Timestamp-aligning and keyword-biasing end-to-end ASR front-end for a KWS system

EURASIP Journal on Audio, Speech, and Music Processing

Table 10 A comparison of using different bias factors and SAD dilation-erosion factors on ASR Model 1. ATWV values are evaluated on the SAD segmented dev dataset. “T”, “G” and “G3” represent CNN-TDNNF aligner, GMM-HMM (4 iteration) aligner and GMM-HMM (3 iteration) aligner, respectively

bias factor	SAD α 0.1	SAD α 0.2	SAD α 0.3	SAD α 0.4	SAD α 0.5
δ	T / G / G3	T / G / G3	T / G / G3	T / G / G3	T / G / G3
	0.5998	0.6092	0.6056	0.6101	0.6004
1/4	0.5919	0.6077	0.6028	0.6086	0.5896
	0.5863	0.5967	0.5904	0.5915	0.5656
	0.6061	0.6043	0.6098	0.6069	0.6013
1/3	0.5919	0.6050	0.6086	0.6058	0.5951
	0.5688	0.5966	0.5984	0.5906	0.5713
	0.6056	0.6048	0.6124	0.6108	0.5999
1/2	0.5955	0.6031	0.6092	0.6072	0.5910
	0.5891	0.5895	0.5993	0.5919	0.5674
	0.6016	0.5998	0.6145	0.6204	0.5988
1	0.5950	0.6007	0.6077	0.6140	0.5972
	0.5897	0.5901	0.5989	0.6004	0.5686
	0.6022	0.6120	0.6206	0.6207	0.5889
2	0.5920	0.6107	0.6102	0.6128	0.5886
	0.5864	0.5794	0.6006	0.6023	0.5649
	0.5898	0.6087	0.6004	0.6019	0.5870
4	0.5788	0.6024	0.6012	0.5993	0.5865
	0.5773	0.5976	0.5955	0.5848	0.5667