MetaVoice provides on‑premise voice agents that can listen and speak simultaneously, enabling natural conversations despite interruptions, overlapping speech, or background noise. Its single 350 ms model handles full‑duplex interaction, understands tone and context, and allows developers to define custom workflows without assembling separate ASR, TTS, or dialogue components. This reduces call drop‑off and improves conversion rates.
Funding
Funding not disclosed
Founders
Product
Problem
Voice agents typically operate in a turn‑taking mode, requiring separate speech‑to‑text, text‑to‑speech, and dialogue components. This fragmented architecture cannot handle interruptions, overlapping speech, or background noise, leading to awkward pauses and high abandonment rates—over 40% of callers hang up within 30 seconds.
Solution
MetaVoice delivers a single, full‑duplex conversational AI model that can listen and speak simultaneously, preserving natural flow even when callers interrupt, talk over the system, or speak amid background sounds. The model processes audio end‑to‑end in about 350 ms, eliminating the need for separate ASR, TTS, turn‑detection, or dialogue frameworks. Users can define custom workflows, tools, and personalities while retaining full control over the conversation. Deployable on‑premises, MetaVoice enables real‑time guardrails to filter or block responses before they are heard and provides deep debugging access to each turn’s reasoning, tool calls, and audio output. This architecture reduces call abandonment and improves conversion by maintaining fluid, human‑like interactions.
Target Audience
MetaVoice is aimed at enterprises and contact‑center operators that run voice‑based customer interactions and need a reliable, on‑premises solution to improve engagement and conversion rates.
Features
- Full‑duplex processing that listens while speaking, handling interruptions, overlap, and background voices without breaking the call
- Single end‑to‑end model delivering speech recognition, language understanding, response generation, and speech synthesis in ~350 ms
- On‑premises deployment for data privacy, low latency, and integration with existing telephony infrastructure
- Built‑in guardrails allowing real‑time filtering or blocking of AI responses before they reach the caller
- Debuggable architecture exposing reasoning, tool calls, and audio for each turn in both text and waveform form
- No requirement to assemble separate ASR, TTS, turn‑detection, or dialogue management components