Mirkovic et al., 2016
target speaker detection
为什么重要
- 路线定位:around-the-ear cEEGrid / target speaker detection。
- 任务或证据:Two-speaker envelope tracking
- 自研用途:Honest AAD baseline and cap/ear layout comparison.
Evidence Matrix Summary
| Field | Value |
|---|---|
| Route / hardware | Bilateral cEEGrid + cap comparison |
| Task / evidence base | Two-speaker envelope tracking |
| Main finding | cEEGrid decoded attended speaker at 69.33% vs cap about 84.8%; spatial placement drove gap. |
| Key limitation | 30-60 s windows; cEEGrid below cap. |
| Use for our system | Honest AAD baseline and cap/ear layout comparison. |
PDF Download
下载 PDF静态路径:/papers/10-mirkovic-2016.pdf
Detailed Reading Card
基本信息
- 年份/出处: 2016, Frontiers in Neuroscience, 10:349.
- DOI: 10.3389/fnins.2016.00349.
- 路线: around-the-ear cEEGrid。
- 研究类型: cEEGrid target speaker decoding / hearing-aid BCI 应用验证。
研究问题
- 目标是测试 concealed around-the-ear cEEGrid 是否能在双说话人竞争场景中用 speech envelope tracking 识别 attended speaker(Abstract; Introduction; PDF pp. 1-2)。
- 应用动机: 如果 EEG 系统能缩小且保持质量,可用于助听器中识别目标说话人并指导 speech enhancement(Abstract; Introduction; PDF p. 1)。
硬件系统
- cEEGrid: 双耳各一片,10 个 flex-printed Ag/AgCl electrodes;R4a ground,R4b reference;L4b offline linked mastoid re-reference,L4a 排除以保持对称,最终 16 channels(Fig. 1; EEG Recordings; PDF p. 3)。
- 放大器: SMARTING wireless mobile 24-channel DC EEG amplifier,24-bit,500 Hz,Bluetooth,置于后脑(EEG Recordings; PDF p. 3)。
- 对照: 96-channel cap 中因 cEEGrid 去掉耳周 12 个,保留 84 channels;BrainAmp 16-bit,nose-tip reference,5000 Hz,0.0153-250 Hz analog filter,LSL downsample 到 500 Hz(EEG Recordings; PDF p. 3)。
- 阻抗: cap 和 ear-EEG 实验前后测量;cEEGrid impedance 0.6-83.6 kOhm,median 13.5 kOhm(Influence of cEEGrid properties; PDF p. 7)。
电极点位 / 布局
- cEEGrid 围绕左右耳贴附,电极 L1-L8/R1-R8 作为分析通道,L4a 排除,L4b/R4b linked mastoids 重参考(Fig. 1; Preprocessing; PDF pp. 3-4)。
- 与 cap 3D 位置同步 digitize,用于位置/角度分析(Methods; PDF pp. 5-6)。
实验设计
- 被试: 20 healthy normal-hearing native German participants,mean age 24.8,8 male,1 left-handed;1 人技术问题排除,1 人任务表现差排除,最终分析 N=18(Participants; Results; PDF pp. 2, 6)。
- 任务: 两个 German fairy tale speech streams 同时播放,5 blocks x 10 min;被试听一个指定 stream,全实验 attended location 固定(Paradigm; PDF p. 2)。
- 刺激: 两名 male professional speakers;silent gaps >500 ms 压缩;用 RMS weighting 平衡强度;HRTF 形成 -45/45 deg azimuth,0.8 m;E-A-RTONE 3A earphones(Stimuli; PDF pp. 2-3)。
- 行为检查: 每个 break 后回答每个 speech stream 10 个 multiple-choice questions;最终被试 attended story 平均 86.1%,unattended <0.1%(Paradigm/Results; PDF pp. 2, 6)。
信号处理流程
- EEG preprocessing: EEGLAB 13.4.4b;cap baseline correction 后 common average reference,2-8 Hz Hann FIR (LP order 100, HP order 500),downsample 64 Hz,切成 50 个 60 s trials(Preprocessing; PDF p. 4)。
- ear-EEG: baseline correction,re-reference 到 linked mastoids L4b/R4b,L4a 移除;同样 2-8 Hz、64 Hz、50 x 60 s trials(Preprocessing; PDF p. 4)。
- speech envelopes: Hilbert transform absolute value,low-pass <8 Hz,downsample 64 Hz(Preprocessing; PDF p. 4)。
- decoding: multivariate linear regression / envelope tracking;time-lag interval 用 -115 到 620 ms 内 45 ms window, 30 ms overlap 搜索;最佳 lag 约 140-200 ms(Optimal Time Lag; PDF pp. 5-6)。
- validation: leave-one-out across 50 trials;训练 Nt-1 decoders,估计剩余 trial 的 attended speech envelope (EASE),比较与 attended/unattended envelope 的 correlation,较高者为分类结果(Envelope Tracking; PDF pp. 4-5)。
