README.md

---
title: IOTEverythin
sdk: static
pinned: false
---

# IOTEverythin

We build small, production-oriented AI that runs where the data lives. Two
independent product lines: compact Indian-English text-to-speech for
conversational agents, and grounded question answering that quotes your
documents instead of hallucinating.

---

# Voice / Text-to-Speech

Compact, production-oriented text-to-speech voices for conversational AI, with a
focus on Indian-English for customer-support and website voice assistants.

## VozVox

These voices power [VozVox](https://www.vozvox.com/), our voice agent platform for
AI-powered conversational agents across phone and web, with natural-sounding speech.

## The Roxi TTS line

A family of Indian-English voices fine-tuned from the open MOSS-TTS models. Two
tiers: the 0.1B models run in real time for live agents, and the 1.7B model is for
pre-rendered or premium-quality audio.

| Model | Base | Best for |
|---|---|---|
| roxi-tts-pro | MOSS-TTS-Local 1.7B | Highest quality and intelligibility, offline or premium |
| roxi-tts-v3.1 | MOSS-TTS-Nano 0.1B | Real-time, current best small voice |
| roxi-tts-v3 | MOSS-TTS-Nano 0.1B | Earlier single-speaker voice |
| roxi-tts-v2 | MOSS-TTS-Nano 0.1B | First Roxi release |
| roxi-tts-v2-onnx | ONNX build of v2 | CPU inference, no transformers dependency |
| voxi-tts | MOSS-TTS-Nano 0.1B | Original prototype voice |

**Focus:** Indian-English accent, natural and telephony-aware; small and fast;
built on commercially permissive Apache-2.0 base models; single-speaker branded
voices for support calls and website assistants.

**Attribution:** Models are built on MOSS-TTS (Apache-2.0). Training data includes
the IIT-Madras Indic TTS English set, used with the required copyright notice shown
on each model card.

---

# Grounded QA

[grounded-pointer-qa](https://huggingface.co/IOTEverythin/grounded-pointer-qa) is an
extractive question-answering model that **cannot hallucinate by construction**: it
only quotes verbatim spans from your documents, abstains when the answer is not
there, and decodes deterministically. Knowledge is hot-swappable: point it at a
folder of `.txt`/`.md`/`.pdf` files, with no retraining.

Built on roberta-base (125M), it runs in milliseconds on consumer hardware: 74.6 EM
on SQuAD v2 in a full-retrieval setting, and 91.7% answer precision in its "right or
silent" mode. A natural grounding layer for support agents that must quote policy
documents instead of improvising.

**Focus:** grounded, verifiable answers; every answer a verbatim quote with a
source, or an honest abstention; local-first and deterministic.