Lattice Free MMI

Lattice-free maximum mutual information (LF-MMI) is a discriminative training criterion used to improve the accuracy of automatic speech recognition (ASR) and other sequence modeling tasks. Current research focuses on integrating LF-MMI into end-to-end (E2E) ASR systems, particularly using architectures like attention-based encoder-decoders and neural transducers, and employing algorithms like stochastic lattice descent for efficient training. This approach offers significant advantages over traditional methods by eliminating the need for lattice generation during training, leading to faster training times and improved performance, as demonstrated by consistent word error rate reductions across various datasets. The resulting improvements in ASR accuracy have broad implications for applications relying on speech processing, such as virtual assistants and speech-to-text systems.