Only about 20 percent of time-frequency units carry audible target speech

Evidence

The ITFS studies found a decrease in target speech intelligibility for SNR criteria exceeding the global SNR of the mixture by 5-10 dB, similar to the local SNR value of about 0 dB (i.e. around 10 dB above global SNR) in our study. An analysis of the local SNRs of all T-F units in ten unfiltered stimuli (see Fig. 3 ) was performed to determine the proportion of T-F units exceeding a certain SNR level. The analysis reveals that on average only about 20 % of the total target T-F units exceeded a local SNR value of -4 dB and about 7 % exceeded a value of 10 dB SNR. — Dijk, Pim van, Physiology, Psychoacoustics and…, 1661

Categorization

• CANDY: “Only 20% of a speech signal’s T-F units carry intelligible content” is a delightful bar-fact about how sparse speech actually is.
• XREF: Connects to gated-noise and sparse-coding literatures showing the redundancy of natural signals, plus parallels with how the visual system reconstructs from sparse features.