BrainWhisperer is Tether’s Brain-to-text project. Tether is earmarking resources to build technologies that push the borders of intracranial electrocortical decoding. The latest result is a variable 98.3% accuracy in translating brain signals, a remarkable neurotechnological breakthrough.
“Do you know where it might have gone? I am an artist, lost in my own vision. I don’t think so anymore.”
Tether Evo’s BrainWhisperer correctly decoded these sentences from phonemes signaled by intracranial Brain Computer Interface (BCI) Implants in research subjects, re-energizing ongoing scientific progress in Brain-to text translation for the speech-impaired and paralyzed.
The brain’s physiology gives room for Neurotechnological efforts to improve understanding of impairments and engineer effective management/remedies through brain-computer connectivity. Tether Evo is exploring this, advancing discoveries, and building on existing progress in this area as part of the Brain OS revolution. Brain OS, the Brain Operative System designed by Tether’s engineering team and built on top of Tether’s QVAC AI platform, aims to create an open-source brain operative system, that connects to the user’s personal brain computer-interface and other wearables, enhancing the human mind with the power of AI, working directly on device to maximize privacy, ensuring that each person remains always in control of his most intimate thoughts, while allowing our minds to become 100 times more expressive and intelligent leveraging the most modern AI techniques.
Brain OS is Tether’s drive to improve abilities for people with impairments through biotechnology, neuroscience, and artificial intelligence. It embraces Michael Oliver’s theory of social disability, with a premise on enablement for people with special needs and the amelioration of impairments.
BrainWhisperer, a key component of the Brain OS, is a hyper-augmented intracortical neural signal-decoding framework with a base architecture built on OpenAI’s ASR (Automatic Speech Recognition) model, Whisper. It complements Whisper’s text-decoding capabilities by tokenizing neural signals and integrating a LoRA (Low-Rank Adaptation)-Fine-Tuned AI model to improve signal transcription and achieve progressively lower WER (Word Error Rate) across separate trial sessions.
Tether’s BrainWhisperer benchmarked
Tether Evo’s BrainWhisperer ranked fourth against 466 participants in the recent Brain-to-Text ‘25 Kaggle Competition with an impressive 1.78% WER, only 0.25% WER short of first place. The Kaggle Competition benchmarks state-of-the-art Brain-to-Text decoding models using standardized core performance parameters.
Tether Evo’s entry for the competition is a multi-stage pipeline Brain-to-Text transcription system designed to maximize accuracy through ensemble learning and sophisticated ranking. It features an ensemble of five models, each for the three commonly used brain-to-text decoding datasets (Willet, Card, and Kunz), trained using Adam Optimizer with a 100-epoch cosine learning rate. A Weighted-Finite-State-Transducer (WFST) converts the phoneme sequence that each model generates into potential transcriptions.
BrainWhisperer’s performance in the competition reflects the series of groundbreaking advances made by the Evo team and outlines possible improvements toward achieving a best-in-class, unconditionally precise Brain-to-text decoding system.
Tether Evo’s ongoing research on BCIs and neural signal decoding has yielded a deeper understanding and greater productivity in brain-machine communication through a series of breakthroughs, including an impressively accurate transcription of BCI implant signals and cross-subject contextual training for signal decoding. Tether’s AI and Data team has also made outstanding progress in developing model fine-tuning solutions for phoneme decoding and building performant health applications, such as their own QVAC Health – an AI-powered health assistance that provides expert wellness recommendations based on user-provided health data.
Near-accurate BCI implants’ signal transcription
Tether Evo achieved a remarkable 99.4% (0.6% WER) accuracy in text decoding by using its Autoregressive transformer model to decode CNN (Convolutional Neural Network) signals. The multi-layer system includes an MFCC (Mel-Frequency Cepstral Coefficients) predictor that converts ECoG or EEG recordings into signals decodable by a Bidirectional Autoregressive Transformer (BART). The BART decodes the neural activity into phonemes and proper words. The phoneme provides a clearer understanding of the translation pattern, enabling integration with AI models to enhance final text decoding.
Cross-subject contextual Training for signal decoding
The calibration stage introduces a significant operational overhead in Brain-to-text decoding. The current Subject-specific calibration standard requires a new calibration for each test subject. Each calibration takes hours to days, imposing time and resource constraints on research teams.
Tether Evo is pioneering research in Cross-subject neural signal decoding, introducing a universal decoding framework that translates signals from separate intercortical recordings.
The Cross-subject contextual Training system is based on the theory of a Common Low-Dimensional structure in neural representations of human speech. That is, the idea that our brain activity follows a similar pattern, with minor differences, when we speak, even though it doesn’t appear so at a glance.
It also uses Hierarchical CTC (Connectionist Temporal Classification) to develop a closed-loop phoneme prediction training module that achieves higher accuracy and faster convergence than Vanilla (standard) CTC.
Using an affine transformation formula, the Cross-subject contextual Training system provides a base structure/library for resolving minor differences that account for salient inter-personal structural variations in neural recordings, enabling the framework to translate recordings from multiple test subjects.
Early test results show promise compared to the accuracy of cutting-edge Brain-to-text (BIT) translation models, with a 7.5% WER on record (against 6.39% WER for cutting-edge BIT models) for standard CTC and a 6.67% WER (against 5.9% WER for cutting-edge BIT models) for Hierarchical CTC.
Pursuing Efficacy and Non-invasiveness
Resolute on achieving optimal brain-machine connectivity and advancing BCI research, Tether Evo aims first to develop an efficient neural signal decoding system and to improve ergonomics. BrainWhisperer and the cross subject signal-decoding framework focus on improving the efficiency of Brain-to-text transcription.
Ergonomics-wise, Tether Evo is exploring non-invasive alternatives for intracranial BCI implants. Contemporary BCI implants are invasive, requiring surgical interaction with the human brain. In contrast, non-invasive implants can be placed on the skin or plugged into the ear. They use Surface Electromyography (sEMG) electrodes to detect electrical signals from muscles, which can be translated into text or machine commands, depending on the placement and complementary technologies.
The Tether Evo team believes in the potential of Non-invasive implants to rapidly improve the user experience, scale the application, and accelerate the advancement of BCI technology. A key bottleneck for non-invasive implants is the refinement of signals due to cross-interference from muscle and brain activity. Tether Evo is collaborating with other research teams to develop more potent non-invasive solutions for BCI technology.
