Tether released BrainWhisperer, an open-source brain-to-text engine that decodes neural signals into written language entirely on local devices. CEO Paolo Ardoino announced the system is now fully integrated into QVAC, the company's open-source on-device AI stack, and available to developers worldwide as a proof-of-concept capability. Developed by Tether Evo, the firm's research division, BrainWhisperer was designed to provide a communication pathway for individuals who have lost the ability to speak due to neurodegenerative conditions such as amyotrophic lateral sclerosis (ALS), addressing the limitations of existing solutions like eye-tracking letter boards that require several minutes to compose a single sentence.
BrainWhisperer Achieves 8.7% Word Error Rate in Validation Tests
BrainWhisperer is a redesigned variant of the Whisper speech-recognition model. Rather than processing audio, the system accepts electrical activity recorded from the brain's speech motor cortex—captured via implanted electrode arrays—and translates those signals directly into text. The approach bypasses the need for physical movement or sound production.
The SDK version of the model operates in under two gigabytes of memory with approximately fifty milliseconds of latency. In validation tests using real neural recordings from a single participant, it achieved an 8.7% word error rate, crossing below the ten percent threshold that researchers treat as the benchmark for real-world usefulness. A more complex variant of the same architecture placed fourth among 466 competing teams in an international brain-to-text challenge, though that version is not end-to-end and therefore was not included in the SDK release.
Tether Emphasizes Local Execution and Privacy Safeguards
Tether emphasizes that local execution is central to the technology's design. Because the decoding occurs entirely on the user's device through QVAC, no neural data is transmitted to external servers—a consideration for a system that reads intention directly from the brain. The company notes that the engine only processes speech the user is actively attempting to produce, and that future safeguards such as mental authentication could further reinforce user control.
Technology Remains Experimental and Requires Surgical Implant
BrainWhisperer remains experimental. The model was trained on data from only four participants, with the SDK version calibrated on a single individual. Adapting the system to a new user would require both a surgical implant and a personalized calibration period, presenting medical, safety, and regulatory hurdles. Tether describes the release as a foundational building block for future assistive products rather than a consumer-ready solution.
The research team indicates that the underlying methodology could eventually extend beyond speech to decode imagined images, sounds, and potentially intended movement. For now, the open-source release provides developers with a working on-device engine that demonstrates both the feasibility of accurate neural decoding and the practicality of running such systems on ordinary hardware without cloud dependency.
FAQ
What did Tether release?
Tether released BrainWhisperer, an open-source brain-to-text engine that decodes neural signals into written language entirely on local devices. CEO Paolo Ardoino announced the system is now fully integrated into QVAC, the company's open-source on-device AI stack.
How accurate is BrainWhisperer in validation tests?
In validation tests using real neural recordings from a single participant, BrainWhisperer achieved an 8.7% word error rate. A more complex variant of the same architecture placed fourth among 466 competing teams in an international brain-to-text challenge.
What are the current limitations of BrainWhisperer?
BrainWhisperer remains experimental. The model was trained on data from only four participants, with the SDK version calibrated on a single individual. Adapting the system to a new user would require both a surgical implant and a personalized calibration period.
