Evaluating the intelligibility benefit of speech modifications in known noise conditions

Improving intelligibility in noise of HMM-generated speech via noise-dependent and -independent methods

“…The differences among the methods is attenuated in CS, that is the gains obtained by the noise-independent method TTS-SS-DRC were attenuated. These results can be converted as equivalent intensity changes (EIC) relative to normal natural speech in a dB scale as proposed in [17]. Calculating this changes with respect to natural speech we found that TTS-SS-DRC was 2.0 dB higher than natural speech in SSN Low SNR, this gain is lower for the Mid SNR condition: 0.7 dB.…”

“…As the speech material and the noise conditions were the same as the ones used in [17] we can directly compare SS-DRC results (OptSII noise estimation here is different than the system used in [17]). SS-DRC applied to a TTS voice improves TTS performance by 4.5 dB for SSN and 2.1 dB for CS (averaged across all SNRs) compared to performance gains of 4.2 dB and 3.1 dB for natural speech.…”

“…A recursive closed-form optimisation scheme calculates, for each time frame, the spectral weights in 21 Bark-scaled subbands which maximise the SII, given the current disturbance spectrum levels, with the additional constraint of an unchanged short-term audio power of the speech signal. Opposed to [10] and the OptSII style used in [17], in this evaluation a moving average noise estimator is used, which is also able to track CS noise.…”

“…We added the seven synthetic styles shown in Table 1 and the natural speech to speech-shaped noise (SSN) and a female competing speaker (CS) at 3 different signal to noise ratios ('Low SNR', 'Mid SNR', and 'High SNR'), estimated in pilot tests to be −9, −4 and +1 dB for SSN, and −21, −14 and −7 dB for CS as used in [17]. In order to obtain listening scores for word accuracy, we performed a listening test with 88 native English speakers.…”

“…However, noise-dependent methods, either for natural or TTS voices, have only recently been proposed and it remains relatively unknown to what extend exploiting spectro-temporal characteristics of the masker is useful. To evaluate a range of enhancement algorithms, both noise-dependent and independent, [17] describes a large scale listening experiment with 5 methods for natural speech and 2 for TTS evaluated under the same conditions (speech and noise material). In this evaluation, it was observed that noise-independent spectral shaping with Dynamic Range Compression (SSDRC) [6] provided the best results of the modifications on natural speech while [10], although noise-dependent, did not perform as well.…”