Background: Speech intelligibility alteration is a frequent consequence of oral/oropharyngeal cancer. The development of automatic speech recognition (ASR) systems could overcome the limitations of perceptual speech assessment. Objective: To predict speech intelligibility after treatment of oral or oropharyngeal cancer using scores from an ASR system. Methods: Spontaneous speech of patients was recorded during a semi-structured interview. Six experts evaluated the subjects' intelligibility perceptually. An ASR system (TDNNf-HMM) trained on healthy adult speech and adapted to phoneme recognition was also used. Automatic scores were computed: phonemic scores, confidence scores. LASSO regression was used to select the parameters from the ASR system that best predicted intelligibility. Results: Spontaneous speech of 25 patients was recorded. LASSO regression led to retain 3 parameters: number of sonants recognized per second, proportion of occlusives, and average confidence score of fricatives. These three parameters present a strong correlation (rs=0.91) with the perceptual score (expert panel). This automatically predicted score is stable and reliable (5-block cross-validation: rs= 0.90). Conclusion: The use of ASR systems in the measurement of intelligibility in ENT oncology is promising. An optimization of these systems for pathological speech would open new perspectives for the determination of fine low-level speech deficits to adapt therapeutic objectives.