EURASIP Journal on Audio, Speech, and Music Processing

Table 4 Recognition results in the first pass obtained on the BNSI evaluation set

From: Context-dependent factored language models

Vocabulary size	WER [%]
	Good-Turing		Knesser-Ney
	Bigram	Trigram	Bigram	Trigram
60K	33.89	30.73	33.84	30.95
100K	31.27	28.08	31.24	28.36
200K	29.70	26.27	29.78	26.53
300K	29.22	25.64	29.28	25.87

Back to article page