Australian researchers are working on a brain-computer interface (BCI) designed to enable people with severe speech impairments to communicate again. The technology, developed by the start-up Fluent, a spin-off from the University of Melbourne, records brain signals generated when a person attempts to speak and uses artificial intelligence to translate these into text or spoken language.
The technology is intended for people with neurological conditions such as amyotrophic lateral sclerosis (ALS) and multiple sclerosis (MS), who, due to their illness, have difficulty speaking or eventually lose their voice completely. Whilst many existing communication aids rely on eye movements, touch or letter boards, the new system is designed to enable communication without uttering a single word or pressing a button. According to the researchers, the new approach offers a less invasive alternative to existing brain-computer interfaces, which typically require complex brain surgery.
AI translates brain signals into words
The implant is not placed inside the brain, but beneath the scalp and on top of the skull, at the level of the motor cortex. This area of the brain controls, amongst other things, the muscles involved in speech. According to co-founder and biomedical engineer Tim Mahoney, every movement of the mouth and jaw produces a unique pattern of electrical activity in the motor cortex. The same patterns also occur when someone with a speech disorder tries to speak.
“If you compare electrical signals to QR codes, then every mouth and jaw movement produces a different code,” explains Mahoney. “Our technology records these codes in the correct sequence and can deduce from them what someone is trying to say.” The recorded brain signals are then analysed by a machine learning model that translates them into text or speech. This creates a means of communication that is entirely controlled by brain activity.
No brain surgery required
A key advantage of the new technology is that the electrodes can be placed outside the skull. During his PhD research, Mahoney demonstrated that the quality of the measured electrical signals is comparable to that of electrodes placed beneath the scalp. This means that invasive brain surgery is not necessary. The initial validation studies took place in specially shielded laboratories at the Aikenhead Centre for Medical Discovery at St Vincent’s Hospital in Melbourne. During these experiments, participants had 144 electrodes attached to their scalps, which recorded brain activity whilst they spoke sentences, mimicked mouth movements or simply imagined that they were speaking.
Based on this data, the researchers compiled the largest English-language dataset of its kind. In collaboration with a Japanese research team, which had access to an even larger dataset, an AI model was developed that was able to select the correct sentence from 128 possible options with 96 per cent accuracy.
Clinical trials
The first clinical trials using Fluent’s implantable electrodes are set to begin later this year. According to Mahoney, it had previously been assumed that such accuracy was only possible with electrodes implanted directly into the brain. The researchers expect that this less invasive approach will make the technology more accessible to a wider group of patients. According to Fluent, the implant has a safety profile that is comparable to, or even more favourable than, that of a cochlear implant.
The University of Melbourne sees Fluent as an example of how university-led innovation can contribute to new treatment technologies for neurological conditions. According to Vice-Chancellor for Research Mark Cassidy, the start-up demonstrates how scientific research can develop into technology that enhances the independence of people with severe disabilities and improves their quality of life.
Stanford innovation
In 2025, researchers at Stanford University demonstrated the first real-time brain-computer interface (BCI) capable of decoding inner speech, the silent voice people use when thinking, with 74% accuracy. Their study involved four participants with complete paralysis who had microelectrodes implanted in the motor cortex. An AI model was trained to distinguish neural activity generated when participants attempted to speak aloud and when they only imagined speaking. Despite weaker signals, the system successfully decoded imagined sentences from a 125,000-word vocabulary.
Researchers also introduced a "thought password" that users could mentally recite to activate decoding, achieving more than 98% reliability and providing an important safeguard against unintended interpretation of thoughts. The findings suggest future BCIs could enable people with severe speech impairments, such as ALS, to communicate more naturally and independently. Improved sensors, hardware and AI algorithms are expected to further increase accuracy and speed.
