A talking and singing robot is being developed based on the mechanical model of the human vocal system by using mechatronics and feedback control technologies. While various ways of vocal sound production have been actively studied so far, a mechanical construction of the vocal system is considered to advantageously realize natural vocalization with its fluid dynamics. Motors are employed for the manipulation of the mechanical system. The robot is able to adaptively learn the relations between motor control parameters and the produced vocal sounds using an auditory feedback with neural networks, by mimicking a human vocalization. This paper presents the construction of the robot and its taking and singing performance, together with the adaptive control for the pitch and phoneme learning. The robot generates vowel and consonant sounds of different pitches by dynamically controlling the vocal cords, vocal tract and nasal cavity.