During the first period of their lives, infants and young birds show comparable phases of vocal development: they first listen to their parents/guardians in order to build a neural representation of the perceived auditory stimulus, and then they begin to produce sounds that progressively approximate the song of their guardian. This learning phase is called the sensorimotor phase and is characterized by the presence of babbling. It ends when the song crystallizes, i.e. when it becomes similar to that produced by adults. There are similarities between the brain pathways responsible for sensorimotor learning in humans and in birds. In both cases, one pathway deals with vocal production and involves direct projections from auditory to motor areas, and another pathway deals with vocal learning, imitation and plasticity.
In birds, these brain circuits are exclusively dedicated to song learning, which in fact motor allows the production of sounds and a sensory response models either how the sound is perceived or how it shapes the reward. The inputs and outputs of these functions are in many an ideal model for exploring the neural mechanisms of vocal learning by imitation. This dissertation aims to build a model of bird song learning by imitation. Many previous studies have attempted to implement imitative learning in computer models and share a common structure. These models include learning mechanisms and possibly exploration and evaluation strategies. In these models, a control function spaces: motor space (motor parameters), sensory space (real sounds), perceptual space (low-dimensional representation of the sound) or goal space (non-perceptual representation of the target sound). The first proposed model is an inverse theoretical model based on a simplified speech learning model where the sensory space coincides with the motor space (i.e. there is no production). Such a simplification allows us to study how to introduce biological assumptions (e.g., a nonlinear response) into a voice learning model and which parameters most influence the computational power of the model. In order to have a complete model (capable of perceiving and producing sounds), we needed a motor control function capable of reproducing sounds similar to real data. We analyzed the ability of WaveGAN (a generating network) to produce realistic canary songs. In this model, the input space becomes the latent space after training and allows the representation of a high
in a lower dimensional variety. We obtained realistic canary songs using only three dimensions for the latent space. Quantitative and qualitative analyses demonstrate the interpolation capabilities of the model, suggesting that the model can be used as a driving function in a speech learning model. The second version of the model is a full speech learning model with a complete action-perception loop (it includes motor space, sensory space and perceptual space). The sound production is realized by the GAN generator obtained previously. A recurrent neural network classifying the syllables serves as a perceptual sensory response. The correspondence between the perceptual space and the motor space is learned by an inverse model. Preliminary results show the impact of learning rate when different sensory response functions are implemented.