I have been studying HMM implementation approaches on ASR for the last couple of weeks. This probabilistic model is very new to me. I am currently using a Python package called Pomegranate to implement an ASR model of my own for the Librispeech dataset. I plan to use my own small-size dataset once I feel comfortable assessing the results of this one with Librispeech.
My problem: Librispeech (and my own dataset) has word-level transcription labels of the audio. The audio files have several word utterances each. Upon generating the MFCCs, I am not sure how to initialize the HMM matrices since the MFCCs, to my knowledge, capture phoneme level context at 10ms windows whereas I am trying to use the current word-level labels. There is no match up to which word each MFCC window belongs. Are the unique words in my corpus to be considered as the individual states in the transition probability matrix? I’m missing the point of how the extracted MFCCs are fed to the model for initialization and/or training?
I’ve been stumped on this for several days and I can’t seem to understand a clear cut explanation in the literature I have read. Any advice and help is very very much appreciated.
