What's in a Voice
- Your voice carries more than words. Bridge2AI-Voice, a project funded by the National Institutes of Health (NIH), is testing whether AI can pick up signs of conditions like Parkinson's, dementia and mood disorders from voice recordings.
- Voice is also one of the hardest kinds of data to make anonymous. The same recording that might flag an illness can identify the speaker.
- The project earns trust three ways: it names AI as the purpose up front, it shares measurements of the sound instead of the recordings, and it's honest about what withdrawal can and can't undo.
- Radio hosts, podcasters and independent voices hold some of the richest everyday speech anywhere. The same three rules can make that audio useful to AI, with consent as the price of entry.
- Box Commons asked NIH's new AI office to make these practices the default for NIH data used to build AI.
+ Jump to Section
What a Voice Carries
When you talk, you share more than words. The pitch, steadiness, pace and breath behind them shift when your body does. Medicine has listened for this for a long time: breath, cough and voice “have long been used for diagnostic purposes.”
Bridge2AI-Voice wants to know how far that listening can go. It's one of the projects in NIH's Bridge to Artificial Intelligence program, paid for by the NIH Common Fund and led from the University of South Florida. Its latest release holds 48,226 recordings, about 190 hours of audio, from 1,370 people at five sites in the United States and Canada, each linked to health information. The project built its study groups around five areas where research has tied voice changes to illness: voice disorders; neurological conditions such as Parkinson's and ALS; mood and psychiatric conditions such as depression; respiratory conditions; and children's voice and speech disorders.
The consent form participants sign states the hope plainly. A voice, it says, “contains audio features that have been linked to diseases such as Parkinson's, dementia, mood disorders, and certain cancers.” The open question is whether AI can learn to hear those features reliably.
It isn't a diagnostic test yet. It's the slow work of building a shared, well-documented dataset so researchers can find out. If the answer is yes, a first check for some of these conditions could start with something as ordinary as talking. Most participants finished every recording in a single session, on a tablet with a headset.
The Catch
What makes voice useful also makes it hard to share. A voice identifies its speaker. Researchers on the project's ethics team, including Jean-Christophe Bélisle-Pipon and Vardit Ravitsky, found in a review of the field that “encryption, anonymization, and de-identification” have “proven to be more challenging with voice and speech data than with other biometric information.”
The project faced the problem every voice dataset faces. People lend their voices only when they trust what happens next. Without that trust there's no dataset, and without a dataset there's no answer.
How They Earned Trust
They said what the AI is for. The consent form doesn't bury it. The study aims to learn whether voice biomarkers “could be screened and differentiated through AI and machine-learning algorithms.” The project also publishes the consent forms from each site, revisions included, in a public repository anyone can read.
They share measurements, not recordings. Bridge2AI-Voice has never released its raw audio openly. Its releases carry derived features instead: spectrograms (pictures of how a sound's pitch and energy change over time) and other measurements of the sound. The latest release holds back “highly identifying features such as speaker embeddings” and blurs its spectrograms on purpose “to make speaker identification more difficult.” Even those measurements go only to credentialed researchers who sign a data use agreement. The recordings themselves sit behind an access committee, and getting them “requires institutional sign off.” The consent form sets up these tiers in advance: low-risk data can be open, while data with a “heightened risk of re-identification,” including “certain voice data recordings,” stays behind controlled access.
They're honest about withdrawal. The consent form doesn't overpromise. It says that once collection is complete, a participant's “voice recording, and other research data cannot be removed from the database.” In practice the project does better than its form: the October 2026 release left out 19 people from the previous version, for reasons that include changes to their consent or data-sharing status.
The harder questions are still open: what happens to copies researchers already downloaded, and to AI models already trained on someone's voice? NIH treats trained models as “Data Derivatives” and has promised “future guidance” on them. Until that guidance arrives, the honest move is to tell people exactly what you can and can't take back. That honesty is the part worth copying.
Where Radio, Podcasts and Independent Audio Fit
Clinics aren't the only places people talk. Community radio, podcasts, call-in shows and independent stations hold years of real people talking in real rooms, with every age, accent, mood and microphone. That everyday speech is what AI needs to sound like a person instead of a press release. Our short version: helping AI voices sound less bland, with your okay.
Broadcast audio has the same catch as clinical audio, because every voice on the air is identifiable. The rules that make Bridge2AI-Voice work carry over almost word for word.
| What the researchers do | What it means for your audio |
|---|---|
| Name AI as the purpose | “We'll tell you plainly that AI will learn from this, and what for.” |
| Share measurements, hold back recordings | “Anything that could point back to you, like your actual voice, stays behind a locked door, and only approved people get a key.” |
| Be honest about withdrawal | “If you change your mind, here's exactly what we can take back, and what we can't.” |
Consent is the price of entry. It isn't the headline. The headline is what a voice can do when people trust where it's going. That's what Box Commons standards are for: clean audio data, with consent and provenance recorded at the moment of collection, not pieced together later.
What We Asked NIH to Do
NIH is setting up a new Office of the Chief AI Officer inside the division that runs Bridge2AI. On October 23, we asked that office to make these practices the default for NIH data used to build AI. Each one is already in use by an NIH-funded team:
- A withdrawal statement that covers future releases, downloaded copies and trained models.
- Tiered release by how identifying the data is, with measurements first for biometric data like voice.
- One machine-readable datasheet that records where the data came from and what participants agreed to.
- Consent forms that name AI as a purpose, published openly.
- A written NIH definition of “ethically sourced,” a term the program relies on but which, its own ethics researchers note, “has not been defined by the NIH.”
NIH's own teams have done the hard part. We asked the new office to make their answers the standard.
A voice carries a lot. The system around it should carry trust too.
Related Filing: Box Commons Comment on the NIH DPCPSI Reorganization (Office of the Chief AI Officer), filed October 23, 2026
Related Analysis: Your Voice Is Not Training Data (Unless You Say So)
Box Commons uses AI-assisted drafting in its publications. Brice Love set the research direction and made the editorial calls for this article. Ax, the AI collaborator named in the byline (built on Claude, by Anthropic), contributed research synthesis and drafting. We checked every quotation against its primary source, and we keep editorial control over the final text.