Services
About
Portfolio
Careers
Blog
Get in touch
Data ManagementAI DevelopmentProduct Discovery

Real-time AI voice avatar with ~1.5s perceived latency

For a brand experience agency we built a voice AI avatar in three weeks: about 1s speech-to-text, 0.5s text-to-speech, ~1.5s perceived reply.

Conversational AI avatar: real-time voice proof of concept

_>About

A brand experience agency asked us to build a real-time voice avatar to represent its client’s product brand on a website. Visitors speak, the avatar answers out loud, and the delay has to feel like talking to a person.

The first phase was a proof of concept: show that the avatar can be embedded in a website and answer questions in real time.

Conversational AI avatar: real-time voice proof of concept - About - Image 1

>_Challenge

Latency was the main problem. Speech-to-text, a language model and text-to-speech all add delay, and a pause of a few seconds breaks the illusion of conversation.

Audio also had to stream in and out between browser and backend without gaps. A later 3D avatar with a moving mouth and eyes would add its own difficulties.

Conversational AI avatar: real-time voice proof of concept - Challenge - Image 1

>>>Process

We had three weeks to deliver a working proof of concept: a 2D, non-animated avatar that answers questions with audio. The team was two backend AI engineers and a project manager, with frontend engineers brought in as needed.

The backend engineers built a streaming architecture that processes audio input and text output in real time. They picked the fastest component for each step of the pipeline and kept testing new releases, ready to swap in anything faster.

To cut perceived latency, they stream a short pre-recorded filler such as “Hmm, let me think” while the language model works. The backend also handles format conversion, assuming PCM audio at 16kHz, so audio flows both ways without gaps.

On the frontend, engineers captured audio in the browser, sent it to the backend with minimal delay, and added authentication so only authorized users could reach the system.

__>Solution

In three weeks the proof of concept transcribed speech in about 1 second, streamed spoken answers back in about 0.5 seconds, and brought perceived latency down to around 1.5 seconds with the filler approach.

For answers, we paired Claude 3 Haiku, which is built for fast responses, with GPT-4 for harder reasoning. The avatar answers quickly when it can and thinks longer when it must.

The client continued the project after the proof of concept. Next on the roadmap: UX techniques beyond fillers to hide latency, conversation memory, knowledge grounding and moderation, and a more resilient audio streaming platform.

If you are weighing a voice interface, this is the budget to design for: roughly a second for each hop, and a plan for what the user hears while the model is thinking.

__> Technologies & Expertise

PythonTypeScriptReact
AI DevelopmentProduct DiscoveryWeb Development

Related projects

_> Projects where we did this work

See all projects
Generative AI meeting insights app built in 6 weeks
Playing

Data Management · AI Development · Web Development

Generative AI meeting insights app built in 6 weeks

In 6 weeks we built a generative AI prototype that turns meeting transcripts into summaries, action items and sentiment, with a chatbot for questions.

View project
mTab Marketplace: data marketplace for market research
Playing

Data Management · Product Design · Product Discovery

mTab Marketplace: data marketplace for market research

Since 2020 we've built mTab Marketplace on Django, React and AWS, so clients buy market research studies without mTab installing each one by hand.

View project
Tessellate: anonymized hiring platform for designers
Playing

Data Management · Product Design · Product Discovery

Tessellate: anonymized hiring platform for designers

We built Tessellate's matching engine, which pairs architects and designers with clients and hides gender and appearance in early hiring stages.

View project

>>>Let's build something great_

If you are working on something similar, send us a short brief. A person on our team reads it and replies within one business day.

Apptension site footer

Taking new projects

Have a product to build, fix or scale?

Apptension is a senior-only software development company headquartered in Poznań, Poland. Since 2012 our engineers have shipped 360+ AI and SaaS products for startups, scale-ups and enterprises across the US, UK and EU: agentic MVPs, RAG and voice AI systems, multi-tenant SaaS and legacy modernization.

Projects shipped
360+
Founded
2012
Senior engineers
30+
Client CSAT
100%

Send a few lines about your product. A person from our team replies within one business day.

Start a projectScope it in Scope Studio
  • Emailcontact@apptension.com
  • Phone+48 533 352 052Poznań, Poland

  • Agentic MVP developmentNew
  • Agentic SDLC implementation
  • Product & SaaS development
  • Legacy modernization
  • Generative AI development
  • SaaS development company
  • Vibe-coded prototype to production
  • End-to-end software development
  • Senior engineering team extension
  • Product consulting & discovery
  • Tech due diligence for investors
  • All services

  • LLM application engineering
  • RAG & retrieval engineering
  • Voice AI agent development
  • MCP server development
  • LangGraph & LangChain agents
  • AI agent evaluation & observability
  • Node.js development company
  • Python & Django development
  • React & Next.js development
  • React Native app development
  • All expertise

  • Open-source SaaS Boilerplate
  • SaaS P&L model template
  • Walkthrough Studio
  • Agentic SDLC playbook
  • AI SDLC maturity assessment
  • Agentic SDLC ROI calculator
  • Founder resources
  • Events & exec dinners
  • Open source
  • Builder's Toolkit

  • Boutique vs large software house
  • Build vs buy vs partner
  • How to choose an AI & RAG partner
  • Team augmentation without chaos
  • Django vs Node.js for SaaS
  • React Native vs Flutter
  • AI in regulated industries
  • All guides

  • About Apptension
  • Case studies
  • Careers
  • Blog
  • Tech Radar
  • Agentic SDLC hub
  • Contact

Senior AI engineers shipping enterprise-grade AI and SaaS with startup speed.

Headquarters

Apptension sp. z o.o.23 Lutego 26
61-743 Poznań, Poland (opens Google Maps)

VAT-ID PL7831720203 · REGON 360404804
KRS 0000534235

Follow Apptension

Client reviews on Clutch

© 2012–2026 Apptension sp. z o.o. All rights reserved.

Privacy policySitemapllms.txt
  • 01Services
  • 02Portfolio
  • 03About
  • 04Careers
  • 05Blog

Services

All services

New · AI-native delivery

Senior engineers with AI coding agents

Ship a production MVP in weeks, or install agents into the team you already have.

Agentic MVPNewA production MVP your customers can rely on. Validate in 1–2 weeks, launch in 4–8, own all of it.Explore
  • Agentic SDLCAI coding agents installed into your delivery process in two weeks, with guardrails and metrics.
Agentic SDLC hub
>_ProductScope StudioFrom an idea to a shareable PRD_> Open_> Open
Emailcontact@apptension.comCall+48 533 352 052

HQ in Poznań, Poland

Start a project