58:
level. The WER is a valuable tool for comparing different systems as well as for evaluating improvements within one system. This kind of measurement, however, provides no details on the nature of translation errors and further work is therefore required to identify the main source(s) of error and to
444:
Whichever metric is used, however, one major theoretical problem in assessing the performance of a system is deciding whether a word has been “mis-pronounced,” i.e. does the fault lie with the user or with the recogniser. This may be particularly relevant in a system which is designed to cope with
440:
There is some debate, however, as to whether Hunt's formula may properly be used to assess the performance of a single system, as it was developed as a means of comparing more fairly competing candidate systems. A further complication is added by whether a given syntax allows for error correction
362:
One problem with using a generic formula such as the one above, however, is that no account is taken of the effect that different types of error may have on the likelihood of successful outcome, e.g. some errors may be more disruptive than others and some may be corrected more easily than others.
349:
experiment, it was shown that, if people were trained under "that matches the optimization objective for understanding", (Wang, Acero and Chelba, 2003) they would show a higher accuracy in understanding of language than other people who demonstrated a lower word error rate, showing that true
452:
For text dictation it is generally agreed that performance accuracy at a rate below 95% is not acceptable, but this again may be syntax and/or domain specific, e.g. whether there is time pressure on users to complete the task, whether there are alternative methods of completion, and so on.
45:
system. The WER metric ranges from 0 to 1, where 0 indicates that the compared pieces of text are exactly identical, and 1 indicates that they are completely different with no similarity. This way, a WER of 0.8 means that there is an 80% error rate for compared sentences.
62:
This problem is solved by first aligning the recognized word sequence with the reference (spoken) word sequence using dynamic string alignment. Examination of this issue is seen through a theory called the power law that states the correlation between
472:( X, Y ) is defined as the minimum of W( P ) / L ( P ), where P is an editing path between X and Y, W ( P ) is the sum of the weights of the elementary edit operations of P, and L(P) is the number of these operations (length of P).
448:
The pace at which words should be spoken during the measurement process is also a source of variability between subjects, as is the need for subjects to rest or take a breath. All such factors may need to be controlled in some way.
327:
49:
The general difficulty of measuring performance lies in the fact that the recognized word sequence can have a different length from the reference word sequence (supposedly the correct one). The WER is derived from the
169:
344:
It is commonly believed that a lower word error rate shows superior accuracy in recognition of speech, compared with a higher word error rate. However, at least one study has shown that this may not be true. In a
209:
The intuition behind 'deletion' and 'insertion' is how to get from the reference to the hypothesis. So if we have the reference "This is wikipedia" and hypothesis "This _ wikipedia", we call it a deletion.
370:
Hunt (1990) has proposed the use of a weighted measure of performance accuracy where errors of substitution are weighted at unity but errors of deletion and insertion are both weighted only at 0.5, thus:
435:
441:
and, if it does, how easy that process is for the user. There is thus some merit to the argument that performance metrics should be developed to suit the particular system being measured.
367:
being tested. A further problem is that, even with the best alignment, the formula cannot distinguish a substitution error from a combined deletion plus insertion error.
223:
659:
456:
The term "Single Word Error Rate" is sometimes referred to as the percentage of incorrect recognitions for each different word in the system vocabulary.
76:
593:
336:
is the number of words in the reference, the word error rate can be larger than 1.0, and thus, the word accuracy can be smaller than 0.0.
613:
377:
654:
649:
525:
Klakow, Dietrich; Jochen Peters (September 2002). "Testing the correlation of word error rate and perplexity".
565:
51:
622:
42:
570:
564:. IEEE Workshop on Automatic Speech Recognition and Understanding. St. Thomas, US Virgin Islands.
346:
38:
542:
534:
351:
617:
594:
Computation of
Normalized Edit Distance and Application:AndrCs Marzal and Enrique Vidal
322:{\displaystyle {\mathit {WAcc}}=1-{\mathit {WER}}={\frac {N-S-D-I}{N}}={\frac {C-I}{N}}}
664:
501:
538:
17:
643:
631:
496:
465:
607:
610:
On the Use of
Information Retrieval Measures for Speech Recognition Evaluation
562:
Is Word Error Rate a Good
Indicator for Spoken Language Understanding Accuracy
64:
546:
486:
445:
non-native speakers of a given language or with strong regional accents.
213:
When reporting the performance of a speech recognition system, sometimes
164:{\displaystyle {\mathit {WER}}={\frac {S+D+I}{N}}={\frac {S+D+I}{S+D+C}}}
55:
491:
464:
The word error rate may also be referred to as the length normalized
364:
634:
Minimizing Word Error Rate in
Textual Summaries of Spoken Language
481:
350:
understanding of spoken language relies on more than just high
389:
386:
383:
260:
257:
254:
238:
235:
232:
229:
88:
85:
82:
625:
Figures of Merit for
Assessing Connected Word Recognisers
430:{\displaystyle {\mathit {WER}}={\frac {S+0.5D+0.5I}{N}}}
380:
226:
79:
429:
321:
163:
205:is the number of words in the reference (N=S+D+C)
468:. The normalized edit distance between X and Y,
363:These factors are likely to be specific to the
37:) is a common metric of the performance of a
8:
627:(Speech Communication, 9, 1990, pp 239-336)
54:, working at the word level instead of the
569:
397:
382:
381:
379:
301:
268:
253:
252:
228:
227:
225:
123:
96:
81:
80:
78:
70:Word error rate can then be computed as:
560:Wang, Y.; Acero, A.; Chelba, C. (2003).
