Amaya Arie

Model selection assistant

Helping language analysts navigate a growing ecosystem of speech-to-text models with explainable AI

Explainable AI • Enterprise Software • Decision Support • Web

As speech-to-text capabilities expanded, language analysts gained access to more model options. While this increased flexibility, it also introduced a new challenge: understanding when to use each model, why it was recommended, and how it compared to alternatives.

We explored how UX design and explainable AI could help analysts navigate this transition from a single-model workflow to a growing ecosystem of model choices.

Client

Laboratory of Analytical Sciences

Year

Jan. - Dec. 2025

Role

UX Researcher

Team

2 PI's and 3 Research Assistants

comparing two models

previewing a models transcription

At a Glance

The Challenge

Analysts were moving from a workflow centered around a single default model to one with multiple model options, increasing decision complexity.

Opportunity

Use explainable AI (XAI) to help a variety of experience-level analysts understand recommendations rather than simply receive them.

Solution

A three-level explanation framework that supports model discovery, evaluation, and comparison.

results

Developed a research-backed framework for explainable model recommendations that was documented for future implementation by stakeholders.

Design process

A research-driven approach to supporting analyst decision-making

This project followed an iterative design thinking process that balanced human-centered research with the technical complexity of explainable AI systems. The process emphasized understanding analyst workflows, decision pressures, and trust-building behaviors before moving into solution design.

Each phase built intentionally on the last. Early research focused on understanding how analysts currently work and how future access to multiple STT models could change their decision-making. Insights from this phase informed clear problem framing and prioritization, which then guided concept exploration, prototyping, and testing. Throughout the process, designs were continuously refined to reduce cognitive load, surface meaningful signals, and support confidence without oversimplifying complexity.

The result was a layered, explainable interface concept shaped through research, iteration, and feedback designed to adapt to varying analyst needs while remaining transparent, flexible, and grounded in real-world use cases.

Understanding analysts current workflow and XAI best practices

Empathize

Identify the core problem

Define

Initial concept development

Ideate

Conduct usability tests

test

Refinement & High-Fi vibe coding

refine

The shift

What happens when one model becomes many?

One Model

Multiple Models

Currently, language analysts rely on one default speech-to-text model to process audio and support their work. As additional models become available, analysts will gain flexibility and the potential for improved performance across different audio conditions. However, instead of simply accepting a recommendation, analysts now needed to evaluate multiple options, understand tradeoffs, and determine which model best aligned with their goals. The challenge became how do we understand AI's logic?

research

Understanding analyst workflows and decision-making

To understand how analysts evaluate model recommendations, we combined interviews, surveys, workflow analysis, and a literature review focused on explainable AI and decision-support systems. Our goal was to understand how analysts used models and built confidence in their decisions.

Informal interviews

Informal interviews with analysts (limited time & availability)

Survey

Analyst feedback on recommendation explanations and decision support

Workflow Mapping

Understanding current model-selection and audio-analysis workflows

Literature Review

Explainable AI, trust calibration, and AI-assisted decision making

exploring current speech-to-text models and analysts workflow

insights

What information helps analysts evaluate a recommendation?

Analysts trusted each other.

System recommendations alone were not sufficient to build confidence. Analysts wanted to understand hie other analysts were utilizing certain models to enhance their workflow. The key was building community through an experience where AI could create isolation and overload.

Analysts trusted each other.

Different decisions required different levels of explanation.

Some decisions required only a quick recommendation, and others required deeper technical understanding. Explanation strategies needed to be flexible to support every scenario.

Different decisions required different levels of explanation.

Some decisions required only a quick recommendation, and others required deeper technical understanding. Explanation strategies needed to be flexible to support every scenario.

Model performance was only part of the story.

Analysts considered factors such as training data, audio conditions, domain relevance, latency, and speaker complexity when evaluating recommendations.

Model performance was only part of the story.

Analysts considered factors such as training data, audio conditions, domain relevance, latency, and speaker complexity when evaluating recommendations.

Analysts needed support comparing tradeoffs.

Recommendations alone were not sufficient to build confidence. Analysts wanted to understand the reasoning behind a suggestion before committing to a model.

Analysts needed support comparing tradeoffs.

Recommendations alone were not sufficient to build confidence. Analysts wanted to understand the reasoning behind a suggestion before committing to a model.

Key finding

Layered explanations explore trust, transparency, and adaptability

One of the most important insights was the value of layered explanations framework for explainable AI. Instead of presenting all system logic at once, explanations can be structured to allow analysts to progressively access deeper levels of detail about a recommendation.

In this framework, high-level signals communicate what the system recommends, additional context explains why a model is appropriate for a specific audio scenario, and deeper layers reveal how the system arrived at that recommendation. This approach supports both quick decision-making and deeper investigation, giving analysts control over how much information they engage with while maintaining transparency and trust.

Level 1

Level 2

Level 3

Quick Suggestions

confidence bar

allows users to access the abstract explanation (Level 2).

marks that the model is recommended by the system.

Level 1

Level 2

Level 3

Quick Suggestions

confidence bar

allows users to access the abstract explanation (Level 2).

marks that the model is recommended by the system.

Level 1

Level 2

Level 3

Quick Suggestions

confidence bar

allows users to access the abstract explanation (Level 2).

marks that the model is recommended by the system.

With the framework established, we began exploring how layered explanations might appear within the analyst workflow. We tested multiple ways information could be revealed across the recommendation experience.

Concept 1

List-based model recommendations

When users are working with a model, the system automatically suggests alternative models based on performance and community usage data, presented in a concise list format. This approach focuses on quick comparison and supports progressively deeper explanation levels as users engage further.

Concept 2

Visual comparison

The system visually compares the current model with top alternatives across key performance metrics, allowing users to quickly identify strengths and weaknesses. This concept emphasizes pattern recognition and supports comparison without relying solely on text.

Concept 3

Task-focused expansion

Based on the user’s current activity, the system suggests not only better-suited models but also potential next tasks. This approach supports forward-thinking exploration beyond the user’s initial query.

key insight

Trust was not tied to seniority, but to how supported analysts felt in understanding and validating model recommendations. Stakeholders found visualizations difficult to interpret and more cognitively demanding than anticipated. While informative, they required additional education and slowed quick decision-making.

Replit

high fidelity vibe coding

While the project primarily focused on research and framework development, I later revisited the concept to explore how the experience could evolve through a more refined visual language.

solution

Supporting analyst decision-making through layered explanations

The first layer helps analysts quickly identify promising models based on their audio conditions and workflow requirements.

Recommendations are paired with concise reasoning and key performance indicators, allowing analysts to understand why a model surfaced without diving into technical details.

Once analysts identify a promising model, they can explore the factors influencing the recommendation.

This layer introduces richer contextual information.

This layer supports side-by-side exploration of model strengths, weaknesses, and tradeoffs.

Rather than identifying a single "best" model, the experience helps analysts determine which model best fits their current situation.

impact

Beyond a single recommendation; Building a 3-Level Explainability Framework

The project established a foundation for supporting analyst decision-making as model ecosystems continue to expand. Through layered explanations, the framework helps analysts move beyond accepting recommendations at face value and toward understanding the reasoning behind them. The resulting framework was documented for stakeholders as a foundation for continued development.

reflection

Layered explanations explore trust, transparency, and adaptability

Working in a highly technical domain reinforced the importance of translating complex systems into experiences that feel understandable and actionable. More than anything, this project showed me that small design decisions around transparency, hierarchy, and trust can have a significant impact on how people engage with AI-powered systems.

reflection