Proceedings Fonetik 2009 - Institutionen för lingvistik

More documents

Recommendations

Info

Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm Universitythe background noise relative to what wouldhave been possible using a 16-bit/sample representation.It is puzzling that such crude signalrepresentations are used by a technology claimingto work on “details”. But the degradation ofthe amplitude resolution becomes even worseas a “filter” that introduces a coarser quantizationusing 3 units’ steps reduces the 256 levelsof the 8-bit representation to only 85 quantizationlevels (ranging from -42 to +42). This verylow sample resolution (something around 6.4-bit/sample), resulting in a terrible sound quality,is indeed the basis for all the subsequent signalprocessing carried out by the LVA-technology.The promise of an analysis of “minute” detailsin the speech waveform cannot be taken seriously.Figure 1 displays a visual analogue of thesignal degradation introduced by the LVAtechnology.Figure 1. Visual analogs of LVA-technology’sspeech signal input. The 256×256 pixels image,corresponding to 16 bit samples, is sampleddown to 16×16 pixels (8 bit samples) and finallydown-sampled to approximately 9×9 pixelsrepresenting the ±42 levels of amplitude encodingused by the LVA-technology.The core analysis procedureIn the next step, the LVA-technology scans thatcrude speech signal representation for “thorns”and “plateaus” using triplets of consecutivesamples."Thorns"According to Nemesysco’s definition, thornsare counted every time the middle sample ishigher than the maximum of the first and thirdsamples, provided all three samples are abovean arbitrary threshold of +15. Similarly, a thornis also detected when the middle sample valueis lower than the minimum of both the first andthe third samples in the triplet and all threesamples are below -15. In short, thorns are localmaxima, if the triplet is above +15 and localminima if the triplet is below -15. Incidentallythis is not compatible with the illustration providedin fig. 2 of the patent, where any localmaxima or minima are counted as thorns, providedthe three samples fall outside the(-15;+15) threshold interval."Plateaus"Potential plateaus are detected when the samplesin a triplet have a maximum absolute amplitudedeviation that is less than 5 units. The±15 threshold is not used in this case but tocount as a plateau the number of samples in thesequence must be between 5 and 22. The numberof occurrences of plateaus and their lengthsare the information stored for further processing.A blind technologyAlthough Nemesysco presents a rationale forthe choice of these “thorns” and “plateaus” thatsimply does not make sense from a signal processingperspective, there are several interestingproperties associated with these peculiar variables.The crucial temporal information is completelylost during this analysis. Thorns and plateausare simply counted within arbitrarychunks of the poorly represented speech signalwhich means that a vast class of waveformscreated by shuffling the positions of the thornsand plateaus are indistinguishable from eachother in terms of totals of thorns and plateaus.Many of these waveforms may even not soundlike speech at all. This inability to distinguishbetween different waveforms is a direct consequenceof the information loss accomplished bythe signal degradation and the loss of temporalinformation. In addition to this, the absolutevalues of the amplitudes of the thorns can bearbitrarily increased up to the ±42 maximumlevel, creating yet another variant of physicallydifferent waveforms that are interpreted asidentical from the LVA-technology’s perspective.222
Proceedings, FONETIK 2009, Dept. of Linguistics, Stockholm UniversityThe measurement of the plateaus appears toprovide only very crude information and is affectedby some flaws. Indeed, the program codeallows for triplets to be counted as both thornsand plateaus. Whether this is intentional or justa programming error is impossible to determinebecause there is no theoretical model behind theLVA-technology against which this could bechecked. In addition, what is counted as a plateaudoes not even have to look like a plateau.An increasing or decreasing sequence of sampleswhere differences between adjacent samplesare less than ±5 units will count as a plateau.Only the length and the duration of theseplateaus are used and because the ±5 criterion isactually a limitation on the derivate of the amplitudefunction, large amplitude drifts can occurin sequences that are still viewed by LVAtechnologyas if they were flat. Incidentally,given that these plateaus can be up to 22 sampleslong, the total span of the amplitude driftwithin a plateau can be as large as 88 units,which would allow for a ramp to sweep throughthe whole range of possible amplitudes (-42 to+42). This is hardly compatible with the notionof high precision technology suggested