Kahani: AI-Native, Creator-First Audio Storytelling

Kahani is the first true AI-native, creator-first audio storytelling platform for India.
We help creators go from text → full, regional-language audio stories — in Hindi, Punjabi, English, Marathi, and Telugu — in seconds, unlocking the next chapter of India’s audio economy.

The Opportunity

India’s audio storytelling market has grown rapidly — fueled by platforms like PocketFM (a portfolio company of Lightspeed Ventures).
However, content creation is still locked into studio-heavy, curated production.
Millions of creators across India, especially in Tier-2 and Tier-3 cities, don’t have access to scalable tools that let them produce high-quality audio in their own language, at their own pace.

Why Now?

Two key shifts make this the perfect time:

1. AI for regional languages is finally production-grade — powered by players like SarvamAI.
2. India’s creator economy is booming — with 200M+ creators, mostly text- and video-first.
   Audio remains an untapped layer that more and more creators, especially from Tier-2/3 cities, are eager to explore.

Our Solution

Kahani solves this by being an AI-native, creator-first tool where anyone can turn a simple text prompt into a 2–10 minute rich audio story — all in their own language — with no studio, no mic, and no wait.
It’s like “Midjourney for audio”, already live in 5 languages and built to scale across India.

Impact

With Kahani:

• Audio storytelling is accessible to all — removing technical, cost, and time barriers.
• Creators can easily generate and iterate on stories in their native language.
• The audio OTT market is projected to grow at 34% CAGR, crossing $1.5B by 2027.
• Over 200M+ monthly active audio consumers in India offer a massive opportunity.
• Kahani is well-positioned to build an independent platform and also complement existing players, creating new synergies across the ecosystem.

Reference:
Grand View Research: Audio Streaming Market Outlook, India

Core Technical Challenges

Building this system introduced several complex challenges, especially around multi-agent coordination and the current capabilities of the voice models we employed.

1. Extensive Prompt Engineering

A significant part of our process involved deep prompt refinement to enable seamless multi-agent collaboration.

Key Details:

• Each agent (e.g., narrator, character 1, character 2, etc.) requires very carefully constructed prompts.
• Poorly tuned prompts can cause agents to produce dialogues that feel disjointed or off-tone.
• The refinement cycle involved several iterations, spanning multiple hours of trial and error.
• Achieving prompt accuracy was crucial for allowing agents to align properly on context and narrative style.

Outcome:
By spending the time to perfect these prompts, we ensured that each agent could “stay in character” and produce high-quality output that enhances the listener’s experience.

2. Sequential Multi-Agent Execution & Context Limits

Our design adopted a sequential multi-agent architecture, where one agent passes its output as context to the next. However, this led to serious context-window constraints.

Challenges:

• Every agent call consumed valuable context length, quickly pushing us toward token limits.
• Degrading the output with each iteration became a major issue — the longer the conversation went, the more context was lost.

Solution Strategies:

• Developed novel context management techniques to minimize token usage without losing key story details.
• Implemented parallelization where possible, allowing some agents to work concurrently and then merge their contributions.
• These optimizations preserved both consistency of output and processing speed, even as the dialogues grew in length and complexity.

3. Expressiveness in Synthetic Voices

Our use case — audio entertainment — places a premium on expressive and dynamic speech. However, our chosen TTS model posed some hurdles.

Current TTS Model:

• Sarvam's Bulbul TTS model was originally tuned for voice agent or support scenarios.
• Its speech style is fairly flat, lacking the rich expressiveness we want for entertainment use cases.

Workarounds Implemented:

• Introduced explicit prompts for pacing, pitch, and loudness to mimic a more dramatic and emotional voice.
• Tuned dialogues to add pauses, emphasis, and tonal variations.
• This yielded a subtle but meaningful improvement in perceived expressiveness.

Future Plans:

• Discussed these requirements with the Sarvam team, who acknowledged this gap.
• They assured us that future models will better support expressive audio for creative industries — a very promising direction for Kahani.