基于模型的分析：调制掩蔽对声码化语音可懂度的影响

文献信息

中文摘要

声码化语音被广泛用于模拟人工耳蜗聆听的某些方面，对正常听力听者的声码化可懂度进行建模，可为电听觉下的听觉处理提供见解。使用三种掩蔽类型（语音形噪声、纯音复合声和噪声调制纯音复合声）和四种处理条件（未处理、带通滤波、纯音声码化以及带模拟电流扩散的纯音声码化）测量句子识别。除带模拟电流扩散的声码化语音外，掩蔽类型在所有条件下均影响可懂度。将带有基于相关性的决策阶段的改进型基于语音的包络功率谱模型拟合到未处理的语音形噪声条件，然后在不重新拟合的情况下应用于其余条件。该模型捕捉了未处理、带通滤波和未模拟电流扩散的纯音声码化语音中依赖于掩蔽类型的可懂度模式。对模型内部表示的分析表明，与声码化相关的可懂度损失主要由较高率时间包络调制的去除驱动，而非由频谱分辨率降低驱动。在模拟电流扩散的情况下，模型预测了未调制掩蔽声的获益丧失，但始终高估了可懂度。这些发现支持以下观点：调制掩蔽并非模拟电听觉中的主要限制，并指出包络能量和音高相关失真应作为未来面向人工耳蜗的可懂度模型的重要目标。

英文摘要

Vocoded speech is widely used to simulate aspects of cochlear implant listening, and modeling vocoded intelligibility in normal-hearing listeners can provide insight into auditory processing under electric hearing. Sentence recognition was measured using three masker types (speech-shaped noise, tone complex, and noise-modulated tone complex) and four processing conditions (unprocessed, bandpass-filtered, tone-vocoded, and tone-vocoded with simulated current spread). Masker type affected intelligibility in all conditions except vocoded speech with simulated current spread. A modified speech-based envelope power spectrum model with a correlation-based decision stage was fitted to the unprocessed speech-shaped-noise condition and then applied, without re-fitting, to the remaining conditions. The model captured masker-dependent intelligibility patterns for unprocessed, bandpass-filtered, and tone-vocoded speech without simulated current spread. Analyses of the model's internal representations suggested that vocoding-related intelligibility loss was driven mainly by the removal of higher-rate temporal-envelope modulations rather than by reduced spectral resolution. With simulated current spread, the model predicted the loss of benefit for unmodulated maskers but consistently overestimated intelligibility. These findings support the view that modulation masking is not the primary limitation in simulated electric hearing and point to envelope-energy and pitch-related distortions as important targets for future cochlear implant-oriented intelligibility models.