Abstract: To address the excessive steady-state speed estimation ripple and slow disturbances recovery of conventional extended state observers (ESOs) in sensorless permanent magnet synchronous motor control, an improved ESO-based speed estimation method is proposed for medium- and high-speed operation. The proposed method integrates sigmoid-gated gain scheduling with hierarchical bandwidth configuration to enable smooth, adaptive adjustment of the error-feedback gain and assign different poles to the physical and extended states, thereby jointly suppressing high-frequency noise propagation to the extended state in both amplitude and frequency domains. A Lyapunov-based analysis establishes the uniform ultimate boundedness of the observer error under bounded disturbances and a common Lyapunov condition. The proposed method is evaluated through simulations and hardware experiments under no-load steady-state conditions and load application and removal at 25% and 50% of the rated load. Compared with the linear extended state observer, the proposed method reduces the steady-state standard deviation by approximately 80% in simulation and 63% and in hardware experiments, while reducing the corresponding peak-to-peak values by approximately 73% and 64%, respectively. Under the two load conditions, the maximum speed drop during load application and the maximum overshoot during load removal are reduced by an average of approximately 52% in simulation, while the recovery time is shortened by approximately 72% on average. In hardware experiments, the corresponding average reductions in speed drop and overshoot are approximately 59% and 62%, respectively. The results demonstrate that the proposed method effectively suppresses steady-state speed estimation ripple and improves dynamic disturbance-rejection performance under the specified medium- and high-speed operating conditions.