I've been studying HMMs lately. I think getting an initial estimate of HMM emission and transition parameters using Segmental K-means training (or Viterbi training) before applying Baum-Welch re-estimation should result in the latter converging better. That's what the HInit tool in HTK does. AFAIK, it's done in the following way:
- Divide all examples (observation sequences) uniformly into as many segments as the number of states.
- Cluster the observations corresponding to each state, and estimate the GMM using the cluster set so that each cluster corresponds to one multivariate Gaussian.
- Do this repeatedly until convergence: get the Viterbi alignment of all examples, use it to get new segments, and estimate the parameters again using the previous step.
Please correct me if I'm wrong. Also, I have two questions:
- What kind of accuracy increase should be expected if using both Viterbi training and Baum-Welch re-estimation, instead of just the latter?
- What kind of accuracy should be expected if only using Viterbi training?
In my experience Viterbi training alone is often enough to get reasonable accuracy, at least in speech recognition.
Instead of doing the more costly Baum-Welch training you can spend your time better elsewhere, e.g. use deep neural networks instead of GMMs or collect more data.
Interesting - I've never tried Viterbi training. Maybe it is worth implementing after all. I plan to do a hybrid DNN-HMM (or whatever it is called now) with pylearn2 in a followup post.
Ask HN: Anybody got a piece of software to split an MP3 by speaker? (or output list of times where each speaker is speaking)
I've got an audiobook with a particularly obnoxious commentator interjecting dumb things every couple of minutes and I'd like to slice those out.
I'd do it manually if it were only a couple of edits, but there are easily 150+ comments spread through 6 hrs of material. Should be easy for a speech recognizer, just two voices with different genders.
You might be able to do this with independent component analysis (ICA). There is a block in sklearn, but you would have to write some Python.
Alternatively, you might look to see if the speakers are in separate channels (if recording is stereo). Then it would be really simple - just take one channel out and resave as mono!
If you have a sample I'd be curious to take a look - sounds like an interesting problem.
ICA helps you separate a superposition of two or more signals. Assuming the speakers are not speaking at the same time, he might rather want to use speaker identification methods to find what parts to mute (e.g. train a binary classifier that operates a spectrogram of the signal).
Good point. I was thinking of two people speaking at the same time, but that would probably be hard to listen to :). Speaker dependent muting is a much more reasonable approach
8 comments
[ 2.7 ms ] story [ 6.2 ms ] thread- Divide all examples (observation sequences) uniformly into as many segments as the number of states.
- Cluster the observations corresponding to each state, and estimate the GMM using the cluster set so that each cluster corresponds to one multivariate Gaussian.
- Do this repeatedly until convergence: get the Viterbi alignment of all examples, use it to get new segments, and estimate the parameters again using the previous step.
Please correct me if I'm wrong. Also, I have two questions:
- What kind of accuracy increase should be expected if using both Viterbi training and Baum-Welch re-estimation, instead of just the latter?
- What kind of accuracy should be expected if only using Viterbi training?
I've got an audiobook with a particularly obnoxious commentator interjecting dumb things every couple of minutes and I'd like to slice those out.
I'd do it manually if it were only a couple of edits, but there are easily 150+ comments spread through 6 hrs of material. Should be easy for a speech recognizer, just two voices with different genders.
Alternatively, you might look to see if the speakers are in separate channels (if recording is stereo). Then it would be really simple - just take one channel out and resave as mono!
If you have a sample I'd be curious to take a look - sounds like an interesting problem.
[1] http://lium3.univ-lemans.fr/diarization/doku.php