byNemesysco. Finally, in addition to the countingof plateaus, the program also computes thesquare root of the cumulative absolute deviationfor the distribution of the plateau lengths.Maybe the intention was to compute the standarddeviation of the sample distribution andthis is yet another programming error but sincethere is no theoretical rationale it is impossibleto discuss this issue.Assessing the speaker’s emotionalstateThe rest of the LVA-technology simply uses theinformation provided by these four variables:(1) the number of thorns per sample, (2) thenumber of counts of plateaus per sample, (3)the average length of the plateaus and (4) thesquare root of their cumulative absolute deviation.From this point on the program code is nolonger related to any measureable physicalevents. In the absence of a theoretical model,the discussion of this final stage and its outcomeis obviously meaningless. It is enough topoint out that the values of the variables used toissue the final statements concerning thespeaker’s emotional status are as arbitrary asany other and of course contain no more informationthan what already was present in thefour variables above.Examples of waveforms that becomeassociated with “LIES” 2Figure 2 shows several examples of a syntheticvowel that was created by superimposing withthe appropriate delays to generate different fundamentalfrequencies, a glottal pulse extractedfrom a natural production.After calibration with glottal pulses simulatinga vowel with a 120 Hz fundamental frequency,the same glottal pulses are interpretedas indicating a “LIE” if the fundamental frequencyis lowered to 70 Hz whereas a raise infundamental frequency from 120 Hz to 220 Hzis detected as “outsmart”. Also a fundamentalfrequency as low as 20 Hz is interpreted as signallinga “LIE”, relative to the 120 Hz calibration.Using the 20 Hz waveform as calibrationand testing with the 120 Hz is detected as “outsmart”.A calibration with the 120 Hz waveabove followed by the same wave contaminatedby some room acoustics is also interpreted as“outsmart”.The illusion of a serious analysisThe examples above suggest that the LVAtechnologygenerates outputs contingent on therelationship between the calibration and the testsignals. Although the signal analysis performedby the LVA-technology is a naive and ad hocmeasurement of essentially irrelevant aspects ofthe speech signal, the fact that some of the "detectedemotions" are strongly dependent on thestatistical properties of the "plateaus" leads tooutcomes that vaguely reflect variations in F0.For instance, the algorithm's output tends to bea “lie” when the F0 of the test signal is generallylower than that of the calibration. The mainreason for this is that the program issues “lie”-warnings when the number of detected “plateaus”during the analysis phase exceeds by acertain threshold the number of “plateaus”measured during calibration. When the F0 is2 The amplitudes of the waveforms used in this demonstrationare encoded at 16 bit per sample.223
Page 1 and 2:
Department of LinguisticsProceeding
Page 3 and 4:
Proceedings, FONETIK 2009, Dept. of
Page 5 and 6:
Page 7 and 8:
Page 9 and 10:
Page 11 and 12:
Page 13 and 14:
Page 15 and 16:
Page 17 and 18:
Page 19 and 20:
Page 21 and 22:
Page 23 and 24:
Page 25 and 26:
Page 27 and 28:
Page 29 and 30:
Page 31 and 32:
Page 33 and 34:
Page 35 and 36:
Page 37 and 38:
Page 39 and 40:
Page 41 and 42:
Proceedings, FOETIK 2009, Dept. of
Page 43 and 44:
Page 45 and 46:
Page 47 and 48:
Page 49 and 50:
Page 51 and 52:
Page 53 and 54:
Page 55 and 56:
Page 57 and 58:
Page 59 and 60:
Page 61 and 62:
Page 63 and 64:
Page 65 and 66:
Page 67 and 68:
Page 69 and 70:
Page 71 and 72:
Page 73 and 74:
Page 75 and 76:
Page 77 and 78:
Page 79 and 80:
Page 81 and 82:
Page 83 and 84:
Page 85 and 86:
Page 87 and 88:
Page 89 and 90:
Page 91 and 92:
Page 93 and 94:
Page 95 and 96:
Page 97 and 98:
Page 99 and 100:
Page 101 and 102:
Page 103 and 104:
Page 105 and 106:
Page 107 and 108:
Page 109 and 110:
Page 111 and 112:
Page 113 and 114:
Page 115 and 116:
Page 117 and 118:
Page 119 and 120:
Page 121 and 122:
Page 123 and 124:
Page 125 and 126:
Page 127 and 128:
Page 129 and 130:
Page 131 and 132:
Page 133 and 134:
Page 135 and 136:
Page 137 and 138:
Page 139 and 140:
Page 141 and 142:
Page 143 and 144:
Page 145 and 146:
Page 147 and 148:
Page 149 and 150:
Page 151 and 152:
Page 153 and 154:
Page 155 and 156:
Page 157 and 158:
Page 159 and 160:
Page 161 and 162:
Page 163 and 164:
Page 165 and 166:
Page 167 and 168:
Page 169 and 170:
Page 171 and 172: Proceedings, FONETIK 2009, Dept. of
Page 203 and 204: Proceedings, FOETIK 2009, Dept. of
Page 205 and 206: Proceedings, FOETIK 2009, Dept. of
Page 221: Proceedings, FONETIK 2009, Dept. of
Page 227: Department of LinguisticsPhonetics
show all

Proceedings Fonetik 2009 - Institutionen för lingvistik

Create successful ePaper yourself

Delete template?

Save as template?