We have raised $7.5M to help grow your business
Answrable
Sign up
← BLOG/ENGINEERING

Improving neural speech cache & full-duplex interruption handling

Benchmarking sub-180ms conversational turn reflexes with Answrable HCM-1 8B parameters.

Audio Research Lab
Audio Research Lab—Speech Modeling
SEP 09, 2026
Answrable
cache compute

Improving neural speech cache on Answrable, Part 1

How our token-level acoustic caching architecture minimizes time-to-first-audio-packet (TTFB) to 142ms while maintaining high fidelity and instant interruption yields.

Agents are the ones building now. They need voice primitives.

Until now, building an AI voice agent required cobbling together disparate vendors: a speech-to-text transcriber, an external LLM completion API, a text-to-speech voice synthesizer, and a legacy SIP trunk aggregator.

This chained architecture created 2,000ms+ roundtrip latencies, robotic speech turn jitter, and clipped conversations whenever callers tried to interrupt. By engineering our own 8-billion parameter foundation model (Answrable HCM-1), we achieved native full-duplex speech-to-speech comprehension with sub-180ms reflexes.

agent-workflow.tsAnswrable HCM-1 8B
import { AnswrableAgent, tool } from "@answrable/core";

export const clinicReceptionist = new AnswrableAgent({
  model: "answrable-hcm-1-8b",
  temperature: 0.2,
  turnReflexLatencyMs: 180,
  systemPrompt: "You are the autonomous voice receptionist for CareClinic...",
  tools: [
    tool({
      name: "verifyInsuranceAndBookSlot",
      description: "Checks real-time calendar availability and schedules patient",
      parameters: { patientId: "string", desiredTime: "string" },
      execute: async ({ patientId, desiredTime }) => {
        return await db.appointments.create({ patientId, slot: desiredTime });
      }
    })
  ]
});

What's included in Answrable GA

HCM-1 Foundation Model

8B parameter speech-to-speech weights with human-grade turn taking.

Direct Carrier SIP

Connect your own carrier numbers with zero transcoding overhead.

With full-duplex interruption handling, callers can speak naturally, pause to think, or change topics mid-sentence without robotic clipping.

What's next for voice telephony

We are expanding our regional edge points of presence across North America, Europe, and Asia Pacific to guarantee sub-50ms carrier jitter worldwide.

In the meantime, we want to see what you build with these tools. Tag us on X, send us feedback, and we're in Discord too.
▶MORE FROM ANSWRABLE