A Random Presentation
1 like7,016 views
The document discusses Muhammad Adil Raja's research project on non-intrusive real-time quality assessment of speech for VoIP using Hidden Markov Models. It describes using HMMs with a mixture density function and a few Gaussians to calculate the probability of an observation sequence and find the most likely state sequence that could have emitted it, based on the Viterbi algorithm and dynamic programming. It asks for recommendations on applications that can access the Viterbi probability, mentioning HTK, HMM Toolbox by Kevin Murphy, and Sphinx4 as options already considered.
A Random Presentation
- 1. Muhammad Adil RajaMuhammad Adil Raja PakistanPakistan ECE DepartmentECE Department Foundation BuildingFoundation Building Wireless Access Research GroupWireless Access Research Group
- 2. My Research ProjectMy Research Project Non-intrusive Real-time quality of speechNon-intrusive Real-time quality of speech assessment for VoIP.assessment for VoIP.
- 4. Hidden Markov Models HMMHidden Markov Models HMM λλ=(A, B,=(A, B, ππ)) Discrete/Continuous.Discrete/Continuous. Mixture density Function.Mixture density Function. A few Gaussians.A few Gaussians. P(O|P(O|λλ)) Expectation MaximizationExpectation Maximization Finding the most likely state-sequence.Finding the most likely state-sequence.
- 5. Most Likely State SequenceMost Likely State Sequence Which could have emitted the ObservationWhich could have emitted the Observation Sequence.Sequence. Its calculated based on Viterbi algorithm.Its calculated based on Viterbi algorithm. It is based on dynamic programming.It is based on dynamic programming. Breaks the problem into small parts andBreaks the problem into small parts and arrives at a solution by enhancing a sub-arrives at a solution by enhancing a sub- problem optimality criteria (Irrelevantproblem optimality criteria (Irrelevant here).here).
- 7. My ProblemMy Problem In the context of speech recognition I amIn the context of speech recognition I am looking for a configurable application fromlooking for a configurable application from which I can access the Viterbi probability.which I can access the Viterbi probability. HTK (Maaassive code and documentation)HTK (Maaassive code and documentation) HMM ToolBox by Kevin Murphy (It isHMM ToolBox by Kevin Murphy (It is generic, training is a huge problem).generic, training is a huge problem). Sphinx4 by Carnegie melon university (ItSphinx4 by Carnegie melon university (It has some potential and I am looking into it)has some potential and I am looking into it)
- 8. Thank you!Thank you! I am looking for information and some niceI am looking for information and some nice ideas.ideas.