Timo Baumann : Home Page > SphinxPerformanceHier wird die Leistung (Geschwindigkeit und Erkennungsgenauigkeit) von Sphinx in unterschiedlichen Konfigurationen festgehalten.
Die aktuellen Ergebnisse beschränken sich auf die Erkennungsgenauigkeit von Sphinx auf dem TIDIGITS-Testmaterial bei Nutzung unterschiedlicher akustischer Modelle und gleichem Languagemodell (1-9,Zero,Oh)*.
Erwartungsgemäß sind die nur auf Ziffern trainierten Modelle am besten, die WSJ-Modelle schlechter und die (mit sehr viel Material trainierten) HUB4-Modelle dazwischen. Die Unterschiede können teilweise auch in unterschiedlichen Aufnahmebedingungen der jeweiligen Korpora begründet sein.
TIDIGITS-Acoustical models:
[java] # ----------------------------- Timers----------------------------------------
[java] # Name Count CurTime MinTime MaxTime AvgTime TotTime
[java] streamDataSourc 45481 0,0000s 0,0000s 0,1640s 0,0001s 2,8440s
[java] premphasizer 45481 0,0000s 0,0000s 0,0140s 0,0000s 0,2160s
[java] windower 45388 0,0000s 0,0000s 0,1200s 0,0001s 3,2960s
[java] fft 714087 0,0000s 0,0000s 0,3610s 0,0000s 35,4510s
[java] melFilterBank 714087 0,0000s 0,0000s 0,1060s 0,0000s 3,3280s
[java] dct 714087 0,0000s 0,0000s 0,1620s 0,0000s 4,2330s
[java] featureExtracti 698071 0,0000s 0,0000s 0,0470s 0,0000s 1,0670s
[java] AM_Load 1 0,5110s 0,5110s 0,5110s 0,5110s 0,5110s
[java] DictionaryLoad 1 0,0030s 0,0030s 0,0030s 0,0030s 0,0030s
[java] grammarLoad 1 0,0040s 0,0040s 0,0040s 0,0040s 0,0040s
[java] compile 1 0,2760s 0,2760s 0,2760s 0,2760s 0,2760s
[java] createGStates 1 0,0190s 0,0190s 0,0190s 0,0190s 0,0190s
[java] collectContex 1 0,0020s 0,0020s 0,0020s 0,0020s 0,0020s
[java] expandStates 1 0,2420s 0,2420s 0,2420s 0,2420s 0,2420s
[java] connectNodes 1 0,0100s 0,0100s 0,0100s 0,0100s 0,0100s
[java] scoring 710083 0,0000s 0,0000s 1,2680s 0,0003s 242,5770s
[java] pruning 706079 0,0000s 0,0000s 0,0280s 0,0000s 0,3220s
[java] growing 710083 0,0000s 0,0000s 2,2200s 0,0001s 79,0840s
[java] # --------------- Summary statistics ---------
[java] Accuracy: 96,390% Errors: 517 (Sub: 321 Ins: 42 Del: 154)
[java] Words: 13159 Matches: 12684 WER: 3,929%
[java] Sentences: 4004 Matches: 3593 SentenceAcc: 89,735%
[java] Total Time Audio: 7099,85s Proc: 280,71s Speed: 0,04 X real time
[java] Mem Total: 126,62 Mb Free: 118,33 Mb
[java] Used: This: 8,30 Mb Avg: 12,52 Mb Max: 18,26 Mb
WSJ-Acoustical models:
[java] # ----------------------------- Timers----------------------------------------
[java] # Name Count CurTime MinTime MaxTime AvgTime TotTime
[java] streamDataSourc 45481 0,0000s 0,0000s 0,5320s 0,0002s 8,1740s
[java] premphasizer 45481 0,0000s 0,0000s 0,0190s 0,0000s 0,2260s
[java] windower 45388 0,0010s 0,0000s 0,0290s 0,0001s 2,7720s
[java] fft 714087 0,0000s 0,0000s 0,2250s 0,0000s 30,7730s
[java] melFilterBank 714087 0,0000s 0,0000s 0,0360s 0,0000s 2,4740s
[java] dct 714087 0,0000s 0,0000s 0,0400s 0,0000s 3,5790s
