A voice is not just sound. It carries identity, memory, personality, the particular way someone laughs or hesitates before a word. For people living with conditions like ALS, amyotrophic lateral sclerosis, the gradual loss of speech is not only a physical limitation. It is the erosion of self. What artificial intelligence is now making possible in this space is worth understanding carefully, because it goes well beyond a technical upgrade to assistive devices.
The Old Way Was Better Than Nothing, and Not Much Else
For decades, people facing voice loss had two main options. The first was message banking: recording common phrases while speech was still possible, creating a library of pre-set responses. The second was voice banking, a more systematic process requiring a person to record around 1,600 specific phrases. Those recordings were then fed into software that could stitch sounds together to produce any word on demand.
Richard Cave, a speech therapist at University College London who works with ALS patients, helped build out the UK’s voice banking initiative starting in 2017. By his own account, after putting roughly 100 people through the process, the results were honest but limited. The voices were not awful, but they were not good either. They captured something of the person, but not the person. That was simply the best available technology at the time.
The physicist Stephen Hawking used a synthesized voice for decades and, by his own description, found it acceptable, even if he acknowledged it sounded a little tinny. Not everyone shares that tolerance. Jules Rodriguez, diagnosed with ALS, described his banked voice in a stand-up routine as sounding like Daft Punk played at quarter speed. The gap between what voice banking produced and what a human voice actually sounds like was wide enough to matter deeply.
Thirty Minutes of Audio, One Voice Clone
The shift came with AI-powered voice cloning. The core change is significant: instead of requiring a person to record hundreds of specific phrases in a controlled setting, modern systems can work from existing audio. Old voicemails, a recorded speech, a podcast appearance, anything that captures a person’s natural voice can serve as input.
Jules Rodriguez used recordings of his own voice to create an AI-generated voice clone in June 2024, through the company ElevenLabs. The result was striking enough that when his wife heard it for the first time, she cried. After months of gradually adjusting to hearing his voice less and less, the clone represented something that had seemed permanently lost.
The technology works by extracting specific features from the source audio: pitch, tone, prosody, breathing patterns. These elements are used to build a model of the voice, which can then generate speech from any typed text. The output includes the small human details that traditional voice banking missed entirely, breaths between words, natural intonations, even stutters. Speech pathologist Kate Nilson, who has tested the technology herself, describes playing a game with colleagues trying to distinguish her real voice from the clone. The difficulty of telling them apart is, by her account, genuinely unsettling.
Laura Sanders, the journalist who reported this story, created her own voice clone using ElevenLabs’ professional option, which required 30 minutes of audio input. The resulting clone could adjust speed, tone, and emotional register, and could even produce speech in other languages. The demonstration is instructive not because it is a novelty, but because it shows how accessible the technology has become.
One detail that speech pathologist Nilson highlights is particularly important: the ability to use old recordings opens the door for people whose speech has already deteriorated by the time they seek help. Under the old model, if a person’s voice had already changed significantly, it was often too late to bank it effectively. AI cloning changes that calculus.
What This Means Beyond the Medical Case
The story of Jules Rodriguez and Patrick Darling, a musician who uses his voice clone to sing and perform, points toward something broader than assistive technology. It raises a question about what a voice actually is and what it means to preserve one.
Traditional voice banking treated the voice as a functional tool: a means of producing intelligible speech. AI cloning treats it as something closer to an identity artifact, something worth capturing in its full texture, including its imperfections. The distinction matters because it changes what the technology is trying to do. It is not just restoring communication. It is restoring presence.
This framing has implications for how AI development in healthcare gets evaluated. The measure of success is not only accuracy or intelligibility. It is whether the person on the other end of the technology feels like themselves. That is a harder target to hit, and the fact that current systems are getting close to it suggests the gap between human and synthetic voice is narrowing faster than most people outside the field realize.
For anyone thinking about AI’s role in medicine, this is a useful case to hold onto. The technology here does not replace the speech therapist, the care team, or the human relationships around a patient. It gives those relationships something to work with that did not exist before.
In Short
AI voice cloning works by extracting the acoustic features of a person’s voice from existing recordings and building a model that can generate new speech in that voice. Unlike older voice banking methods, which required hundreds of controlled recordings and produced results that sounded mechanical, current systems can use informal audio and reproduce the natural rhythms, breaths, and intonations of a real voice. For people living with conditions like ALS, this represents a meaningful shift: not just a communication aid, but a way of preserving something that feels like identity. The technology is already in use, and the gap between synthetic and human voice is closing.
Based on reporting from Science News.