A self-powered wearable sensing-actuation system based on soft magnetoelasticity is introduced to enable assisted speaking without relying on vocal folds. The system, with a lightweight mass of approximately 7.2 g and a skin-like modulus of 7.83 × 10⁵ Pa, can detect extrinsic laryngeal muscle movements and convert them into high-fidelity electrical signals. These signals are then translated into speech using a machine learning algorithm with an accuracy of 94.68%. The system also includes an actuation component that produces voice signals without vocal fold vibration. The device is highly stretchable (164% maximum stretchability), waterproof, and comfortable to wear, making it suitable for patients with voice disorders during recovery. The system can capture three-dimensional laryngeal muscle movement signals, enabling accurate speech generation. The device has been tested for its acoustic performance, showing a sound pressure level (SPL) above the normal speaking threshold across the entire human hearing range. It is also resistant to sweat and can function effectively in damp environments. The system has been validated for its ability to recognize and classify laryngeal muscle movement signals, allowing for the selection of corresponding voice signals for output. The device has been tested with eight participants, achieving an overall accuracy of 94.68%. The system provides a feasible solution for patients with voice disorders to communicate during recovery, offering a non-invasive and comfortable alternative to existing solutions. The device's performance is supported by its unique design, which combines a magnetoelastic layer with a kirigami structure to enhance sensitivity and stretchability. The system has been shown to be durable, with no significant degradation in performance after 24,000 cycles of operation. The device's ability to generate clear and distinguishable voice signals, even under challenging conditions, highlights its potential for practical application in voice disorder treatment.A self-powered wearable sensing-actuation system based on soft magnetoelasticity is introduced to enable assisted speaking without relying on vocal folds. The system, with a lightweight mass of approximately 7.2 g and a skin-like modulus of 7.83 × 10⁵ Pa, can detect extrinsic laryngeal muscle movements and convert them into high-fidelity electrical signals. These signals are then translated into speech using a machine learning algorithm with an accuracy of 94.68%. The system also includes an actuation component that produces voice signals without vocal fold vibration. The device is highly stretchable (164% maximum stretchability), waterproof, and comfortable to wear, making it suitable for patients with voice disorders during recovery. The system can capture three-dimensional laryngeal muscle movement signals, enabling accurate speech generation. The device has been tested for its acoustic performance, showing a sound pressure level (SPL) above the normal speaking threshold across the entire human hearing range. It is also resistant to sweat and can function effectively in damp environments. The system has been validated for its ability to recognize and classify laryngeal muscle movement signals, allowing for the selection of corresponding voice signals for output. The device has been tested with eight participants, achieving an overall accuracy of 94.68%. The system provides a feasible solution for patients with voice disorders to communicate during recovery, offering a non-invasive and comfortable alternative to existing solutions. The device's performance is supported by its unique design, which combines a magnetoelastic layer with a kirigami structure to enhance sensitivity and stretchability. The system has been shown to be durable, with no significant degradation in performance after 24,000 cycles of operation. The device's ability to generate clear and distinguishable voice signals, even under challenging conditions, highlights its potential for practical application in voice disorder treatment.