兽研[内容请查看原文]
引言
声音象征一词指的是声音与特定感知元素之间的关联。心理学界对这一主题的兴趣日益增长，研究表明声音象征效应广泛影响人类，不仅体现在语言发展中[1]，还体现在行为和认知方面[2,3]。声音象征最著名的例子或许是Bouba–Kiki效应[4]。该效应描述了将伪词"Bouba"与圆形形状相关联、将"Kiki"与尖刺形状相关联的倾向[5]。这种形状与声音之间的自发匹配已通过外显匹配任务和内隐联想测验反复得到证实[6]。此外，该效应已在来自不同语言和正字法背景的成人[7,8]以及前语言期婴儿[1]中观察到，表明它是一种广泛存在的效应，出现在发育的早期阶段。然而，在非人类动物中研究Bouba–Kiki效应的结果喜忧参半。在非人类灵长类动物（如黑猩猩和大猩猩）中证明该效应的尝试未能成功[9]，这强化了该效应实际上是人类特有现象的观点。与此同时，有证据表明3日龄雏鸡（Gallus gallus）表现出圆形和尖刺形状与"Bouba"和"Kiki"声音之间的自发关联[10]。后一项研究表明，Bouba–Kiki效应可能并非与声音象征相关的语言特异性现象，而是跨模态感官属性之间系统性映射的一个实例，即跨模态对应[5]。跨模态对应在不同物种之间共享，如黑猩猩[11]、猴子[12]、狗[13,14]、鸡[10,15]和陆龟[16,17]。它们被认为具有生态学意义，因为它们能够使个体基于两个维度在自然界中相关的可能性来预测关联[8]。例如，在音高–大小对应的情况下[13,16]，这可能反映了以下事实：由于发声道的物理限制，小动物发出的声音音调高于大动物，而较大的动物发出较低音调的声音[18,19]。基于统计规律和感官信息共现的类似解释，可能也是Bouba–Kiki效应的基础。圆形物体在撞击或在表面滚动时，与相同大小的尖刺物体相比，通常产生频率更低、更连续的声音[20]。
在本研究中，我们考察成年家犬（Canis familiaris）中的Bouba–Kiki效应，旨在提供来自非人类哺乳动物的新证据，并加深我们对声音–形状关联功能的理解。由于家犬使用与人类类似的机制来处理和回应人类语言，它们构成了研究声音象征的有前景的比较模型。少数家犬能够像人类婴儿一样，在单次接触后学会将新词与新物体相关联（快速映射）[21,22]。最近的研究表明，少数狗（被称为"天赋词汇学习者"）还能够从无意中听到的言语中学习物体的标签，这是一种存在于成人和婴儿人类中的社会认知过程[23]。因此，在家犬中证明Bouba–Kiki效应不仅将为其在非人类物种中的存在提供新证据，还将为未来旨在加深对其潜在机制理解的研究奠定基础。为此，我们向狗呈现投影在两个屏幕上的视觉形状（圆形或尖刺形），并配以"Bouba"或"Kiki"声音。我们通过人工智能（AI）计算模型分析狗的角视野和探索行为，并结合人工编码，考察了它们的自发注视偏好。优先注视范式已被广泛用于考察视觉注意以及视觉与听觉刺激之间的对应关系，不仅用于人类（如婴儿[24–27]），也用于多种非人类物种，包括雏鸡[10]、恒河猴[28,29]、马[30,31]、簇绒卷尾猴[32]、倭黑猩猩[33]和家犬[19,34–36]。该范式假设注视时间是衡量个体对特定刺激的兴趣及感知或认知加工的有效指标[37]。我们预测，如果狗像人类和雏鸡一样对Bouba–Kiki效应作出反应，它们将（1）优先注视与"Bouba"声音配对的圆形形状，以及与"Kiki"声音配对的尖刺形状；（2）对其自发匹配听觉刺激的形状表现出更多探索行为。
我们测试了51只成年家犬（见补充材料表S1），其主人能够看到形状并听到声音（研究1）。为排除所谓的"聪明汉斯效应"可能导致的偏差，即主人无意的信号影响狗的反应，我们额外测试了一组19只受试犬，其主人佩戴播放音乐的耳机以防止听到测试声音（见补充材料表S2）（研究2）。在两项研究中，每只狗均接受连续四次试验，每次持续30秒。"Bouba"和"Kiki"试验的顺序在受试者之间进行了平衡。
[内容请查看原文]
材料与方法
[内容请查看原文]
受试者
[内容请查看原文]
研究一
我们测试了51只中型至大型中头型家犬：27只雄性（未去势17只，去势10只）和24只雌性（未去势9只，去势15只），年龄在1至12岁之间（平均6.3岁）（详见补充材料表S1）。犬只通过帕尔马大学数据库和社交媒体平台招募。
