Subjective quality testing of a new source model of LPO vocoders

Our source model, reported at an earlier meeting, excites the LPO speech spectrum with a low‐frequency band of pulses mixed with a high‐frequency band of noise. Pulses are low‐pass filtered and noise is high‐pass filtered at the same frequency, to yield a flat source spectrum. The cutoff frequency of the filters, a continuous variable, replaces the usual binary voiced/voiceless decision. Thirty‐six phoneme‐specific test sentences were processed through a single high‐quality vocoder (5 kHz bandwidth, 11 poles, no quantization, 100 frames/s), which was excited in turn by both the usual pulse/noise source and by the new source. Subjects rated the resulting speech separately on eight‐point buzziness and breathiness scales. The results show that the new source model greatly reduces perceived buzziness, occasionally at a cost of slightly increased breathiness. Any remaining inadequacies can probably be ascribed to the algorithm that extracts the cutoff frequency during analysis, rather than to the model itself....