Fine-tuning alignments

By default and standard in ASR, the frame step between feature frames is set to 10 ms, which limits the accuracy of MFA to a minimum of 0.01 seconds. When the --fine_tune flag is specified, the aligner does an extra fine-tuning step following alignment. The audio surrounding each interval’s initial boundary is extracted with a frame step of 1 ms (0.001s) and is aligned using a simple phone dictionary combined with a transcript of the previous phone and the current phone. Extracting the phone alignment gives the possibility of higher degrees of accuracy (down to 1ms).

Warning

The actual accuracy bound is not clear as each frame uses the surrounding 25ms to generate features, so each frame necessary incorporates time-smeared acoustic information.