结果
- cap 验证: 当前 84-channel cap 解码 84.78%;重分析既往数据 88.02%,无显著差异;96-channel 既往数据为 89.23%(Scalp verification; PDF p. 6)。
- cEEGrid 解码: 双 cEEGrid 16 channels 准确率 69.33%,2/18 低于 chance;单侧 8 channels 为 64.88%,5 人在或低于 chance;cap-EEG abstract 报告 84.8%(Abstract; Fig. 6; PDF pp. 1, 7)。
- cap 与 ear performance score 正相关 r=0.57, p=0.012,但 ear score 显著低于 cap(Overall performance; Fig. 5; PDF p. 6)。
- electrode layout: 16 个 wide scalp channels 可达 85.67%,不显著低于 84-channel;temporal cap 16 channels 为 82.78%;posterior 16 channels 最低 63.78%(Influence of Electrode Position; Fig. 6; PDF p. 7)。
- impedance/placement: impedance 与 good/bad performers 无系统差异;upper vs lower cEEGrid channel impedance 无显著差异;cEEGrid placement angles 与表现无关(Fig. 7; PDF p. 7)。
局限
- cEEGrid 性能低于 cap;作者认为主要来自 electrode placement/spatial location,而不是电极尺寸、阻抗或放大器差异(Discussion; PDF pp. 9-10)。
- speech decoding 需要约 30-60 s 数据才能可靠决策,信息传输率低,注意切换捕捉慢(Discussion; PDF p. 10)。
- cEEGrid 非最佳 scalp 位置,未来需要 cEEGrid-dedicated signal processing 和更快动态算法(Discussion/Conclusion; PDF pp. 10-11)。
- 可能存在 blink、lateral eye movement、jaw/head EMG 等伪迹,但高频功率分析未支持这些是性能差异主因(Discussion; PDF p. 10)。
对自研的启发
- hearing-aid steering 是 cEEGrid 的自然应用场景,但要诚实报告相对 cap 的性能差距。
- envelope tracking baseline 必须包含 cap 对照、不同 electrode layout 对照、单耳/双耳对照和 impedance/位置分析。
- 真实产品化瓶颈不是“能否 above chance”,而是决策窗口太长、注意切换慢和佩戴位置不是最优。
下一批建议
下一批建议处理 11 Goverdovsky 2016、13 Pacharra 2017、14 Sterr 2018、15 Denk 2018、16 Mikkelsen 2019。这批会把路线扩展到 generic in-ear 24/7、visual cognition、sleep PSG、hearing-device ERP 和 sleep/wake machine learning。
Metadata
| Field | Value |
|---|---|
| ID | p10_mirkovic_2016_target_speaker |
| Title | Target Speaker Detection with Concealed EEG Around the Ear |
| Year | 2016 |
| Category | 03_ceegrid_task_validation |
| Route | around-the-ear cEEGrid |
| Stage | target speaker detection |
| Status | processed |
| Source integrity | ok |
| Pages | 11 |
| OCR status | not_needed |
Evidence Groups
| Group | Hits | Pages |
|---|---|---|
| hardware | 12 | p. 1, p. 2, p. 3 |
| electrode_layout | 12 | p. 1, p. 2, p. 3, p. 4 |
| experiment | 12 | p. 1, p. 2 |
| signal_processing | 12 | p. 1, p. 2, p. 3, p. 4 |
| results | 12 | p. 1, p. 2, p. 4, p. 5, p. 6 |
| limitations | 12 | p. 1, p. 2, p. 3, p. 5, p. 7, p. 9 |
Local Evidence Sources
- Source PDF path:
US-pdf/Target Speaker Detection with Concealed EEG Around the Ear.pdf - Public PDF path:
/papers/10-mirkovic-2016.pdf - Categorized PDF path:
library/pdfs_by_category/03_ceegrid_task_validation/10_2016_mirkovic_et_al_target_speaker_detection_with_concealed_eeg_around_the_ear.pdf - Extracted text path:
library/texts/03_ceegrid_task_validation/10_2016_mirkovic_et_al_target_speaker_detection_with_concealed_eeg_around_the_ear.txt - Detailed card source:
library/DETAILED_PAPER_CARDS_BATCH_2.md - Page-level evidence index:
library/EVIDENCE_INDEX.md
Close Reading Checklist
- Verify exact figures, tables, page numbers, and statistics against the local PDF before formal citation.
- Keep missing parameters as
Not reportedunless the PDF or supplementary material confirms them. - Mark any cross-paper synthesis as interpretation rather than a single-paper claim.