Table 8.

The CER (%) and WER (%) results of different E2E ASR models were trained using the Uzbek language speech corpus. The impact of language model (LM), speed perturbation (SP), and spectral augmentation (SA) are also reported.

Model	LM	SP	SA	Valid		Test
				CER	WER	CER	WER
E2E-LSTM	×	×	×	13.8	43.1	14.0	44.0
	√	×	×	14.9	30.0	14.3	31.4
	√	√	×	13.7	27.6	14.4	30.6
	√	√	√	12.6	24.9	12.0	27.0
DNN-HMM	×	×	×	12.8	34.7	10.2	32.1
	√	×	×	10.3	20.5	8.6	24.9
	√	√	×	6.9	18.8	7.5	23.5
	√	√	√	6.9	19.9	8.1	24.9
RNN-CTC	×	×	×	13.3	35.8	9.7	32.3
	√	×	×	12.2	27.2	9.1	24.3
	√	√	×	10.9	25.1	8.7	23.9
	√	√	√	8.3	24.7	7.9	22.3
E2E − Transformer	×	×	×	12.3	35.2	9.4	31.6
	√	×	×	11.7	25.7	8.7	23.9
	√	√	×	10.7	23.9	8.4	23.0
	√	√	√	9.9	21.4	7.6	21.0
E2E-Conformer	×	×	×	12.7	37.6	10.7	35.1
	√	×	×	11.5	27.5	9.7	26.3
	√	√	×	9.2	21.7	7.5	21.2
	√	√	√	7.8	18.1	5.8	17.4
E2E − T (CTC + Attention)	×	×	×	12.1	33.2	9.8	30.3
	√	×	×	9.6	19.4	7.9	22.7
	√	√	×	6.4	17.9	7.4	20.3
	√	√	√	5.7	15.2	5.41	14.3