Transcranial magnetic stimulation (TMS) during confrontational naming can induce anomia or speech arrest and paraphasias and those can be used to locate language areas. This method is qualitative and time-consuming since a trained professional needs to analyze a long video-file of the mapping session. For data analysis we adapted an existing speech analysis method [1] to cope with TMS noise, which exceeds 100 dB SPL with each discharge of the coil. Our algorithm deciphered what was said and the reaction time (RT) of each utterance with a resolution of up to 10ms, leading to a fast and quantitative analysis. To develop the noise reduction in combination with the speech analysis algorithms, semi-synthetic data were created, since it is impossible for the human ear to discern speech onset in the loud TMS-noises. During a confrontational naming task (without TMS), voice-recordings were obtained and RTs were manually annotated. Separately, multiple realistic recordings of rTMS noise were recorded, using a figure8 coil. Semi-synthetic data were then created by fusion of the rTMS-noises (at the onset) and the voicerecordings with some random delay. Over 1000 recordings were created. Actual patient data were recorded from three male patients planned for neurosurgery near language areas. Setup was in accordancewith published guidelines [2] with two no-TMS recordings as baseline and TMS-mapping of language areas using a 5Hz-1s stimulation at an intensity of 120% of resting motor threshold. Stimulation targets were recorded (Brainsight, Rogue Research, Canada) for offline analysis. All three had direct cortical stimulation (DCS) during awake surgery which we used as gold-standard. The study was approved by the local Ethics Committee of the University Hospitals Leuven. All patients gave written informed consent. The algorithm contrasted the data likelihood for the model of the expected response with the data likelihood for a generic model of words. An internal penalty parameter avoided the generic model from winning inappropriately. In this task, we adjusted the weight to recognize the correct response in order to maximize the true negatives correctly (but reducing true positives). Speech enhancement front-end suppressed background noises and rTMS noise [3,4]: recordings were decomposed in a weighted sum of thousands of spectro-temporal exemplars of speech and noise and the noise components were suppressed. The speech recognizer yielded temporal alignments with differing offsets depending on the beginning phoneme of the response: these biases were compensated in the post-processing stage. The algorithm could handle synonyms and new pictures, since the word models were built by joining phoneme models automatically. As output a

中文翻译：

---

自动语音分析以改进基于 TMS 的语言映射：算法和概念证明

对抗性命名过程中的经颅磁刺激 (TMS) 可引起失语或言语停滞和失语症，这些可用于定位语言区域。这种方法是定性和耗时的，因为受过训练的专业人员需要分析映射会话的长视频文件。对于数据分析，我们采用了现有的语音分析方法 [1] 来处理 TMS 噪声，该噪声在线圈每次放电时超过 100 dB SPL。我们的算法以高达 10 毫秒的分辨率破译了所说的内容和每个话语的反应时间 (RT)，从而进行快速和定量的分析。为了结合语音分析算法开发降噪，创建了半合成数据，因为人耳不可能在嘈杂的 TMS 噪声中辨别语音开始。在对抗性命名任务（没有 TMS）期间，获得了录音并手动注释了 RT。另外，使用 8 字形线圈记录了 rTMS 噪声的多个真实录音。然后通过融合 rTMS 噪声（在开始时）和具有一些随机延迟的录音来创建半合成数据。创建了 1000 多个录音。从计划在语言区附近进行神经外科手术的三名男性患者中记录了实际患者数据。设置符合已发布的指南 [2]，以两个无 TMS 录音作为基线和 TMS 映射语言区域，使用 5Hz-1s 刺激，强度为静息运动阈值的 120%。记录刺激目标（Brainsight，Rogue Research，加拿大）用于离线分析。所有三个人在清醒手术期间都进行了直接皮质刺激（DCS），我们将其用作金标准。该研究得到了鲁汶大学医院当地伦理委员会的批准。所有患者均给予书面知情同意。该算法将预期响应模型的数据似然与单词通用模型的数据似然进行了对比。内部惩罚参数避免了通用模型不恰当地获胜。在这个任务中，我们调整了权重以识别正确的响应，以正确地最大化真负例（但减少真正例）。语音增强前端抑制背景噪声和 rTMS 噪声 [3,4]：录音被分解为数千个语音和噪声的频谱时间样本的加权和，并且噪声分量被抑制。语音识别器根据响应的开始音素产生具有不同偏移的时间对齐：这些偏差在后处理阶段得到补偿。该算法可以处理同义词和新图片，因为单词模型是通过自动加入音素模型构建的。作为输出