Research & Innovation

Pushing the boundaries of Voice AI

Vodex is an AI-native voice automation company building speech intelligence for real-world, high-stakes conversations in mortgage, insurance, and collections.

Talk To Our Expert Schedule a Demo

Debt Collection

Where this work began

The team's background is foundational AI/ML and speech synthesis. Vodex is the product of years spent inside these models, not a wrapper around someone else's.

2020

First experiments in speech synthesis

The team's background is foundational AI/ML and speech synthesis. Vodex is the product of years spent inside these models, not a wrapper around someone else's.

2021

Retrieval and language modeling

Work with FAISS, Haystack-RAG, and BLOOM built the retrieval and language foundations that conversational agents depend on.

Today

An AI-native voice automation company

Vodex builds speech intelligence for real-world, high-stakes conversations in the mortgage, insurance, and collections industries.

Running a collections operation? See Vodex for Debt Collection

Speech synthesis

Why we built our own TTS

Existing text-to-speech fell short on naturalness, expressiveness, and control. Production conversation needs behaviors off-the-shelf models often lack: laughing, sighing, pausing, expressing uncertainty. So we trained our own.

Orpheus core architecture

Our pretrained TTS model is built on the Orpheus core architecture, tuned for production conversation rather than demos.

21,000+ hours of training data

Trained on more than 21,000 hours of diverse, expressive speech, so the model has heard how people actually talk.

Zen-Tokenizer inside

Our proprietary neural audio codec is integrated at the audio tokenization layer, not bolted on afterward.

Narrowband and wideband

Supports both 8kHz telephony pipelines and 16kHz wideband audio from the same stack.

Running a collections operation? See Vodex for Debt Collection

Named Voices

Voices that feel human

Each production voice is a distinct speaker with its own character, trained for a specific kind of conversation.

Shreya

The most soft-spoken and emotionally expressive of our voices. Ideal for empathetic agentic use cases where tone carries the conversation.

Shweta

Trained to sound like a professional audiobook narrator. Smooth, clear storytelling with steady pacing.

Aastha

A latency-optimized voice that preserves expressiveness while minimizing response time. Built for real-time back-and-forth.

Running a collections operation? See Vodex for Debt Collection

Latency

Fast enough to interrupt

A voice agent that pauses too long breaks the illusion. We keep pushing time-to-first-byte down through codebook structure refinement and distributed deployment techniques.

Learn More About Integrations

189ms

Average time-to-first-byte (TTFB) today

<80ms

Target response time in production

Roadmap

What comes next

We draw inspiration from Kyutai (Moshi's full-duplex and MIMI tokenizer work), Snac's neural codecs, Canaophy Labs, Sesame Labs and their CSM stack, and Carson et al. (2025). Agentic voice AI is a long-term mission, and we want company on the road.