Next: Posterior Probabilities to Up: System Description Previous: The Acoustic Vector

The Phone Probability Level

Figure 4 shows the input and output representation of the recurrent network for a sentence from the TIMIT database. The top part of the diagram shows the MEL+ acoustic features. The top twenty channels represent the power at mel-scale frequencies up to 8kHz. The bottom three channels represent the power, pitch and degree of voicing. Some features, like the high frequency fricative energy in /s/ and /sh/ and the formant transitions are clearly visible. The lower part of the diagram shows the output from the recurrent network. Each phone has one horizontal line with the width representing the posterior probability of the phone given the model and acoustic evidence. The vowels are placed at the bottom of the diagram and the fricatives at the top. As the TIMIT database is hand aligned, the dotted vertical lines show the boundaries of the known symbols. The identity of these hand aligned transcriptions is given on the top and bottom line of the diagram. Further information concerning this representation can be obtained from [20].

  
Figure 4: Input and output of the recurrent network for a TIMIT sentence ``she had your dark suit in greasy wash water all year''.

The complete sentence is ``She had your dark suit in greasy wash water all year''. Some of the phone labels may be read from the diagram directly; for example, the thick line in the bottom left is the initial silence and is then followed by a /sh/ phone half way up the diagram. Indeed, by making ``reasonable'' guesses and ignoring some of the noise, the first few phones can be read off directly as /sh iy hv ae dcl d/ which is correct for the first two words. Thus, the problem of connected word recognition can be rephrased as that of finding the maximum likelihood path through such a diagram while taking into account lexical and grammatical constraints.

Next: Posterior Probabilities to Up: System Description Previous: The Acoustic Vector