Gait speed is an important parameter to characterize people's daily mobility. For real-world speed measurement, inertial sensors or Global Navigation Satellite System (GNSS) can be used on wrist, possibly integrated in a wristwatch. However, power consumption of GNSS is high and data are only available outdoor. Gait speed estimation using wrist-mounted inertial sensors is generally based on machine learning and suffers from low accuracy due to the inadequacy of using limited training data to build a general speed model that would be accurate for the whole population. To overcome this issue, a personalized model was proposed, which took unique gait style of each subject into account. Cadence and other biomechanically-derived gait features were extracted from wrist-mounted accelerometer and barometer. Gait features were fused with few GNSS data (sporadically sampled during gait) to calibrate the step length model of each subject through online learning. The proposed method was validated on 30 healthy subjects. For walking, it has achieved a median [Interquartile Range] RMSE, bias and precision of 0.05 [0.04-0.06], 0.00 [-0.01 0.00], and 0.06 [0.05 0.07] (m/s), respectively. For running, the errors are 0.14 [0.11 0.17], 0.00 [-0.01 0.02], and 0.18 [0.14 0.23] (m/s), respectively. Results demonstrated that the personalized model provided similar performance as GNSS. It used 50 times less training GNSS data than non-personalized method and achieved even better results. This parsimonious GNSS usage allowed extending battery life. The proposed algorithm met requirements for applications which need accurate, long, real-time, low-power, and indoor/outdoor speed estimation in daily life.