Introduction

On October 14, Gnani AI launched Timbre v3.0, a closed text-to-speech model for Indian languages that adds inline audio tags — short cues like <laugh>, <sigh>, <whispers>, and <pauses> that steer how a line is delivered instead of being spoken aloud. It is live today in the Gnani app / playground and on the Vachana TTS API as model id timbre-v3.0. There are no open weights.

This sits next to today’s Sarvam Bulbul V4 Flash story rather than replacing it. Bulbul Flash is about sub-250 ms speed and price for Indian voice agents. Timbre v3.0’s headline is expressive direction: the same voice can laugh, whisper, or hold a beat mid-sentence across 10 languages, with first audio claimed under 150–200 ms and streaming playback.

Primary sources: the Gnani launch blog (October 14), the Timbre v3 product page, and TTS inference docs. Vendor X also posted the day-3 launch clip the same afternoon.

What Shipped

  • Model id: timbre-v3.0 (alongside still-available timbre-v2.5 for everyday agent calls without tags).
  • Voices: 13 production voices (examples: Shivani, Avinash, Lakshmi, Kalpana, Shubham, Prakash, Devansh, Yashraj, Chetana, Vasudha, Kavitha, Riyaa, Tarika). Each voice speaks all 10 supported languages.
  • Languages: Hindi, English, Tamil, Telugu, Kannada, Malayalam, Marathi, Punjabi, Bengali, Gujarati — plus auto and Hinglish (hi-en) routing in the API.
  • Audio tags (13): <laugh>, <sigh>, <whispers>, <pauses>, <coughs>, <excited>, <exhales>, <giggle>, <shouting>, <slow>, <sniffles>, <yawns>, <clears throat> — write them in angle brackets at the moment you want the effect.
  • Latency: product page markets under 150 ms time to first byte; the blog and FAQ say under 200 ms with streaming before synthesis finishes. Treat both as vendor claims, not independent benches.
  • Surfaces today: Gnani app / playground; REST POST https://api.vachana.ai/api/v1/tts/inference; streaming and realtime WebSocket paths in the same stack; Python SDK gnani-vachana. Pipecat and LiveKit are promised later, not day-0.
  • Free tier: product page offers 2,500 free TTS minutes, no card required.

How Audio Tags Work

You keep ordinary punctuation for base delivery. Timbre v3.0 already adapts tone from meaning and punctuation; tags only pin the moments that need direction. Place a tag at the start of a line or between clauses so it shapes the words that follow — do not park a tag at the end of a line with nothing after it. Gnani’s own guidance: one or two tags per segment, avoid stacking opposite cues (whisper + shout), and keep laughter out of collections / fraud / claims scripts until compliance reviews them.

A paste-ready shape from the OpenAPI text field:

Let me check. <pauses> Found it <laugh>

API callers pass that string as text with model: "timbre-v3.0" and a v3 voice (for example Shivani or Avinash). Voices are not shared across models — a v2.5 voice will not work on v3.0.

Mid-sentence code-switching (Indian language ↔ English without an accent seam) and domain number reading (cards, policy IDs, PIN codes, Indian numbering) are also part of the v3 pitch. Those matter for BFSI, insurance, and healthcare agents as much as the tags do.

Where Creators Fit

Beyond contact-center agents, Gnani markets the same tags to video editors, freelancers, audiobook / podcast producers, course creators, and game dialogue. Editing a tag can replace a re-record when you only need a different take of the same line. That is still closed API audio, not a local Comfy or HF workflow.

Conclusion

If you need Indian-language TTS with explicit expressive control today, Timbre v3.0 is a same-day API and playground drop with a clear model id, voice list, and tag vocabulary. If you only need fast, cheap agent speech without tags, Bulbul V4 Flash (and Timbre v2.5) remain the closer fits. Open-weights readers: skip this one and keep watching HF / Comfy for base-model drops.

—Anabel ♡