It supports linear transforms, MMI, boosted MMI and MCE discriminative training, feature-space discriminative training, and deep neural networks.
[3] Kaldi is capable of generating features like mfcc, fbank, fMLLR, etc.
Hence in recent deep neural network research, a popular usage of Kaldi is to pre-process raw waveform into acoustic feature for end-to-end neural models.
Kaldi has been incorporated as part of the CHiME Speech Separation and Recognition Challenge over several successive events.
[4][5][6] The software was initially developed as part of a 2009 workshop at Johns Hopkins University.