Pattern Comparison Techniques By Sarita Jondhale 1 Pattern

Pattern Comparison Techniques By Sarita Jondhale 1

Pattern Comparison Techniques n n n The output of the front end spectral analysis is in the form of vectors. The test pattern T is the set containing many vectors. The reference pattern R is the set containing many vectors. The goal of pattern comparison stage is to determine the dissimilarity of each vector in T to each vector of R The reference pattern should be such that there should be minimum dissimilarity By Sarita Jondhale 2

Pattern Comparison Techniques n To determine the global similarity of T and R we will consider the following problems: n n n T and R generally are of unequal length w. r. t. time duration due to different speaking rates across different talkers T and R need not line up in time in any simple or well prescribed manner this is because different sounds cannot be varied in duration to same degree. Vowels are easily lengthened or shortened but consonants cannot change in duration We need a way to compare a spectral vectors By Sarita Jondhale 3

Speech Detection n n Also called as End point detection The goal of speech detection is to separate speech signal with a background signal. The need of speech detection occurs in many applications in telecommunications For automatic speech recognition, end point detection is required to isolate the speech of interest so as to be able to create a speech pattern or template. By Sarita Jondhale 4

Speech Detection n n Speech must be detected so as to provide the best patterns for the recognition Best patterns means which provides highest recognition accuracy By Sarita Jondhale 5

Speech Detection n n Accurate detection of speech is a simple problem when speech is produced in a relatively noise free environment It becomes difficult task when the environment is noisy By Sarita Jondhale 6

Speech Detection n n First factor: during speech, talker produces sound like lip smacks, heavy breathing and mouth clicks Mouth click with speaking: The mouth click is produced by opening the lips prior to speaking or after speaking, the noise of clicking is separate from the speech signal and the energy level is comparable to speech energy signal By Sarita Jondhale 7

By Sarita Jondhale 8

By Sarita Jondhale 9

Speech Detection n Heavy breathing with speaking: unlike the mouth click the heavy breathing noise is not separated from the speech and therefore makes accurate end point detection quite difficult By Sarita Jondhale 10

Speech Detection n n Second factor: environmental noise The ideal environment for talking is the quite room with no acoustic noise signal generators other than that produced by the speaker. By Sarita Jondhale 11

Speech Detection n n Ideal environment is not possible practically Have to consider speech produced n n In noisy backgrounds (fans, machinery) In non stationary environments (presence of door slams, irregular road noise, car horns) With speech interference ( as from TV, radio, or background conversations) And in hostile circumstances ( when the speaker is stressed) By Sarita Jondhale 12

Speech Detection n These interfering signals are some what like speech signals therefore accurate end point detection become difficult By Sarita Jondhale 13

Speech Detection n Third factor: distortion introduced by the transmission system over which speech signal is sent. By Sarita Jondhale 14

Speech Detection n The methods for speech detection is broadly classified into three approaches n n n The explicit approach The implicit approach The hybrid approach By Sarita Jondhale 15

The explicit approach By Sarita Jondhale 16

The explicit approach n n n The speech signal is first measured and feature measurement is made The speech detection method is then applied to locate and define the speech events The detected speech is sent to the pattern comparison algorithm, and finally the decision mechanism chooses the recognized word By Sarita Jondhale 17

The explicit approach n n For signals with a stationary and low level noise background, the approach produces reasonably good detection accuracy The approach fails often when the environment is noisy or the interference in non stationary By Sarita Jondhale 18

The implicit approach By Sarita Jondhale 19

The implicit approach n n This approach detects the speech detection problem simultaneously with the pattern matching and recognition-decision process It recognizes that the speech events are almost always accompanied by a certain acoustic background By Sarita Jondhale 20

The implicit approach n n n The unmarked signal sequence is processed by the pattern matching module in which all possible end points sets are considered The decision mechanism provides ordered list of the candidate words as well as corresponding speech locations The final result is best candidate and its associated end points. By Sarita Jondhale 21

The implicit approach n n Depending on the word recognized the boundary locations could inherently be different with the implicit method (feedback) With explicit method only a single choice of boundary locations is made By Sarita Jondhale 22

The implicit approach By Sarita Jondhale 23

The implicit approach n Advantages & disadvantages n n Requires heavy computations But offers higher detection accuracy than the explicit approach By Sarita Jondhale 24

The hybrid approach n n n This is the combination of both implicit and explicit approaches Uses the explicit method to obtain several end points sets for recognition processing and implicit method to choose the alternatives The most likely candidate word and the corresponding end points as in implicit approach, are provided by the decision box. By Sarita Jondhale 25

The hybrid approach By Sarita Jondhale 26

The hybrid approach n n Computational load is equivalent to explicit method And accuracy comparable to implicit method By Sarita Jondhale 27