Cutting Voice AI Cost Per Minute from Rs. 6 to Rs. 1.2

When I joined Justdial's AI Initiative, cascading Voice AI cost sat at Rs. 6/min — well above what a vendor like Squadstack charged at Rs. 2.5/min. Getting to Rs. 1.2/min without losing accuracy meant treating the pipeline as one system, not three vendor API calls stitched together.

Where the cost actually lives

A cascading pipeline (STT → LLM → TTS) burns money in places that aren't obvious from a pricing page:

1. Redundant STT calls from retry logic that doesn't know the difference between "the model was wrong" and "the network hiccuped."
2. Full-context LLM calls for turns that need only a small slice of conversation history.
3. Off-the-shelf TTS priced for general-purpose voices when call-center audio only ever needs a handful of consistent personas.

What moved the number

Fine-tuning IndicF5 TTS on our own Justdial call-center audio replaced a per-call vendor TTS cost with a fixed compute cost on owned infrastructure, and gave us voices that actually sounded like our brand instead of a generic assistant.

On the STT/LLM side, trimming context aggressively per turn and caching intent classification for repeat query patterns cut the most expensive calls without touching accuracy — in some cases improving it, because a smaller, well-scoped prompt hallucinates less than a bloated one.

The part that's easy to miss

Cost and latency optimization are the same project. Every trick that cuts round-trip cost also cuts the 400ms budget between a caller finishing a sentence and hearing a response. Optimize for one and you mostly get the other for free — which is the only reason this was possible in a month, not a quarter.