YouTube-to-Docs

What's the Problem?

State agencies and researchers rely heavily on text-based records of public meetings. Regularly hosted on YouTube, these public meetings, legislative hearings, and training videos are subject to strict legal mandates for digital accessibility. This includes requirements to provide accurate and easily accessible transcripts, as well as multi-language translation services for limited English proficient (LEP) individuals. 

While YouTube offers native, auto-generated captions, these are consistently inaccurate. They often misinterpret regulatory language, technical state acronyms, and speaker identities. Public servants spend extensive hours manually making new transcripts or reviewing faulty transcripts. Only after establishing this textual baseline can staff begin to index speakers, draft notes, and extract key insights. 

This manual workflow creates several problems for public sector agencies:

• High Administrative Cost: A typical two-hour agency briefing takes eight hours of human labor to transcribe. State employees spend thousands of hours on manual transcription across several sets of videos.
• Poor Quality Transcripts: YouTube transcribes millions of videos a day on a lower quality, cheaper model. As a result, its auto-generated captions can mislabel speakers, use inaccurate technical terms, and confuse agency-specific acronyms. This makes the video content inaccessible to people who rely on clean text summaries or assistive reading tools. 
• Information Bottlenecks: Maryland's civil servants lacked a quick, reliable way to search through public hearing videos, which can last hours. Without accurate transcripts as a content guide, this causes policy analysts to miss key details and critical context from past discussions. 

Under Maryland policy, all technology procurements, software adoptions, and platform approvals must be approved by the State's Department of Information Technology (DoIT). This helps ensure compliance with state security, data privacy, and budgetary guidelines. 

In the case of improving transcript services, Maryland agencies had two options: Pursue a long procurement cycle to buy third-party transcription software or build a tool in-house. 

DoIT's Solution

DoIT is actively transitioning from a traditional “buyer” approach (which relies on external vendor contracts) to a "builder" mindset that leverages state talent for information technology (IT) solutions where it makes sense. 

YouTube-to-Docs is a direct product of this cultural shift. This tool was developed in-house using agentic engineering principles. This means that its creation balances traditional software engineering best practices with the speed and flexibility that AI offers. 

Ray Bell, the State of Maryland's AI Product Director, designed and developed YouTube-to-Docs with reduced manual effort while upholding strict standards for software quality and security. Bell began by collecting requirements from an agency that initially requested to procure external transcription software. When they confirmed during AI office hours that this software did not meet DoIT's data protection standards, Bell chose to build a solution himself. 

As a result, YouTube-to-Docs was built to serve specific state worker needs. Instead of reviewing video footage for hours, agencies can extract accurate, structured transcription data. This approach removes the friction of manual transcription by quickly delivering verbatim transcripts, timestamped outlines, and text summaries.

The development team built three separate access points directly into the repository. Non-technical employees can use a clean, web-based app layout, while developers can use an Application Programming Interface (API) or a Model Context Protocol (MCP) server. API and MCP infrastructure allow AI assistants to expose the tool's features natively, letting developers retrieve transcripts using natural language right inside their workspaces.

How Does YouTube-to-Docs Work?

YouTube-to-Docs takes video URLs, unique IDs, playlists, channel feeds, and comma-separated batches as acceptable forms of input. Then, by pulling raw audio or existing transcripts, the tool can quickly return new, accurate transcripts. While the tool’s default is grabbing native YouTube transcripts to improve them, it can also create original transcripts. 

Data is saved directly to each user’s designated workspace (e.g., Microsoft SharePoint, Google Workspace, local folders) as a text file, which users can put through additional processing using their DoIT-provided AI tools of choice. This allows users to directly query the transcript material and extract specific details on demand.  

YouTube-to-Docs' features include video summaries, infographics, and supporting multi-language translations by applying bidirectional speech-to-text and text-to-speech pipelines. As an open-source, cloud-, AI-, and vendor-agnostic tool, it runs on a local machine. Large language models (LLMs) from Amazon Web Services (AWS), Google Cloud Platform (GCP), and Azure are all compatible with YouTube-to-Docs.

YouTube-to-Docs has four distinct interfaces to fit agency workflows. This gives users with diverse technical skills and comfort levels their choice of how to use this tool:

• A Web-Based Application: The Maryland Web Design System (MDWDS) library is a standardized framework of design tools, guidelines, and pre-built UI components developed by the Maryland Digital Service. YouTube-to-Docs leverages MDWDS on a web browser, which ensures design consistency across the State’s digital services and makes it accessible. 
• A Command Line Interface (CLI): A CLI is a text-based user interface (as opposed to clicking on icons with a mouse) that interacts with a computer’s operating system, software, or files. This allows users to run YouTube-to-Docs from their own terminal. 
• A Model Context Protocol (MCP) Server: An MCP server acts as a USB-C port for AI applications. It can bridge AI models (e.g., LLMs) with external data sources, applications, and developer tools. This integrates YouTube-to-Docs with a wide range of AI coding tools. 
• AI Skills: “AI skills” package an AI agent’s capabilities using plain text markdown files so users can upload instructions on demand. These markdown files containing specific context and rules for proper use provide YouTube-to-Docs with specific criteria for video transcription and analysis. 

