Testing the Universal Baby Language Hypothesis - Automatic Infant Speech Recognition with CNNs

This paper presents an application of convolutional neural networks (CNN) for the recognition of the so-called “Dunstan baby language” that consists of five “words” or phonemes used by babies of age under 3 months to communicate their needs before they start crying. The model was derived from a CNN architecture which was successfully applied by the authors for voice-based emotion detection. The input of the neural network is the spectrogram obtained from the audio records of babies' voices and is processed as a two-dimensional image. The architecture was trained for a set of 250 small duration recordings and was tested for other 65 recordings with a recognition rate of 89%. The length of all audio files is less than 1 second; the recordings were extracted from certified Dunstan language recordings. The most important original contribution of the paper is the recognition of the actual “baby words” (and not the baby cry as was done before). This architecture offers an efficient tool for the verification of the “universal baby language” hypothesis, according to which the language of infants does not depend on culture, family, etc.