A formant modification method for improved ASR of children’s speech

Differences in acoustic characteristics between children’s and adults’ speech degrade performance of automatic recognition systems when trained using are used to recognize speech. This degradation is due the mismatch training testing. One main sources difference vocal tract resonances (formant frequencies) adult child speakers. The present study aims reduce formant frequencies by modifying formants better correspond carried out warping linear prediction (LP) spectrum computed from warped LP spectra a frame-based manner with corresponding residuals synthesize whose structure closer that When testing an ASR system speech, reduces spectral improves Experiments were conducted narrowband (8 kHz) wideband (16 speakers WSJCAM0 PF_STAR databases, respectively, recognizing models proposed method gave relative improvements 24% 11% for DNN TDNN models, For technique 27% 13% respectively. was also compared two speaker adaptation methods: length normalization (VTLN) speaking rate (SRA). comparison showed best method. We combined VTLN SRA, found further reduction WER. Moreover, our experiments noisy various types additive noise signal-to-noise ratios performs well degraded