[java] featureExtracti 698071 0,0000s 0,0000s 0,0680s 0,0000s 1,0020s
[java] AM_Load 1 4,0670s 4,0670s 4,0670s 4,0670s 4,0670s
[java] DictionaryLoad 1 0,0130s 0,0130s 0,0130s 0,0130s 0,0130s
[java] grammarLoad 1 0,0470s 0,0470s 0,0470s 0,0470s 0,0470s
[java] compile 1 0,1500s 0,1500s 0,1500s 0,1500s 0,1500s
[java] createGStates 1 0,0120s 0,0120s 0,0120s 0,0120s 0,0120s
[java] collectContex 1 0,0020s 0,0020s 0,0020s 0,0020s 0,0020s
[java] expandStates 1 0,1270s 0,1270s 0,1270s 0,1270s 0,1270s
[java] connectNodes 1 0,0070s 0,0070s 0,0070s 0,0070s 0,0070s
[java] scoring 710083 0,0000s 0,0000s 1,9580s 0,0004s 302,2580s
[java] pruning 706079 0,0000s 0,0000s 0,1360s 0,0000s 0,4400s
[java] growing 710083 0,0000s 0,0000s 1,7380s 0,0002s 137,2490s
[java] # --------------- Summary statistics ---------
[java] Accuracy: 90,356% Errors: 1628 (Sub: 1080 Ins: 359 Del: 189)
[java] Words: 13159 Matches: 11890 WER: 12,372%
[java] Sentences: 4004 Matches: 2929 SentenceAcc: 73,152%
[java] Total Time Audio: 7099,85s Proc: 399,34s Speed: 0,06 X real time
[java] Mem Total: 126,62 Mb Free: 61,85 Mb
[java] Used: This: 64,78 Mb Avg: 66,63 Mb Max: 72,37 Mb
HUB4-Acoustical models:
[java] # ----------------------------- Timers----------------------------------------
[java] # Name Count CurTime MinTime MaxTime AvgTime TotTime
[java] streamDataSourc 45481 0,0000s 0,0000s 0,1130s 0,0001s 4,2320s
[java] premphasizer 45481 0,0000s 0,0000s 0,0080s 0,0000s 0,2020s
[java] windower 45388 0,0000s 0,0000s 0,0360s 0,0001s 3,2390s
[java] fft 714087 0,0000s 0,0000s 0,4210s 0,0001s 36,5030s
[java] melFilterBank 714087 0,0000s 0,0000s 0,0410s 0,0000s 2,9190s
[java] dct 714087 0,0000s 0,0000s 0,0350s 0,0000s 3,8510s
[java] featureExtracti 698071 0,0000s 0,0000s 0,0480s 0,0000s 1,2540s
[java] AM_Load 1 4,7850s 4,7850s 4,7850s 4,7850s 4,7850s
[java] DictionaryLoad 1 0,0040s 0,0040s 0,0040s 0,0040s 0,0040s
[java] grammarLoad 1 0,0030s 0,0030s 0,0030s 0,0030s 0,0030s
[java] compile 1 0,1480s 0,1480s 0,1480s 0,1480s 0,1480s
[java] createGStates 1 0,0090s 0,0090s 0,0090s 0,0090s 0,0090s
[java] collectContex 1 0,0030s 0,0030s 0,0030s 0,0030s 0,0030s
[java] expandStates 1 0,1290s 0,1290s 0,1290s 0,1290s 0,1290s
[java] connectNodes 1 0,0060s 0,0060s 0,0060s 0,0060s 0,0060s
[java] scoring 710083 0,0000s 0,0000s 1,1940s 0,0006s 458,3170s
[java] pruning 706079 0,0000s 0,0000s 0,0010s 0,0000s 0,3020s
[java] growing 710083 0,0000s 0,0000s 0,5280s 0,0003s 210,0190s
[java] # --------------- Summary statistics ---------
[java] Accuracy: 92,226% Errors: 1056 (Sub: 527 Ins: 33 Del: 496)
[java] Words: 13159 Matches: 12136 WER: 8,025%
[java] Sentences: 4004 Matches: 3167 SentenceAcc: 79,096%
[java] Total Time Audio: 7099,85s Proc: 624,46s Speed: 0,09 X real time
[java] Mem Total: 126,62 Mb Free: 41,05 Mb
[java] Used: This: 85,58 Mb Avg: 85,57 Mb Max: 91,39 Mb
timo, 06/04/07 01:10 (GMT)
Add a new page under this one