[内容请查看原文]
研究二
基于事后功效分析（功效=0.858），我们额外测试了19只中头型犬：8只雄性（未去势6只，去势2只）和11只雌性（未去势5只，去势6只），年龄在1至11岁之间（平均5.28岁）（详见补充材料表S2）。受试犬通过犬只训练中心招募。
[内容请查看原文]
实验设计与装置
研究一和研究二在所有实验程序上完全相同，唯一区别在于研究二中犬主对呈现给犬只的听觉刺激不知情，因为他们佩戴了播放摇滚音乐的耳机。
实验在意大利帕尔马大学犬类行为学实验室进行。实验在一个室内房间（3×4米）中进行，房间由一块木板分隔为两个区域（图1）。木板上设有两个开口，左右各一，两个相同的屏幕（27英寸E2723H，DELL，TX，USA）放置于此以显示视觉刺激（Bouba形状或Kiki形状[8]，图1）。每个屏幕后方各放置一个扬声器（JBL Flip 6蓝牙音箱）以播放听觉刺激（"Bouba"音或"Kiki"音[8]）。装置后方设有实验人员专用区域，实验人员操作一台连接屏幕、扬声器和GoPro 10的电脑，GoPro 10置于木板顶部用于视频录制。装置前方3米处放置一把供犬主使用的椅子。
[图：实验设置：使用两个屏幕，以及通过狗头部检测和角度进行计算机分析的步骤]
[内容请查看原文]
实验刺激与程序
实验通过专用Python程序运行，该程序同步听觉和视觉刺激的呈现以及视频录制。
视觉刺激包括一个黑色尖刺形状和一个黑色圆润形状（图2），基于先前研究[7,8,10]，这些研究证明了Bouba–Kiki跨模态关联在多种文化和文字系统中均存在。听觉刺激包括"Bouba"和"Kiki"的重复发音；这些声音最初由先前研究[8]的第一作者——一位受过训练的波兰女性语音学家——发音。每个听觉刺激在每次试验的30秒内呈现15次。图3以频率（Hz）为纵轴、时间（s）为横轴绘制能量分布（dB），突出了伪词"Kiki"和"Bouba"对应的独特声学特性。"Kiki"频谱图以高频频谱特征为特点，而"Bouba"频谱图则呈现低频频谱特征。
[图：代表圆润形状（左侧）和尖刺形状（右侧）的视觉刺激。]
[图：对比频谱图，频率（Hz）作为时间（s）的函数绘制，以表示能量分布（dB）]
实验前，犬主填写知情同意书。到达实验室后，犬只被允许自由探索房间3分钟，以熟悉装置和环境。在犬只进入前，共放置六份食物奖励在装置附近的地面上，每个屏幕前方两份，装置中央部分两份。在探索期间，实验人员向犬主解释实验程序。探索阶段结束后，犬主坐下并将犬只唤至身边。犬主固定住犬只，使其肩部朝向装置中央，确保犬只身体正直、面向装置中心。在研究一中，犬主仅被要求直视装置中央；而在研究二中，犬主佩戴耳机以听不到播放的声音（"Bouba"或"Kiki"）。本研究采用被试内设计，每只犬在一个约15分钟的会话中经历四次相同的测试试验。听觉刺激（"Bouba"或"Kiki"）的顺序在被试间进行了平衡和交替，视觉刺激的空间位置（左或右）也进行了交替。
声音和视觉刺激的可能组合如下：
[内容请查看原文]
BOUBA_bk："Bouba"音，Bouba形状在右侧屏幕，Kiki形状在左侧屏幕；
[内容请查看原文]
BOUBA_kb："Bouba"音，Kiki形状在右侧屏幕，Bouba形状在左侧屏幕；
[内容请查看原文]
KIKI_kb："Kiki"音，Kiki形状在右侧屏幕，Bouba形状在左侧屏幕；
[内容请查看原文]
KIKI_bk："Kiki"音，Bouba形状在右侧屏幕，Kiki形状在左侧屏幕。
每次测试会话以"Bouba"或"Kiki"音 paired with 两种可能的图像位置之一（"bk"或"kb"）开始。随后的试验呈现另一种听觉刺激。八种组合的顺序在被试间进行了平衡（补充材料表S1–S3）。
在每次试验中，听觉刺激（“Bouba”或“Kiki”）播放30秒，共重复15次，同时呈现视觉刺激。
试验分为两个阶段：
[内容请查看原文]
自发注视阶段（20秒）：图形用红色边框标出，狗被保持在固定位置。
[内容请查看原文]
探索阶段（10秒）：边框变为绿色，示意主人放开狗，允许其自由探索刺激物。
