Learn how Microsoft research uncovers backdoor risks in language models and introduces a practical scanner to detect tampering and strengthen AI security.
Abstract: This paper proposes an unsupervised, batch-type, class-based language model adaptation method for spontaneous speech recognition. The word classes are automatically determined by maximizing ...