OBJECTIVE: A novel analytic approach for task-related high-gamma modulation (HGM) in stereo-electroencephalography (SEEG) was developed and evaluated for language mapping. METHODS: SEEG signals, acquired from drug-resistant epilepsy patients during a visual naming task, were analyzed to find clusters of 50-150 Hz power modulations in time-frequency domain. Classifier models to identify electrode contacts within the reference neuroanatomy and electrical stimulation mapping (ESM) speech/language sites were developed and validated. RESULTS: In 21 patients (9 females), aged 4.8-21.2 years, SEEG HGM model predicted electrode locations within Neurosynth language parcels with high diagnostic odds ratio (DOR 10.9, p < 0.0001), high specificity (0.85), and fair sensitivity (0.66). Another SEEG HGM model classified ESM speech/language sites with significant DOR (5.0, p < 0.0001), high specificity (0.74), but insufficient sensitivity. Time to largest power change reliably localized electrodes within Neurosynth language parcels, while, time to center-of-mass power change identified ESM sites. CONCLUSIONS: SEEG HGM mapping can accurately localize neuroanatomic and ESM language sites. SIGNIFICANCE: Predictive modelling incorporating time, frequency, and magnitude of power change is a useful methodology for task-related HGM, which offers insights into discrepancies between HGM language maps and neuroanatomy or ESM.