每次试验后休息1分钟，期间允许狗探索房间。休息结束时，主人召回狗并为其摆好姿势进行下一次试验。所有四次试验均重复相同程序。
[内容请查看原文]
行为编码
通过AI模型计算首次注视其中一个屏幕的行为及其潜伏期。每次试验前20秒内狗的自发注视行为由AI模型分析（见§2e）。注视屏幕的时长（秒）由AI模型计算（见§2e）。AI模型能够处理和分析超过91%的录制帧。该性能与其他动物分析研究相当；因此，AI模型被证明是一种有效且可靠的方法[38,39]。
每次试验最后10秒的探索行为（“探索bouba”、“探索kiki”、“无探索”——见补充材料表S4）由一名编码员（E.B.）通过BORIS（行为观察研究交互式软件[40]）手动编码。一名外部行为编码员还分析了50个视频，以评估与主要编码员的评分者间信度；组内相关系数（ICC）（RStudio v. 2024.09.1）显示信度良好——ICC在0.883至0.937之间（p值<0.001）。
[内容请查看原文]
角度视觉的计算分析
在前20秒内，评估了对屏幕的优先注视，以确定形状与声音之间的关联。开发了一个AI模型来分析狗的视频，以确定其注视方向，这是研究注意力和决策过程的关键指标。该模型的工作流程首先对视频数据进行标准化，通过以每秒30帧的采样率确保时间分辨率的一致性。这种标准化是必要的，以消除因视频帧率不同而产生的不一致性，这些不一致可能会给行为分析引入噪声。然后对每一帧进行处理以检测狗的头部朝向，因为头部朝向可作为注视方向的代理指标。为此，我们采用了YOLOv8模型，这是一种最先进的目标检测算法，经过优化可在不同光照条件和头部朝向下高效定位头部。YOLOv8模型[41]在包含280张图像的自定义数据集上进行了微调，每张图像均手动标注了狗头部的边界框。这些图像是从我们实验中捕获的帧集中随机抽取的。这些标注经过精心整理，为在各种场景下检测头部位置提供了可靠的基础。在确定头部位置后，模型应用Canny算法[42]检测其边缘。边缘勾勒出头部边界，随后用于生成分割掩码——表示头部内部区域。对该分割进一步分析以拟合头部周围的椭球体。我们使用椭球体而非正方形或圆形，是因为椭球体具有长轴和短轴（正方形和圆形不具备），有助于确定狗头部的方向。椭球体拟合[43]过程由应用于分割像素的二维高斯分布驱动，确保椭球体的轴与头部的自然朝向对齐。同时，模型识别帧内的屏幕，代表感兴趣的视觉刺激。这些屏幕通过搜索指定RGB（红、绿、蓝）颜色范围内至少500像素的像素块来检测，颜色范围定义为(220, 220, 220)至(255, 255, 255)之间。选择该颜色范围和像素数量是为了可靠地将屏幕与帧中的其他元素区分开来。所提出的模型假设存在两个屏幕，并提取其边界梯形以定义其空间范围。最后一步涉及计算椭球体长轴（反映头部朝向）与连接椭球体中心至每个屏幕梯形中心的线之间的角度。如果这些向量之间的角度小于30°，则该帧被分类为狗朝向相应屏幕。该角度阈值基于经验观察选择，以平衡注视分类的灵敏度和特异性。通过估计头-鼻线的方向以及评估双眼视角来评估注视右侧或左侧形状的时间。根据Maggs等人[44]的研究，我们决定仅考虑狗的双眼视觉来确定它们是否在看屏幕。在计算双眼角度后，如果双眼视角包含屏幕，则定义狗正在看屏幕。我们记录了狗双眼角度指向其中一个屏幕的帧数。基于这些假设，我们将狗正在看屏幕的帧视为狗的目光与至少一个屏幕之间存在交集的帧。
[内容请查看原文]
统计分析
从计算分析中，我们提取了首次注视两个屏幕之一的朝向、从试验开始到首次注视的潜伏期，以及双眼角度指向其中一个或另一个屏幕的帧数。从行为编码中，我们计算了在10秒探索期内对尖刺形或圆形形状的探索行为的发生率和持续时间。
为控制潜在的普遍侧向偏好，使用单样本t检验[45]分析了装置右侧和左侧的平均注视比例。为检查独立于听觉刺激的对两种形状之一的偏好，我们分析了：(i) 使用χ2检验[46]分析首次注视的方向；(ii) 使用广义线性混合效应模型（GLMM）分析首次注视潜伏期，采用伽马分布和对数链接，包括形状作为固定效应和受试者作为随机截距，使用‘glmmTMB’包[47]拟合；(iii) 使用beta分布的GLMM（函数‘glmmTMB’，包‘glmmTMB’[47]）分析注视两种形状的总时间。