How is YouTube-to-Docs Compliant with Maryland's AI Governance Policies?

YouTube-to-Docs was designed and developed entirely in-house by Maryland state employees, not an outside vendor. This means DoIT retains full ownership over how the tool processes public data and maintains control over its security. This strategy supports the mandates for secure, trustworthy technology outlined in Gov. Wes Moore’s Executive Order on Artificial Intelligence.

In accordance with the State’s Responsible AI Policy, YouTube-to-Docs acts as an administrative assistant, not an autonomous decision-maker. Users must maintain active oversight and verify generated summaries against their sources before using them in any official capacity.

YouTube-to-Docs complies with Maryland's Responsible AI Implementation Guidance by restricting its scope to Level 1 Public Data. The software only processes public YouTube URLs like legislative hearings and town halls. This reduces the risk of exposing sensitive data or compromising confidential State networks. The application’s data remains within DoIT's isolated tenant and isn’t used to train public AI models.

The web application interface was built using the Maryland Web Design System (MDWDS) library. Using these pre-approved components guarantees that the frontend meets Maryland's digital accessibility standards for citizens who use assistive technologies to access state services.

Key Impacts

YouTube-to-Docs' infrastructure is functional, stable, and secure. Agencies can quickly scale YouTube-to-Docs into daily operations as needed. 

Internal testing confirmed that the tool provides three clear advantages when used for state workflows: 

• Time Savings: Processing a standard two-hour public meeting previously required hours of manual transcription and summarizing. YouTube-to-Docs automates the bulk of this manual pipeline and generates clean text in minutes, meaning the transcript is instantly ready for human review.
• Streamlined Video Analysis: Rather than scrolling through hours of video to find key details, Maryland's state staff can use basic keyboard shortcuts (such as Ctrl+F) to instantly search the compiled text files.
• Accessibility: YouTube-to-Docs’ language models support multi-lingual processing and can quickly transcribe audio or text into different languages. This allows a diverse user base to benefit from the tool's features.
• Cost Reductions: Agencies don't need to procure expensive, third-party video transcription software. Instead, they can quickly access a free service that was built with and for them.

Constraints and Limitations

YouTube-to-Docs is designed to work within a local environment rather than being deployed as a wide-access application. This is because public cloud service providers often face strict Internet Protocol (IP) address blocks when pulling data from YouTube. 

To bridge this gap, a developer or administrator must host the repository on a local workstation or server. Once initialized, non-technical staff can seamlessly connect to the tool's capabilities through its browser-based interface.

Running the application is a deliberate security measure to safeguard state infrastructure. This setup avoids commercial cloud IP restrictions while securing sensitive open authorization client IDs and API keys. To maintain operational safety, developers using YouTube-to-Docs must do so with DoIT-approved internal infrastructure, designated LLM endpoints, and vendor environments.

Currently, YouTube-to-Docs supports the following languages:

• Arabic

• Chinese

• Dutch

• French

• German

• Hindi

• Italian

• Japanese

• Korean

• Portuguese

• Russian

• Spanish

• Turkish

• Ukrainian

• Vietnamese

YouTube-to-Docs only handles public YouTube URLs. Processing alternative video hosting platforms is out of scope at this time.

Demo of YouTube-to-Docs

The following screenshots show YouTube-to-Docs' various features. They use the Maryland Department of the Environment's 26-NT-3016 Wetland Permit Application Hearing as their case study, and Gemini 3.1 Flash-Lite as the LLM: 

Grabbed and cleaned-up YouTube transcript, CLI view:

Identifying the speakers, CLI view:

Generated summary, CLI view:

Generated infographic, CLI view:

Generated alt-text for infographic, CLI view:

Transcript translation into French, CLI view:

Interface for video upload, web-based view:

Grabbing and downloading the transcript, web-based view: 

To see YouTube-to-Docs in action, watch the following YouTube-to-Docs video demonstration.

What's Next?

The roadmap for YouTube-to-Docs focuses on defining a long-term deployment strategy alongside key feature expansions. 

The tool's administrators at diverse agencies can monitor live performance benchmarks (e.g., processing times, accuracy scoring, and saved administrative hours) to ensure that YouTube-to-Docs acts as intended. The team that built it will also monitor API token consumption and local computing expenses to prevent budget drift.

To maximize their tool's reach across government operations, the team is exploring ways for it to process any standard video recording and support more languages. This expansion would remove dependency on public YouTube links and allow agencies to securely process video content on a more diverse set of hosting environments. In tandem, it would open access to Marylanders who speak even more languages.

Maryland state employees and contractors who want to explore this repository or learn how to leverage similar architectures can sign up for free office hours every Monday. DoIT’s AI Enablement Team is also available to help answer questions regarding Maryland’s responsible AI use policy or to assist in rolling out secure AI utilities across state agencies.