517:
7:
27:Computer language processing metric
25:
660:Evaluation of machine translation
199:is the number of correct words,
181:is the number of substitutions,
1:
539:10.1016/S0167-6393(01)00041-3
193:is the number of insertions,
187:is the number of deletions,
59:focus any research effort.
681:
632:Zechner, K., Waibel, A.
431:
323:
165:
18:Single Word Error Rate
608:McCowan et al. 2005:
432:
324:
166:
67:and word error rate.
527:Speech Communication
378:
224:
215:word accuracy (WAcc)
77:
52:Levenshtein distance
655:Machine translation
584:Nießen et al.(2000)
43:machine translation
650:Speech recognition
623:Hunt, M.J., 1990:
616:2019-02-24 at the
427:
347:Microsoft Research
319:
161:
39:speech recognition
425:
317:
296:
217:is used instead:
159:
118:
16:(Redirected from
672:
596:
591:
585:
582:
576:
575:
573:
557:
551:
550:
522:
436:
434:
433:
428:
426:
421:
398:
393:
392:
352:word recognition
332:Note that since
328:
326:
325:
320:
318:
313:
302:
297:
292:
269:
264:
263:
242:
241:
170:
168:
167:
162:
160:
158:
141:
124:
119:
114:
97:
92:
91:
21:
680:
679:
675:
674:
673:
671:
670:
669:
640:
639:
618:Wayback Machine
604:
599:
592:
588:
583:
579:
559:
558:
554:
524:
523:
519:
515:
510:
478:
462:
399:
376:
375:
360:
342:
303:
270:
222:
221:
142:
125:
98:
75:
74:
31:Word error rate
28:
23:
22:
15:
12:
11:
5:
678:
676:
668:
667:
662:
657:
652:
642:
641:
638:
637:
629:
620:
603:
600:
598:
597:
586:
577:
552:
533:(1–2): 19–28.
516:
514:
511:
509:
506:
505:
504:
502:ROUGE (metric)
499:
494:
489:
484:
477:
474:
461:
458:
438:
437:
424:
420:
417:
414:
411:
408:
405:
402:
396:
391:
388:
385:
359:
356:
341:
338:
330:
329:
316:
312:
309:
306:
300:
295:
291:
288:
285:
282:
279:
276:
273:
267:
262:
259:
256:
251:
248:
245:
240:
237:
234:
231:
207:
206:
200:
194:
188:
182:
172:
171:
157:
154:
151:
148:
145:
140:
137:
134:
131:
128:
122:
117:
113:
110:
107:
104:
101:
95:
90:
87:
84:
26:
24:
14:
13:
10:
9:
6:
4:
3:
2:
677:
666:
663:
661:
658:
656:
653:
651:
648:
647:
645:
636:
635:
630:
628:
626:
621:
619:
615:
612:
611:
606:
605:
602:Other sources
601:
595:
590:
587:
581:
578:
572:
571:10.1.1.89.424
567:
563:
556:
553:
548:
544:
540:
536:
532:
528:
521:
518:
512:
507:
503:
500:
498:
497:NIST (metric)
495:
493:
490:
488:
485:
483:
480:
479:
475:
473:
471:
467:
466:edit distance
460:Edit distance
459:
457:
454:
450:
446:
442:
422:
418:
415:
412:
409:
406:
403:
400:
394:
374:
373:
372:
368:
366:
358:Other metrics
357:
355:
353:
348:
339:
337:
335:
314:
310:
307:
304:
298:
293:
289:
286:
283:
280:
277:
274:
271:
265:
249:
246:
243:
220:
219:
218:
216:
211:
204:
201:
198:
195:
192:
189:
186:
183:
180:
177:
176:
175:
155:
152:
149:
146:
143:
138:
135:
132:
129:
126:
120:
115:
111:
108:
105:
102:
99:
93:
73:
72:
71:
68:
66:
60:
57:
53:
47:
44:
40:
36:
32:
19:
633:
624:
609:
589:
580:
561:
555:
530:
526:
520:
469:
463:
455:
451:
447:
443:
439:
369:
361:
343:
333:
331:
214:
212:
208:
202:
196:
190:
184:
178:
173:
69:
61:
48:
34:
30:
29:
340:Experiments
644:Categories
508:References
354:accuracy.
65:perplexity
566:CiteSeerX
547:0167-6393
487:F-Measure
308:−
287:−
281:−
275:−
250:−
614:Archived
476:See also
56:phoneme
568:
545:
492:METEOR
365:syntax
174:where
665:Rates
513:Notes
543:ISSN
482:BLEU
535:doi
416:0.5
407:0.5
41:or
35:WER
646::
541:.
531:38
529:.
574:.
549:.
537::
470:d
423:N
419:I
413:+
410:D
404:+
401:S
395:=
390:R
387:E
384:W
334:N
315:N
311:I
305:C
299:=
294:N
290:I
284:D
278:S
272:N
266:=
261:R
258:E
255:W
247:1
244:=
239:c
236:c
233:A
230:W
203:N
197:C
191:I
185:D
179:S
156:C
153:+
150:D
147:+
144:S
139:I
136:+
133:D
130:+
127:S
121:=
116:N
112:I
109:+
106:D
103:+
100:S
94:=
89:R
86:E
83:W
33:(
20:)
Text is available under the Creative Commons Attribution-ShareAlike License. Additional terms may apply.