💬 NLP

Berta: an open-source, modular tool for AI-enabled clinical documentation

The paper introduces Berta, an open-source, modular AI scribe platform deployed at scale within Alberta Health Services that significantly reduces costs and ensures data sovereignty by integrating with existing infrastructure while generating thousands of clinical sessions across numerous facilities.

Original authors: Samridhi Vaid, Mike Weldon, Jesse Dunn, Sacha Davis, Kevin Lonergan, Henry Li, Jeffrey Franc, Mohamed Abdalla, Daniel C. Baumgart, Jake Hayward, J Ross Mitchell

Published 2026-03-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Samridhi Vaid, Mike Weldon, Jesse Dunn, Sacha Davis, Kevin Lonergan, Henry Li, Jeffrey Franc, Mohamed Abdalla, Daniel C. Baumgart, Jake Hayward, J Ross Mitchell

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a doctor's day. For every hour they spend talking to a patient, they spend nearly two hours staring at a computer screen, typing up notes. This is the "pajama time" problem: doctors are so buried in paperwork that they're working late into the night, which leads to burnout and less time for actual patients.

Enter Berta.

Think of Berta not as a robot doctor, but as a super-efficient, invisible intern who sits in on every conversation, listens intently, and instantly types up a perfect draft of the medical notes for the doctor to review.

Here is the story of how this tool was built, how it works, and why it's a big deal, explained simply:

1. The Problem with the "Big Brands"

Right now, if a hospital wants this kind of "AI intern," they have to buy it from big commercial companies.

The Cost: It's expensive, like renting a luxury apartment. It costs between $99 and $600 per doctor, every single month.
The Catch: These commercial systems are "black boxes." You pay them, but you don't own the data. The conversations happen on the company's servers, not the hospital's. It's like sending your private family photos to a stranger's house to have them printed; you lose control over who sees them or how they are used.
The Risk: If the company raises prices or goes out of business, the hospital is stuck.

2. The Berta Solution: "Bring Your Own Intern"

The team behind this paper (from the University of Alberta and Alberta Health Services) decided to build their own version. They called it Berta.

Think of Berta as open-source software. Instead of buying a finished, locked-up product, they built the "blueprints" and the "engine" and gave them away for free.

It's Modular: Imagine building a car. Commercial systems are like buying a Tesla; you can't change the engine. Berta is like a kit car. You can choose your own engine (the AI brain) and your own transmission (the voice-to-text part). If you want to swap parts later, you can.
It Lives at Home: Berta is designed to run entirely inside the hospital's own secure computer network (their "digital fortress"). The patient's voice and the doctor's notes never leave the building. This keeps the data safe and under the hospital's control.

3. The Big Test Drive

To see if this worked, they launched a massive trial in Alberta, Canada.

The Scale: They didn't just test it in one office. They rolled it out to 198 doctors working in 105 different places, ranging from big city emergency rooms to small rural clinics.
The Volume: Over eight months, these doctors used Berta for 22,000 patient visits. That's over 2,800 hours of recorded conversations!
The Growth: At first, only a few doctors used it. By the end, usage exploded. Doctors loved it because it saved them time.

4. The Price Tag: A Fraction of the Cost

Here is the most shocking part:

Commercial Cost: ~ $300–$ 600 per doctor/month.
Berta Cost: Less than $30 per doctor/month.

That's a 70% to 95% savings. Why? Because they aren't paying a vendor's profit margin. They are just paying for the electricity and computing power needed to run the software. A tech-savvy doctor could even run this on their own computer for free!

5. Why This Matters (The "So What?")

This paper proves that you don't need to be a giant corporation to use advanced AI.

Data Sovereignty: The hospital keeps its own data. They can use it to improve care, train better doctors, or research new treatments without asking a vendor for permission.
Customization: Because it's open-source, a hospital in a rural area can tweak Berta to understand local slang or specific medical terms that big commercial tools might miss.
Freedom: Hospitals aren't locked into a contract. If they want to change the AI model or the interface, they can.

The Bottom Line

Berta is like giving every hospital the recipe and the ingredients to bake their own high-tech cake, rather than forcing them to buy an overpriced, pre-made cake from a bakery that won't tell them what's in it.

It shows that AI can be a helpful tool for doctors that is cheap, safe, and owned by the people who use it, rather than a luxury product owned by a tech giant. The team has now released the code for free, so any hospital in the world can download it and start building their own version today.

