You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Laya voice router

convaiinnovations/laya (multilingual, mmBERT-base) fine-tuned to decide, from a caller's partial speech-to-text and the conversation so far, which read-only lookups a voice agent is about to need and with which closed-set arguments. English, Hinglish, Hindi. Load with laya.load("this-repo").

Training mix: template-generated voice-agent turns over six business domains (complete, two-in-one, mid-sentence, split-across-turns, small talk; ASR-style noise), MASSIVE intents (Hindi, English, six Indic languages), LocalLLaMA/typed-decisions, no real call data yet.

	held-out domain, per question	held-out domain, all questions right	7 production routing cases	latency (median)
base	0.636	0.107	3/7	32.9 ms
fine-tuned	0.962	0.817	7/7	33.9 ms

The held-out domain shares its sentence templates' style with the training domains, so treat that column as an upper bound; performance on real callers is only established by shadow-testing against live traffic.

Downloads last month: -; Downloads are not tracked for this model. How to track

Safetensors

Model size

0.3B params

Tensor type

F16

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for IAMCB/laya-voice-router

Base model

convaiinnovations/laya

Finetuned

(106)

this model