Binary time-frequency (TF) masks can be applied to separate speech from noise. Previous studies have shown that with appropriate parameters, ideal TF masks can extract highly intelligible speech even at very low speech-to-noise ratios (SNRs). Two psychophysical experiments provided additional information about the dependence of intelligibility on the frequency resolution and threshold criteria that define the ideal TF mask. Listeners identified AzBio Sentences in noise, before and after application of TF masks. Masks generated with 8 or 16 frequency bands per octave supported nearly-perfect identification. Word recognition accuracy was slightly lower and more variable with 4 bands per octave. When TF masks were generated with a local threshold criterion of 0 dB SNR, the mean speech reception threshold was −9.5 dB SNR, compared to −5.7 dB for unprocessed sentences in noise. Speech reception thresholds decreased by about 1 dB per dB of additional decrease in the local threshold criterion. Information reported here about the dependence of speech intelligibility on frequency and level parameters has relevance for the development of non-ideal TF masks for clinical applications such as speech processing for hearing aids.
展开▼
机译：可以将二进制时频（TF）掩码用于将语音与噪声分离。先前的研究表明，有了合适的参数，理想的TF掩码即使在非常低的信噪比（SNR）时也可以提取出高度可理解的语音。两次心理物理实验提供了有关清晰度对定义理想TF掩模的频率分辨率和阈值标准的依赖性的其他信息。听众在使用TF蒙版之前和之后都发现了噪音中的AzBio句子。每个八度产生8或16个频带的掩码支持近乎完美的识别。字识别准确度略低，每个八度音阶有4个带，变化更大。当使用本地阈值标准SNR为0 dB生成TF掩码时，平均语音接收阈值为-9.5 dB SNR，而对于噪声中未经处理的句子，则为-5.7 dB。语音接收阈值每降低本地阈值标准中的dB，大约降低1dB。此处报告的有关语音清晰度对频率和级别参数的依赖性的信息与开发用于临床应用（如助听器的语音处理）的非理想TF蒙版具有相关性。
展开▼
