Output details

Source Separation of Convolutive and Noisy Mixtures Using Audio-Visual Dictionary Learning and Probabilistic Time-Frequency Masking

Speech-on-speech audio mixtures in typical (reflective) acoustic conditions face the permutation problem (i.e., pulling out the wrong voice in place of the desired target speaker), which is a major obstacle in the wide-spread deployment of source separation techniques. The proposed method solves it using visual information about a speaker and a practical machine learning technique to help extract the relevant speech, which has the potential to improve the real-world performance of many speech technologies in a vast array of application scenarios including: biometric security, automatic speech recognition, speaker localisation, speech enhancement for hearing aids, audio re-mastering, surveillance and telecommunications.