Sarvam Bulbul v3 TTS

Sarvam Bulbul v3 TTS — Text-to-Speech for Indian Languages

What is Sarvam Bulbul v3 TTS?

Sarvam Bulbul v3 TTS is a text-to-speech model from Sarvam AI that turns written text into natural, expressive speech across Indian languages. It covers 11 languages — 10 Indian languages plus Indian-accented English — with a library of 37 studio-recorded speaker voices. Rather than reading text as a flat sequence, Bulbul v3 is built on an LLM-based text-analysis layer that infers emphasis, pauses, tone, and pacing from context, so the audio matches the intent of what is being said. That makes it a strong fit for Indian speech, where people switch languages mid-sentence, accents vary by region, and names, numbers, and abbreviations have to be pronounced correctly. On Segmind you call it with a single synchronous API request and get audio back directly.

Key Features

• •11 languages: Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, Odia, and Indian English.
• •37 speaker voices covering male and female timbres, each able to voice every supported language.
• •Code-mixed input (Hinglish, Tanglish) handled natively when Indic words are written in native script.
• •Up to 2,500 characters per request for long-form narration and content.
• •pace control from 0.5x to 2.0x and a temperature control for expressiveness.
• •Eight output codecs (wav, mp3, aac, flac, opus, linear16, mulaw, alaw) and sample rates from 8 kHz telephony up to 48 kHz studio audio.

Best Use Cases

Bulbul v3 was designed for production voice systems. It is well suited to voice agents and IVR prompts, BFSI call-center flows that read out EMIs and account details, EdTech tutors explaining concepts in a student's native language, audiobook and YouTube narration, gaming characters, and accessibility or government-service announcements. In Sarvam's published listener-preference study it ranked highest for 8 kHz telephony audio, which makes it a practical default for call-center and voice-agent workloads. For studio-grade output, use the higher 44.1 kHz or 48 kHz sample rates.

Prompt Tips and Output Quality

Always write Indian-language words in their native script — romanized Indic text noticeably degrades quality. Code-mixed English inside native-script sentences is fine. Set language_code to the script's actual language so Hindi and Marathi (both Devanagari) are not read in the wrong language. Comma-format long numbers for accurate reading, keep pace near 1.0 for natural delivery, and raise temperature for more emotive lines or lower it for consistent, predictable narration. Bulbul v3 has no SSML support, so split text at natural pause points for coarse control.

FAQs

Which languages does Sarvam Bulbul v3 TTS support? Eleven: Hindi, Bengali, Tamil, Telugu, Kannada, Malayalam, Marathi, Gujarati, Punjabi, Odia, and Indian English.

How many voices are available? 37 speaker voices, all recorded by professional voice artists, and every speaker can voice all eleven languages.

Does it handle Hinglish and code-mixed text? Yes. Mixing English words into native-script Indic sentences works well; avoid fully romanized Indic input.

What is the maximum input length? Up to 2,500 characters per request.

Which sample rate should I use? Use 8 kHz for telephony, the 24 kHz default for general use, and 44.1 kHz or 48 kHz for studio-quality output.

Does it support SSML? No. Use the pace parameter and sentence breaks for prosody control instead of SSML tags.