Out-Of-Vocabulary Word Error Rate (OOV-WER) Measure

An Out-Of-Vocabulary Word Error Rate (OOV-WER) Measure is a Word Error Rate Measure that measures the performance of OOV detection system.

• Example(s):
  • …
• Counter-Example(s):
  • Position Independent Word Error Rate (PER),
  • NIST Score.
  • Word Accuracy (WAcc) Measure,
  • Phonetic Levenshtein Distance.
  • ASR Latency.
  • BLEU Score.
• See: OOV Detection System, Recall, Speech Recognition, Machine Translation, Levenshtein Distance, Phoneme.

---

---

References

2006

• (Schultz & Kirchhoff, 2006) ⇒ Tanja Schultz, and Katrin Kirchhoff (Eds.). (2006). "Multilingual Speech Processing". Elsevier.
  • QUOTE: This is due to their morphologic structure and leads to rapid vocabulary growth and a high rate of Out-of-Vocabulary (OOV) words. OOV words are those that a speech processing system is expected 10 handle but have ever been observed in training data (e.g., unseen inflections of words that are important for the application). Words that do not occur in training data are usually not included in the search vocabulary of the speech recognition system. Thus, they cannot be recognized by the system. Consequently, each OOV word leads 10 at least one recognition error. As a word error can reduce the effective reach of the language model, we usually observe 1.2-1.5 word errors per OOV word.

    Table 4.2 gives the size of vocabulary and resulting OOV rates for 10 GlobalPhone languages. The OOV rates differ significantly between these languages (...)

Table 4.2: Out-of-vocabulary rates for ten languages.

Language	Vocabulary	OOV Rate
Korean	64K	34.0%
Turkish	64K	13.5%
German	61K	4.4%
Portuguese	60K	4.3%
English	54K	0.3%
Korean (segmented)	64K	0.2%
Chinese (sepmented)	60K	0%
Croatian	31K	13.6%
Spanish	30K	5.2%
French	30K	4.7%
Japanese (segmented)	22K	3.0%

---