为探究在播放“Kiki”或“Bouba”声音后的前20秒内，犬的注视是否朝向尖刺形或圆形，使用了带有beta分布的GLMM（函数“glmmTMB”，包“glmmTMB”[47]）。使用beta回归是因为所考虑的响应变量为犬注视尖刺形或圆形的帧数占犬注视屏幕总帧数的比例（比例介于0和1之间）。声音（二水平因子“Kiki”/“Bouba”）作为主要预测变量纳入。试验、性别和年龄作为控制预测变量纳入，受试者ID作为随机效应以控制同一个体的重复观察。
还使用GLMM来探究试验次数（检验预测变量）是否对注视屏幕时长的持续时间产生负面影响，以性别和年龄作为控制预测变量，受试者ID作为随机效应。
为分析犬在每个试验最后10秒内的探索模式，首先将大于10秒的探索值设为10秒，然后使用带有二项分布的GLMM（函数“glmmTMB”，包“glmmTMB”[47]）。犬是否探索尖刺形或圆形（发生，二项响应）作为响应变量，声音作为主要检验预测变量纳入，而试验、性别和年龄作为控制预测变量纳入。最后，使用带有gamma分布的GLMM（函数“glmmTMB”，包“glmmTMB”[47]），以评估声音刺激对朝向屏幕的探索行为的影响。作为对主要预测变量的整体检验并避免“隐蔽多重检验”[48]，我们将上述完整模型与缺少检验预测变量但在其他方面相同的零模型进行比较。我们使用似然比检验，将完整模型与一次缺少一个固定效应的简化模型进行比较，以检验各个固定效应的作用[49]。对于这些检验以及完整-零模型比较，我们使用了似然比检验[50]。每次比较的置信区间（CIs）使用通用R函数“confint”计算。通过一次从数据中剔除一个个体，并将对这些子集拟合的模型所得的估计值与对每个模型完整数据集所得的估计值进行比较，来估计模型稳定性，使用R函数“glmmTMB.stab”。这些结果显示所有模型均具有良好的稳定性（详情见补充材料）。所有模型均进行了零膨胀和一膨胀检验。未检测到膨胀的证据，这表明所选择的模型分布是合适的。所有统计分析均在R v. 4.5.1中完成。完整的包版本信息提供在Open Science Framework（OSF）存储库中。结果在p ≤ 0.05时被认为具有统计学显著性。
对于研究1和研究2，模型的稳定性使用Roger Mundry（德国莱比锡马克斯·普朗克进化人类学研究所）提供的函数进行评估，所有模型均显示出可接受的稳定性。
[内容请查看原文]
结果
[内容请查看原文]
研究1
[内容请查看原文]
对右侧或左侧的一般侧偏
单样本t检验显示对右侧或左侧屏幕没有显著的侧偏（右侧：均值 = 0.499，95% CI [0.490, 0.508]，t50 = −0.24，p = 0.81；左侧：均值 = 0.501，95% CI [0.492, 0.510]，t50 = 0.24，p = 0.81）。
[内容请查看原文]
首次注视、其潜伏期以及不考虑声音的总体注视偏好
The dogs’ first look was directed equally towards one of the two shapes (
[内容请查看原文]
[内容请查看原文]
[内容请查看原文]
注视屏幕
在20秒试验中，犬注视屏幕的平均时间为5.8秒（5.8 ± 2.1秒）。未发现试验对注视屏幕时间的影响（p = 0.24；补充材料，表S5）。
[内容请查看原文]
优先注视任务
AI模型计算了犬的双眼视觉角度以及该角度包含尖刺形或圆形的帧数比例。注视帧数的比例受到所播放听觉刺激的影响（完整-零模型比较：χ² = 215.69，p < 0.001）。具体而言，播放“Bouba”声音时，犬注视圆形的时间更长（p < 0.001，Es. −0.76 ± 0.04，95% CI [−0.84, −0.69]），而播放“Kiki”声音时，犬注视尖刺形的时间更长（p < 0.001，Es. 0.76 ± 0.04，95% CI [0.69, 0.84]）（图4和图5）。未发现试验、性别、年龄和顺序的影响。
[图：当“Bouba”（蓝色）声音播放时，狗的双目视觉朝向描绘圆润或尖刺形状屏幕的注视率]
[图：狗在听到“Bouba”声音时看向圆润形状（研究1——主人能够看到形状并听到声音）]
