pk-Ctone的开发与评估：一种用于增强人工耳蜗普通话声调感知的计算高效策略

文献信息

中文摘要

背景： 传统人工耳蜗（CI）策略传递的声调信息有限。尽管基于基频（F0）的策略（如将F0提取与幅度包络调制相结合的C-tone）可改善普通话声调感知，但其高计算需求阻碍了其集成到功率和资源受限的设备中。本研究评估了一种词汇声调增强CI策略pk-Ctone，该策略采用基于局部极值的计算高效F0检测算法。方法：在两个普通话单音节语料库上进行声学评估，以评估F0检测的粗尺度稳健性和细尺度保真度。量化了F0提取所需的计算操作，并开展了一项有62名参与者参加的临床试验，比较使用pk-Ctone与标准包络基线的普通话语音识别。结果：结果显示，与C-tone相比，pk-Ctone将F0提取的计算操作减少了30倍以上。尽管这种效率提升在细尺度精度上有所损失，但pk-Ctone实现了更好的粗尺度稳健性，以准确率和F1分数衡量。在临床上，pk-Ctone在单音节声调识别（+10.1个百分点，pp）、单音节词（+19.2 pp）和双音节词（+13.0 pp）方面均产生了显著改善。结论：这些发现表明，pk-Ctone通过保持粗尺度F0跟踪稳健性同时降低计算成本，尽管在细尺度精度上有所损失，仍保持了临床有效性，为功率和资源受限的CI处理器提供了一种实用解决方案。

英文摘要

Conventional cochlear implant (CI) strategies convey limited tonal information. Although fundamental frequency (F0)-based strategies, such as C-tone, which combines F0 extraction with amplitude envelope modulation, improve Mandarin tone perception, their high computational demands hinder integration into power- and resource-constrained devices. This study evaluates a lexical tone enhancement CI strategy, pk-Ctone, that employs a computationally efficient local extrema-based F0 detection algorithm. Acoustic evaluations on two Mandarin monosyllabic corpora assessed both the coarse-scale robustness and fine-scale fidelity of F0 detection. Computational operations required for F0 extraction were quantified, and a clinical trial with 62 participants compared Mandarin speech recognition using pk-Ctone versus a standard envelope-based baseline. Results showed that pk-Ctone reduced the computational operations for F0 extraction by more than 30-fold compared with C-tone. Although this efficiency gain incurred a penalty in fine-scale precision, pk-Ctone achieved improved coarse-scale robustness, as quantified by Accuracy and F1-score. Clinically, pk-Ctone yielded significant improvements in recognition of monosyllabic tones (+10.1% points, pp), monosyllabic words (+19.2 pp), and disyllabic words (+13.0 pp). These findings indicate that pk-Ctone, by preserving coarse-scale F0 tracking robustness while reducing computational cost, maintains clinical effectiveness despite some loss in fine-scale precision, offering a practical solution for power- and resource-constrained CI processors.