Text to Speech for Indian Languages

Before any audio is generated, the input text goes through a normalization stage that expands abbreviations, formats numbers, and resolves ambiguities. In English, this is relatively straightforward. In Indian languages, it gets complicated fast. Consider a sentence like "Dr. Lal PathLabs ka appointment 18 Jan ko 7:30 AM pe hai." The system needs to know that "Dr." should be spoken as "Doctor" (not "D-R"), that "18 Jan" becomes "athaarah January," that "7:30 AM" becomes "saadhe saat baje subah," and that "PathLabs" is a proper noun that should not be translated.

Sarvam's normalizer handles Indian names, addresses (including PIN codes and landmark references), currency amounts in rupees, phone numbers in the Indian 10-digit format, and mixed-script text where Devanagari and Latin characters appear in the same sentence. This preprocessing step is what makes the difference between robotic output and speech that sounds like a real person reading the text aloud.

Prosody is the rhythm, stress, and intonation pattern of speech. It is what makes a question sound like a question and a statement sound like a statement. It is also what separates a natural-sounding voice from a flat, robotic one. Hindi prosody is fundamentally different from English: stress patterns are more evenly distributed across syllables, and the pitch contour of a Hindi sentence follows a different arc than its English translation. When a TTS model trained on English data generates Hindi, it often applies English stress patterns to Hindi words, producing audio that is technically intelligible but sounds off to native speakers.

Bulbul V3 uses an LLM-based text analysis layer that automatically infers where to emphasize, when to pause, and how to modulate tone and pacing — before any audio is generated. The model has separate prosodic representations for each of the 11 supported languages, so Hindi narration has the rising-falling pattern characteristic of North Indian speech, while Tamil output follows the more syllable-timed rhythm of Dravidian languages. This per-language prosody modeling is why Sarvam achieved the highest listener preference in the Josh Talks blind evaluation — 20,000+ votes from 500+ annotators across 11 languages confirmed that Bulbul V3 sounds more natural than ElevenLabs, Cartesia, and other competitors.

An estimated 350 million Indians speak English as a second language, and the vast majority of them mix it freely with their primary language. A customer support agent might say "Aapka order dispatch ho gaya hai, expected delivery by tomorrow evening." A teacher might explain "Friction ka coefficient zyada matlab zyada resistance, simple." This is not broken grammar. It is normal Indian communication, and any TTS system that cannot handle it will produce jarring, unnatural output every time a language boundary appears.

Sarvam handles code-switching at the model level, not through a pipeline that detects language boundaries and switches between separate engines. When Bulbul V3 encounters a Hinglish sentence, it generates audio in a single pass with smooth transitions between Hindi and English segments. No pauses, no voice quality changes, no accent shifts. The same applies to Tanglish, Benglish, Marathi-English, and other mixed-language patterns. This is particularly valuable for voice agents and IVR systems where callers naturally speak in mixed languages.

Sarvam offers 35+ distinct voices across all 11 supported languages, sourced from professional voice artists and available at up to 48 kHz full-band quality. Voices are organized into categories — Edtech, Customer Support, Advertising & Announcements, Storytelling & Narration, and Social Media Content — so you can pick a voice that matches your use case rather than browsing a generic list. Each voice has a unique personality: Arjun works well for authoritative banking communications, Meera suits warm customer service interactions, and Ritu brings energy to expressive storytelling. Beyond voice selection, you can adjust pace (0.5x to 2x), pitch, and emotional expressiveness through simple API parameters. For enterprises that need a branded voice, Sarvam offers consent-based voice cloning with built-in safeguards — provide a short speech sample (30-60 seconds), give explicit consent, and the system creates a custom voice that maintains Bulbul V3's natural expressiveness. Browse the full catalog at /text-to-speech/voices.