1. Problem Statement

The paper addresses the critical administrative burden in modern healthcare, where physicians spend nearly two hours on electronic health records (EHR) for every hour of direct patient care. This "pajama time" contributes to physician burnout, reduced job satisfaction, and workforce attrition. While Ambient AI scribes (systems that transcribe conversations and draft notes) have emerged as a solution, current commercial implementations present significant barriers:

High Cost: Subscription fees range from $99 to over $600 per physician per month.
Data Opacity & Sovereignty: Commercial systems operate as "black boxes" on vendor-controlled infrastructure, preventing institutions from auditing model behavior, customizing workflows, or retaining data for quality improvement and research.
Vendor Lock-in: Organizations face risks of pricing escalations, service discontinuation, and an inability to adapt the technology to local regulatory or clinical needs.

2. Methodology

The authors developed Berta, an open-source, modular platform designed to be deployed entirely within an institution's secure infrastructure.

Co-Design Process: The system was co-designed by a multidisciplinary team (emergency physicians, ML engineers, analysts) using an iterative, user-centered approach. Feedback from the Alberta Health Services (AHS) pilot was integrated directly into the open-source release.
System Architecture:
- Frontend: Built with Next.js (Vercel), providing a web and mobile interface for clinicians to record/upload audio and review generated notes.
- Backend: Built with FastAPI (Python), exposing RESTful APIs for logic, data processing, and integration.
- Modularity: The core innovation is the decoupling of components, allowing institutions to swap out specific engines:
  - ASR (Automatic Speech Recognition): Supports WhisperX, OpenAI Whisper, NVIDIA Parakeet (via MLX), and Amazon Transcribe.
  - LLM (Large Language Model): Supports local inference engines (Ollama, vLLM, LM Studio) and commercial endpoints (OpenAI API, Amazon Bedrock).
- Deployment: Capable of running from a single workstation to GPU servers within secure Virtual Private Clouds (VPC) or enterprise clouds (e.g., Snowflake AI Data Cloud).
Pilot Implementation:
- Context: Deployed within Alberta Health Services (AHS), Canada's largest provincial health system, integrated with their existing Snowflake AI Data Cloud infrastructure.
- Configuration: Used WhisperX for transcription and a private-cloud GPT-4o instance for note generation.
- Data Governance: All clinical audio and text remained strictly within the AHS firewall; no data was exported to external vendors.
- Workflow: Clinicians record patient encounters $\rightarrow$ ASR transcribes audio $\rightarrow$ LLM generates a structured draft based on configurable templates $\rightarrow$ Clinician reviews/edits $\rightarrow$ Note is manually transferred to the EHR (copy-paste).

3. Key Contributions

First Provincial-Scale Deployment: This represents the first large-scale integration of an AI scribe into existing health system infrastructure (Snowflake) without relying on a commercial vendor's proprietary stack.
Open-Source Release: The full codebase is released under the Apache 2.0 license, providing a reproducible foundation for other health systems.
Data Sovereignty: The architecture ensures that sensitive patient data never leaves the institutional environment, enabling full control over data governance, privacy compliance, and secondary use for research.
Customizability: Institutions can modify note templates, prompts, and terminology (e.g., regional slang, Indigenous community names) to fit specific clinical workflows.

4. Results

The system was operational for eight months (November 2024 – July 2025) with the following outcomes:

Adoption: Used by 198 emergency physicians across 105 urban and rural facilities.
Volume: Generated 22,148 clinical sessions and over 2,800 hours of clinical audio.
Growth: Monthly usage scaled from 680 sessions (Nov 2024) to 5,530 sessions (July 2025).
Cost Efficiency: Operating costs averaged less than $30 per physician per month. This represents a 70–95% reduction compared to commercial alternatives ( $99–$ 600/month).
Customization: 42% of users customized at least one documentation template.
Expansion: Based on pilot success, AHS approved an expansion to 850 physicians.

5. Significance and Implications

Economic Viability: Demonstrates that high-volume clinical AI documentation is financially sustainable at a fraction of the commercial cost, making it accessible to resource-constrained systems and low-to-middle-income countries.
Research & Quality Improvement: By retaining data internally, the system creates a valuable research asset (2,800+ hours of audio) for analyzing communication patterns and evaluating transcription accuracy, which is impossible with opaque vendor systems.
Regulatory Alignment: The architecture (on-premises processing, human-in-the-loop review, auditable logs) aligns with regulatory frameworks like the NHS England 2025 guide for ambient documentation.
Strategic Independence: Eliminates vendor lock-in risks, allowing organizations to avoid pricing escalations and service discontinuations.
Future Roadmap: The authors plan to integrate directly with EHRs (currently manual copy-paste) and explore richer contextual inputs (e.g., prior patient history) while maintaining strict privacy controls.

In conclusion, Berta proves that an open-source, modular approach can successfully replace proprietary AI scribes, offering a cost-effective, secure, and adaptable solution that empowers healthcare institutions to control their own data and clinical workflows.