Lessons learnt from building an AI Voice Assistant for scientific labs - Voice AI Space Amsterdam

Overview of the Voice AI Lab Startup

This is a voice AI startup built for scientific labs, helping scientists document experiments and interact with lab systems hands-free. The founder explains how the company started by solving a painful problem in drug discovery and lab research: scientists spend too much time manually recording notes, updating systems, and managing documentation instead of focusing on experiments.

The product uses voice AI to automate documentation, improve data quality, and make scientific records more reproducible. Instead of forcing scientists to type notes or manually update legacy databases, the system allows them to speak naturally while working in the lab.

Main Functions

Voice-Driven Lab Documentation:
The platform lets scientists record experiment details through speech. This reduces manual note-taking and helps create more accurate, reproducible scientific records.

Legacy Database Integration:
The system connects voice workflows with older lab databases and existing pharma infrastructure. This allows teams to adopt voice AI without replacing their current systems.

System Architecture

The system is designed for challenging lab environments where traditional voice AI does not work perfectly out of the box. Scientific labs can be noisy, sterile, and full of specialized terminology that general speech recognition models may misunderstand.

To solve this, the team used LLMs for intent recognition but also built custom models for domain-specific speech accuracy. Since real scientific voice data can be limited or sensitive, they trained models using synthetic data to improve recognition of scientific terms, experiment names, and lab-specific vocabulary.

The product also had to work with legacy databases, meaning the voice layer needed to sit on top of existing systems instead of forcing labs to rebuild their workflows from scratch.

Key Learnings

Trust Is the Biggest Barrier:
Scientists initially resisted voice AI because mistakes in lab documentation can have serious consequences. In critical research environments, users need to trust that the system understands them correctly.

Visual Feedback Is Essential:
Early voice interfaces lacked clear visual feedback, which made users doubt whether the transcription was accurate. A strong voice UI should not only listen but also show users what it understood and guide them through the process.

Generic Speech Models Are Not Enough:
Scientific terminology is complex, inconsistent, and often missing from general ASR systems. Custom models trained with synthetic data helped fill this gap and improve domain-specific accuracy.

Voice UI Still Lacks Standards:
Unlike web or mobile apps, there is no universal design standard for voice applications. The ideal lab voice system needs to guide users both visually and verbally so they know what to say, what the system captured, and what action will happen next.

Market Timing Matters:
The founder notes that first-mover advantage is not always guaranteed. Sometimes competitors educate the market first, creating a last-mover opportunity for companies that enter later with a better, more refined product.

Technical Details and Q&A

Technology Stack:
The system uses voice AI, speech recognition, LLM-based intent recognition, synthetic data, custom domain models, and integrations with legacy lab databases.

Reasoning / Agent Logic:
LLMs help identify user intent from spoken requests. However, because lab workflows require high precision, the company did not rely only on general-purpose LLMs. They built custom models to better understand scientific vocabulary and reduce errors.

Synthetic Data:
Synthetic data was used to train the system when real-world lab speech data was difficult to collect. This helped the model learn scientific terminology, experiment-related phrases, and domain-specific commands.

User Experience Features:
The product supports hands-free interaction, allowing scientists to continue working while speaking commands or recording experiment details. The system also needs strong visual feedback so users can confirm what was captured and avoid hidden transcription mistakes.

Lab Environment Challenges:
Labs introduce difficult real-world conditions, including background noise, sterile workflows, user hesitation, gloves, equipment sounds, and strict documentation requirements. These conditions made speech recognition and user adoption harder than in normal office settings.

User Personas:
Different users responded differently to voice AI. Tech-savvy users were more willing to tolerate early mistakes, while traditional pharma users expected speed, accuracy, and reliability from the start.

Debugging and Testing:
Debugging voice applications was difficult because there are fewer established frameworks for stress-testing voice workflows. The team had to test how the system behaved under noisy conditions, unclear commands, domain vocabulary, and real lab usage.

Security and Authentication:
Security is not the main focus of this transcript, but the product operates in sensitive scientific and pharma environments. This means integrations, documentation accuracy, and controlled access to research data are important considerations.

Quality Assurance:
Quality is measured through transcription accuracy, correct intent recognition, workflow completion, and the reliability of scientific records. The system must reduce administrative burden while improving reproducibility and data quality.

Important Keywords and Definitions

• Voice AI:
  Technology that allows computers to understand and respond to spoken language.

• Synthetic Data:
  Artificially generated training data used when real-world examples are limited, private, or difficult to collect.

• Intent Recognition:
  The AI’s ability to understand what the user wants to do based on their spoken request.

• Reproducible Data:
  Scientific records that can be repeated and verified by others.

• Last-Mover Advantage:
  The benefit of entering a market later after competitors have already educated users and exposed common mistakes.

• Legacy Databases:
  Older systems that companies already use and that new tools must integrate with carefully.

• Sterile Environments:
  Controlled lab spaces where cleanliness, contamination control, and precise workflows matter.

• Agentic Process:
  An AI workflow where the system can take actions to complete tasks, not just answer questions.

• Text-to-Speech:
  Technology that converts written text into spoken audio.

• Debugging Voice Apps:
  The process of finding and fixing errors in how voice systems hear, interpret, and respond to spoken commands.

• Analogy:
  Building voice AI for labs is like teaching a robot chef to work in a busy kitchen while wearing gloves. It must understand commands without relying on hands, ignore background noise, and never confuse important ingredients. Early attempts failed because the robot did not understand the kitchen’s special language. Success came from training it with thousands of synthetic examples until it could follow instructions reliably without disrupting the work.