What the models actually do
Five bounded tasks that make up most of a regulated call. Each is small enough to run locally and specific enough to benchmark.
EchoLeads · voice stack
Hinglish bench
One sentence, one use case, one model pairing, one price. Add a line the way a caller would actually say it, pick the pairing, and read the cost of speaking it. The tuned model is named for the use case it was trained on.
Speak the line and it lands in the box — edit before you run it.
Hindi tokenEnglish tokenTTS share of costWord tagging is a lookup, not a classifier — treat it as a reading aid.
Rates and assumptions
LLM cost assumes a reply about the length of the input line, billed on output tokens at roughly four characters per token. TTS is billed per character spoken. Edit any rate to reprice every run.
- Multilingual & script fidelityCode-switch handling · correct script per language run
- Objection handlingClassify the pushback · respond in approved language
- Compliance guardCheck before speaking · log the verdict
- Call structuringSchema-valid output into your CRM
- Voice latencyThe engineering that makes the other four fit in a turn
Each capability is trained per industry. The insurance objection model and the real estate objection model share architecture and nothing else.