Inspiration

The inspiration for Handsfree comes from my daily anxiety about not having enough time to read papers. I wish I could read more papers during my PhD training, but it is really hard to manage my time when experiments are so time-consuming.

I do not want to listen to music or audiobooks for ten hours straight while running experiments all day, so I have always wanted something that could read scientific papers aloud to me without demanding all of my attention.

I previously tried Google’s NotebookLM, but I found that its podcast-style summaries removed too many important scientific details. I also do not feel comfortable trusting authors’ conclusions without examining the figures, because that is often where the most important evidence is presented.

What It Does

We built Handsfree to read scientific papers aloud so that I can learn about and think through research while doing experiments. You can think of “Handsfree” as freeing your hands for experiments so that you can use your time more efficiently.

The website can also be used in many other situations, such as while running, commuting, or doing other tasks that do not require your full attention.

Handsfree reads papers aloud and explains their figures. It uses verbal cues to distinguish between text written by the authors and AI-generated explanations of visual content.

When the narration reaches an important section, I can pause and examine the figure myself. This allows me to combine listening with my own visual interpretation and decide whether I truly understand or trust the experiment, without completely disrupting the experimental workflow I am already working on.

How We Built It and the Challenges We Ran Into

I started building Handsfree by giving Codex a list of the features I wanted. The first version was already functional, but the website did a terrible job of recognizing figures and segmenting them into reasonable panels.

I asked Sol to research tools that could solve this issue. It looked into several open-source options and selected a model that was small enough to run efficiently while still completing the task well.

Another major challenge came from text-to-speech. Sol found a good free TTS system on the first attempt, but the initial version had problems with pacing, pauses, and verbal cues. These details are important for making long listening sessions comfortable and easy to follow.

After testing the system ourselves, we gave Sol additional requirements based on the problems we encountered. We also asked it to think through broader use cases and edge cases rather than overfitting the solution to one specific paper or disrupting features that were already working well.

We also noticed that, as the context length accumulated, Sol’s performance seemed to decrease. This made it harder to fix issues gradually because new changes could unintentionally affect parts of the system that had already been working.

Accomplishments That We’re Proud Of

I tested Handsfree today while building an experimental rig, and it did a very good job of explaining a paper to me. It met the requirements I originally had in mind.

All of the current features are free to use, and no API key is required. The text-to-speech output is also natural enough to provide a comfortable listening experience during long experiments.

I am especially proud that Handsfree does not simply summarize a paper. It preserves the scientific details, explains the figures, and still gives the listener the opportunity to pause and evaluate the evidence independently.

What We Learned

We learned that reading a scientific paper aloud is not the same as converting its text into speech. A useful scientific listening experience depends on structure, pacing, figure interpretation, and clear distinctions between the authors’ claims and the AI’s explanations.

We also learned that figures cannot be treated as optional supplementary material. In many papers, the figures contain the actual evidence needed to understand whether the conclusions are justified. A system that ignores or oversimplifies them may sound convincing while leaving out the most important part of the paper.

Another lesson was that small details in text-to-speech design matter much more during long listening sessions. Pauses, transitions, verbal cues, and pronunciation may seem minor when testing a short paragraph, but they strongly affect whether someone can comfortably follow a full scientific paper.

Finally, we learned that building with coding agents still requires careful human testing and iteration. The agents were useful for researching tools and implementing features quickly, but we still needed to identify failures, define better requirements, and make sure that fixing one problem did not break something that was already working.

What’s Next for Handsfree

We would like to add authentication features so that Handsfree can access papers available through a user’s institutional subscriptions and retrieve articles directly from journal websites.

We also want Handsfree to become a platform for scientific discussion. Users could share comments, critiques, alternative experimental approaches, and ideas for future research. In this way, the platform could function like an online journal club and encourage more communication around scientific papers.

The platform could also be useful for college students who are interested in science and want to learn how early-career researchers evaluate experiments and interpret scientific evidence.

Finally, we may add optional API-key support to give users more choices for text-to-speech voices and figure-explanation models.

Built With

• codex