When Casey Harrell's ALS had progressed to the point where he could barely move, surgeons placed an array of 256 electrodes over the surface of his brain. That was nearly two years ago. Since then, he's expressed close to 2 million words—not by typing, not by eye-tracking a keyboard, but by thinking them. A computer reads the neural activity from the part of his brain that once controlled his mouth and tongue, and speaks for him.

The technology keeping Harrell connected to his family and his full-time job represents one of the most significant shifts in how we treat paralysis. After decades of research, neural interfaces have finally crossed from laboratory curiosity to practical communication tool.

From 15 Words to Full Conversations

The progress has been swift and recent. In 2021, researchers at UC San Francisco published a study of a man they called "Bravo-1," paralyzed for 16 years after a brainstem stroke. Their system could decode 50 words from his brain activity—but only those 50 words, and only at about 15 words per minute with a 25% error rate. Useful, perhaps, but frustratingly slow for someone trying to have a conversation.

Four years later, the same research team achieved something different. Their new system, published in March 2025, decodes speech at nearly 48 words per minute with a full vocabulary—not a preset list of phrases, but any word the user wants to say. With a limited 50-word vocabulary, the speed jumps to 91 words per minute. Normal conversation runs about 130 words per minute, which means these devices are approaching the rhythm of natural speech.

The accuracy improved even more dramatically: over 99% success rate, compared to the earlier 75% accuracy. The difference between three-quarters correct and nearly perfect is the difference between laborious guesswork and actual conversation.

What the Brain Reveals About Speech

These systems work by reading neural activity from the sensorimotor cortex, the part of the brain that controls movement. When we speak, our brains send signals to dozens of muscles in the tongue, lips, jaw, and larynx. Even when someone is paralyzed and can't move those muscles, the brain still generates the neural patterns for speech.

Surgeons implant a thin pad containing 64 to 256 microelectrodes on the surface of the brain, positioned over the region that controls speech production. These electrodes detect electrical activity from thousands of neurons. Deep learning algorithms—similar to the AI that powers Siri or Alexa—translate those patterns into words.

The system then synthesizes speech using recordings of the patient's pre-paralysis voice, when available. For the 47-year-old woman who participated in the 2025 streaming speech study, researchers used recordings from her wedding video, made before a stroke took her ability to speak 18 years earlier. When the computer speaks for her, it sounds like her.

The Training Problem and Its Solution

Early systems required extensive training on specific words and phrases. The breakthrough came from training on massive datasets—over 23,000 speech attempts across 12,000 different sentences, including more than 1,000 unique words drawn from social media and movie transcripts.

But the real achievement is generalization. The systems can now decode words and sentences they've never encountered during training. A 45-year-old man with ALS, whose speech had become severely slurred, achieved 97.5% accuracy after just 16 hours of calibration. Within a day and a half, his vocabulary expanded from 50 words to 125,000 words with over 90% accuracy.

This means users aren't limited to a predetermined vocabulary. They can express new thoughts, use slang, adapt to changing circumstances—all the things that make language useful for actual communication rather than just requesting basic needs.

Living With a Neural Interface

Harrell, the ALS patient who's been using his system for nearly two years, has logged over 3,800 hours across 400-plus days. He uses it for more than 12 straight hours at a time without assistance. His average speed is 56 words per minute at 92% accuracy—fast enough to maintain employment despite his paralysis.

The system includes eye-tracking technology so he can navigate screens and select options. It has a privacy mode that prevents recording when he wants certain thoughts to remain private. The neural patterns have remained stable over the months and years without needing recalibration—the brain's speech signals are consistent even when the body can't respond.

Through his device, Harrell shared: "It is very sweet to have the ability to look at my wife's eyes when she hears my voice and conjures up a sweet memory."

The 80-Millisecond Gap

The latest system decodes and synthesizes speech in less than 80 milliseconds—less than a tenth of a second from neural signal to spoken word. This speed matters because conversation requires back-and-forth exchange. When there's a multi-second delay between thinking and speaking, dialogue becomes stilted and unnatural.

The 2025 system can also produce speech indefinitely without interruption, unlike earlier devices that worked in short bursts. This streaming capability allows for the flow and rhythm of natural conversation. Users can interrupt themselves, change direction mid-sentence, or speak for extended periods—all things that eye-gaze typing systems make difficult or impossible.

What These Devices Still Can't Do

Current systems can't yet convey emotion through tone, pitch, or volume. The synthesized voice sounds flat compared to natural speech, which carries meaning through inflection and emphasis. Researchers are working on this, but it requires decoding not just what someone wants to say but how they want to say it.

The technology also requires brain surgery to implant the electrodes, which limits who can access it. The devices themselves are expensive and require technical support. And while the neural patterns remain stable, the hardware can fail or need replacement.

Most importantly, these systems only work for people whose speech-control brain regions are intact. Someone with damage to those areas—from certain types of stroke or brain injury—may not generate clear enough signals for the algorithms to decode.

When Thought Becomes Sound Again

Twenty-seven participants are currently enrolled in the BrainGate 2 clinical trial testing these devices across multiple research sites. The technology is still experimental, still being refined, still limited to people enrolled in research studies.

But the trajectory is clear. In four years, these systems went from 50-word vocabularies at 15 words per minute to unlimited vocabularies at nearly 50 words per minute. Error rates dropped from 25% to less than 1%. Training time decreased from weeks to hours.

For the first time, people who have been silent for years or decades can speak at speeds approaching normal conversation, with accuracy high enough for reliable communication. The interface between brain and computer has become fast enough, accurate enough, and stable enough to restore something we rarely think about until it's gone: the ability to turn a thought into a voice.