10,000 characters. V4 pricing has not been separately listed yet; expect it to follow the same structure or carry a slight premium. Enterprise volume discounts are available on request.

The broader stack

Sarvam AI is building a full product layer spanning chat LLMs, speech recognition, text-to-speech, translation, and document intelligence, all optimized for how Indian languages are actually used: native script, romanized WhatsApp Hindi, code-mixed Hinglish, and 22 scheduled languages. Bulbul V4 sits within that stack as the layer most users will encounter first.

Getting words right was the V3 story. V4 focuses on getting feeling right, and for a country where voice is the primary interface for hundreds of millions of people, that shift carries real weight in markets where English-first models have never had to think about it.