Speech & Audio
Spoofed and deepfake speech detection held near earlier results in a harder challenge
In a 2021 paper, researchers describe ASVspoof 2021, the fourth edition of a challenge series that tests ways to protect voice-verification software, the kind of software that confirms who someone is from their voice, from spoofing and manipulation. The challenge kept the earlier logical access and physical access tasks and added a new deepfake speech detection task. Despite added variability in recording channels and audio compression, results on the logical access and deepfake tasks stayed close to results from previous ASVspoof editions. The physical access task showed that detecting attacks in real, variable physical spaces is still difficult.
What did the challenge test?
The challenge tested three tasks: logical access, physical access, and deepfake speech detection. It introduced new databases for each task. The paper describes the evaluation metrics, four challenge baselines, and the evaluation platform, along with a summary of results.
What did the results show?
The logical access and deepfake speech tasks produced results close to earlier ASVspoof editions, even though channel and compression variability made conditions harder. The physical access task showed difficulty in detecting attacks in real, variable physical spaces. ASVspoof 2021 was the first edition in which participants received no matched training or development data. That change reflected real conditions where the nature of spoofed and deepfake speech cannot be predicted with confidence. The authors describe the results as extremely encouraging and say they show substantial progress in the field in recent years.
What are common questions?
Why was there no matched training or development data?
This was the first edition without matched training or development data. The change reflected real conditions where the nature of spoofed and deepfake speech cannot be predicted with confidence.
What was new in ASVspoof 2021?
The challenge added a deepfake speech detection task, while continuing the logical access and physical access tasks. It also introduced new databases for the three tasks.
Which task still proved hard?
The physical access task showed difficulty. The results pointed to trouble detecting attacks in real, variable physical spaces.
Limits and what's next
The abstract describes one challenge edition, ASVspoof 2021, summarized in a single paper. It covers three tasks, uses new databases for each task, and compares results with earlier ASVspoof editions. The design removed matched training and development data, so conditions matched real use more closely. Because this is one challenge summary, the abstract reports no independent replication. The physical access result points to a remaining gap: detecting attacks in real, variable physical spaces is still difficult.
Source: Junichi Yamagishi et al., “ASVspoof 2021: accelerating progress in spoofed and deepfake speech detection,” 2021.