When mismatched training data outperform matched data

Emmanuel Vincent (1)

1 MULTISPEECH - Speech Modeling for Facilitating Oral-Based Communication

My talk will focus on robustness to background noise in distant-microphone speech recordings. I will introduce deep learning based techniques for speech enhancement and for acoustic modeling of speech. I will then report the results of a study on the impact of environment and microphone mismatches on the recognition accuracy. This study reveals that mismatched training data can sometimes outperform matched data. I will suggest a way to optimize the training set in order to exploit this finding.

Domaines: Traitement du signal et de l'image [eess.SP]

Emmanuel Vincent. When mismatched training data outperform matched data. Systematic approaches to deep learning methods for audio, Sep 2017, Vienna, Austria. ⟨hal-01588876⟩