Voice agents are replacing traditional call center recordings with dynamic, context-aware speech. A banking voice agent can read back a customer's account balance in Hinglish — "Aapka EMI payment of rupees 12,345 due hai by 15th March" — without pre-recording any of those specific sentences. In BFSI deployments, Bulbul V3 handles loan collection calls with financial terminology (EMI, credit records, late charges) across Hindi, Kannada, and other languages. Healthcare voice agents confirm appointments with complex medical terms like "Comprehensive Thyroid Profile with Anti-TPO Antibodies test" without mispronouncing them. Sarvam's sub-250ms streaming latency means callers hear responses without awkward pauses.

IVR systems use TTS to generate dynamic menu prompts that change based on context. Instead of maintaining thousands of pre-recorded audio files for every possible menu option in every language, telecom and banking companies generate prompts on the fly. A single API call creates "Aapka current balance hai rupees 12,345. PIN change karne ke liye 1 dabayein" in natural Hindi.

Voice notifications deliver OTP codes, appointment reminders, delivery updates, and payment confirmations as spoken calls. For the hundreds of millions of Indian users who are more comfortable with voice than text, spoken notifications have significantly higher engagement than SMS. TTS makes it possible to personalize every call without recording each message individually.

YouTube creators in India are using TTS to produce videos faster. Educational channels, news aggregators, and storytelling accounts generate Hindi, Tamil, or Telugu narration from scripts without needing a recording studio. A creator who publishes daily can write a script and have broadcast-ready audio in seconds.

Podcast production becomes accessible to anyone with a script. Writers, journalists, and educators can turn articles into audio episodes in any Indian language. The AI voices are natural enough for extended listening, which matters for podcast formats where listeners spend 20-40 minutes with a single voice.

Audiobook creation at scale is now feasible for Indian language publishers. Recording a full audiobook traditionally takes weeks and costs lakhs. With TTS, a 300-page book can be converted to audio in hours. Sarvam's expressive voices with emotion control produce audiobooks that listeners actually enjoy, not the flat robotic output that gave early TTS audiobooks a bad reputation.

Voiceover for explainer videos, product demos, corporate presentations, and documentary narration can be generated in 11 languages from a single script. Production teams that previously needed separate voice artists for each language can now localize content in minutes.

Dubbing and localization teams use TTS to create first-draft voiceovers for video content that needs to reach multiple Indian language audiences. A marketing video produced in English can be localized to Hindi, Tamil, and Telugu in minutes. Professional dubbing studios use TTS as a reference track; content teams with smaller budgets use it as the final output.

E-learning platforms use TTS to narrate courses in regional languages. India's National Education Policy emphasizes mother-tongue instruction, and TTS makes it economically viable to offer the same course in 11 languages without recording 11 separate voiceover tracks. Students retain more information when they learn in their first language.

Corporate training content reaches a wider workforce when narrated in regional languages. A compliance training module for a bank with branches across India can be generated in Hindi, Marathi, Tamil, and Bengali from a single script. Updates to policies or procedures are reflected instantly without re-recording.

Presentations with embedded narration are more engaging than slide decks alone. Sales teams, trainers, and educators add TTS narration to their slides so the content can be consumed asynchronously without a live presenter.

Advertising teams produce radio spots, digital audio ads, and video ad voiceovers at scale. A national campaign that needs to run in 8 languages can generate all the audio variants from a single script, test different voices and tones, and iterate in hours instead of weeks.

Accessibility is one of the most important applications of text to speech. For visually impaired users, TTS enables access to websites, documents, and digital services. For users with reading difficulties, it provides an alternative way to consume written content. India has over 60 million people with visual impairments and hundreds of millions who are more comfortable with spoken information than written text. Sarvam TTS supports screen reader integration and can narrate content in the user's preferred Indian language at adjustable speeds.

Sign up at sarvam.ai/try/tts-api to get your API key. Install the Python SDK with pip install sarvamai or the Node.js SDK with npm install sarvamai. Both SDKs use an OpenAI-compatible interface, so if you have integrated any LLM API before, the pattern will feel familiar. Your first TTS generation takes under 5 minutes from signup to working audio.