Fig. 1. Scheme of the parameterization front-end.

Fig. 2. Usual scheme for computing the non-linearly compressed filter-bank energies for a given frame. Sometimes, a LP modeling block is inserted at the end (like in PLP).

Fig. 3. Scheme of the FF computation with the filter z−z⁻¹.

Fig. 4. Approximation of the TI digits estimated variance with the inverse square magnitude of the filter 1−0.5z⁻¹.

Fig. 5. Sine shape of the cepstral lifter corresponding to the frequency filter z−z⁻¹.

Fig. 6. Simulated modulation spectra of time-filtered feature sets for: (a) del(7) and acc(11) filters; (b) TF1(15), TF2(15) and TF1(8) filters.

Fig. 7. Isolated digit recognition rate for FF2 using: (1) the static feature set, (2) one time-filtered feature set (TF1), (3) two time-filtered feature sets (TF1 and TF2), and (4) the root non-linearity instead of the logarithm with two features sets. Also, results from the standard three-feature set MFCC front-end are included as reference.

Fig. 8. (a) 2-D modulation spectrum computed over the whole adult portion of the TI isolated digit database; (b) mismatch in the 2-D modulation spectrum between clean and noisy speech for SNR=10 dB.

Fig. 9. The quotient Variance_between/Variance_within for isolated digits.

Fig. 10. Recognition rates for clean and noisy (SNR=10 dB) digits for different time and frequency filters. The black bar corresponds to the pair of time-frequency filters TF1–FF1 and TF2–FF2.

Table 1. Percentage of connected digit recognition errors for MFCC and three different frequency-filtered parameter setsa

Table 2. Percentage of connected digit recognition errors for LPCC and two different frequency-filtered parameter sets

Table 3. Single digit recognition performance corresponding to four parameterizations: MFCC, the usual FF2 technique, the modified FF2 that discards F(Q), and FF1

Table 4. Average performances for the TI connected digit Aurora task for DSR standardization

Table 5. Recognition accuracy for the Aurora task