[YX]

It just occurred to me, is it because the ‘particular’ type of pruning we talked about for DTW is actually the same as the Viterbi Criterion (i.e. throw away the bad ones when they meet)?
And for the Viterbi algorithm, the real pruning is implemented by beam search, so we pretty much ‘throw away bad ones before they even meet’.

[YX]

Thank you for your explicit explanation!

[YX]

Thank you for your kind reply. I’m aware of the differences. So, does this mean that the statement is not correct, since it does not mention the most important part of voicing (i.e. vibration).

[YX]

Hi, I have a question regarding the objective function of these two algorithms.

I’m not sure if this this correct, but my understanding is that both algorithms can be used to weigh the empirical mean/variance of each possible alignment, the only difference is that while Viterbi has only 0 or 1 for weighing, B-F uses the true probability distribution to weigh each alignment.

The objective function for Viterbi is “argmax over w P(W|O) ∝ argmax over w P(O|W)P(W)”. But what is the objective function for B-F? Also, do all weights used in B-F have to add up to 1?

[YX]

Does the fundamental frequency of a complex wave (provided this frequency is present as one of the component waves) necessarily have the largest amplitude on the spectrum?

On page 54, it is noted that ‘the wave at the top left of figure 4.13 has a fundamental frequency of 100 Hz…Within each cycle there are six peaks, corresponding to a wave with six times the fundamental frequency. We may therefore expect the 600 Hz to have a relatively high amplitude.’

I don’t really get why the 600 Hz component should have a high amplitude in this case. Could you please explain it? Cheers!

Attachments:

You must be logged in to view attached files.

[Author]

Posts