Fu Quchen ; Teng Zhongwei ; White Jules ; Powell Maria ; Schmidt Douglas C.

Abstract

Voice assistants, such as smart speakers, have exploded in popularity. It iscurrently estimated that the smart speaker adoption rate has exceeded 35% inthe US adult population. Manufacturers have integrated speaker identificationtechnology, which attempts to determine the identity of the person speaking, toprovide personalized services to different members of the same family. Speakeridentification can also play an important role in controlling how the smartspeaker is used. For example, it is not critical to correctly identify the userwhen playing music. However, when reading the user's email out loud, it iscritical to correctly verify the speaker that making the request is theauthorized user. Speaker verification systems, which authenticate the speakeridentity, are therefore needed as a gatekeeper to protect against variousspoofing attacks that aim to impersonate the enrolled user. This paper comparespopular learnable front-ends which learn the representations of audio by jointtraining with downstream tasks (End-to-End). We categorize the front-ends bydefining two generic architectures and then analyze the filtering stages ofboth types in terms of learning constraints. We propose replacing fixedfilterbanks with a learnable layer that can better adapt to anti-spoofingtasks. The proposed FastAudio front-end is then tested with two popularback-ends to measure the performance on the LA track of the ASVspoof 2019dataset. The FastAudio front-end achieves a relative improvement of 27% whencompared with fixed front-ends, outperforming all other learnable front-ends onthis task.