[内容请查看原文]
探索行为
在试验的探索阶段，当播放“Kiki”声音时，与“Bouba”条件相比，犬更频繁且持续时间更长地探查尖刺形（频率：完整-零模型比较：χ² = 7.8，p = 0.005；Es. 1.07 ± 0.4，p值 = 0.007，95% CI [0.31, 1.89]；持续时间：完整-零模型比较：χ² = 14.41，p = 0.0001，Es. 1.9 ± 0.6，p值 = 0.001，95% CI [0.89, 3.289]）（图6）。相反，对圆形的探索发生和持续时间均未受到所播放“Bouba”声音的影响。未发现试验、性别、年龄和顺序的影响。
[图：狗在听到“Kiki”声音时探索尖刺形状（研究1——主人能够看到形状并听到声音）]
[内容请查看原文]
研究2
[内容请查看原文]
对右侧或左侧的一般性侧向偏好
单样本t检验显示，对右侧或左侧屏幕均无显著的侧向偏好（右侧均值 = 0.498，95% CI [0.485, 0.511]）和（左侧均值 = 0.502，95% CI [0.489, 0.515]）（右侧：t(18) = –0.32，p = 0.75；左侧：t(18) = 0.32，p = 0.75）。
[内容请查看原文]
首次注视、其潜伏期及无论声音如何的整体注视偏好
The dogs’ first look was directed equally towards one of the two shapes (
[内容请查看原文]
[内容请查看原文]
[内容请查看原文]
注视屏幕
狗注视屏幕的平均时间为8.8秒（8.8 ± 1.7秒）。未发现试验次数对注视屏幕时间的影响（p = 0.4；补充材料，表S6）。
[内容请查看原文]
优先注视任务
AI模型的结果显示，当播放“Bouba”声音时，狗注视圆形形状的时间更长（全模型与零模型比较：χ² = 137.42，p < 0.001，Es. −0.97 ± 0.05，95% CI [−1.07, −0.88]），而当播放“Kiki”声音时，狗注视尖刺形状的时间更长（全模型与零模型比较：χ² = 137.42，p < 0.001，Es. 0.97 ± 0.05，95% CI [0.88, 1.07]）（图7）。
[图：当“Bouba”（蓝色）声音播放时，狗的双目视觉朝向描绘圆润或尖刺形状屏幕的注视率]
[内容请查看原文]
探索行为
对两个形状的探索发生率和持续时间均未受到所播放声音的影响（“Bouba”声音——发生率：全模型与零模型比较：χ² = 1.06，p = 0.30；持续时间：全模型与零模型比较：χ² = 1.62，p = 0.20；“Kiki”声音——发生率：全模型与零模型比较：χ² = 0.23，p = 0.63；持续时间：全模型与零模型比较：χ² = 0.6，p = 0.45）。未发现性别、年龄、顺序和试验的影响。
[内容请查看原文]
讨论
在研究1和研究2中，狗均未表现出对右侧或左侧屏幕的侧化，且首次注视同等指向两个形状。它们还表现出在听到“Bouba”声音时注视圆形形状的时间更长，在听到“Kiki”声音时注视尖刺形状的时间更长，从而表现出一种自发的声音-形状关联，这与先前在人类和雏鸡中报道的Bouba-Kiki效应一致。这加强了以下观点：Bouba-Kiki效应是不同物种间共享的更广泛跨模态对应关系的一部分，而非人类独有且与语言相关。与这一观点一致，大量证据表明人类和狗共享若干跨模态对应关系，包括音高-大小、音高-空间和音高-亮度关联[13,14,18]。它们能够从声音线索推断另一只狗的体型——将高音吠叫与小型犬关联，将低音吠叫与大型犬关联[18,19]。同样，它们更关注遵循这一模式的视听组合，例如高音与位于较高位置的刺激配对[13]。Korzeniowska及其同事[14]证明，狗对一致的音高-大小配对反应更快，对发出高音的小型刺激的定向比对不一致组合更快。
有趣的是，在研究1中，狗还表现出：(i) 首次注视尖刺形状时的潜伏期更短，且与听觉刺激无关；(ii) 总体上对尖刺形状存在普遍的注视偏好， regardless of the auditory stimuli；(iii) 在听到“Kiki”声音时对尖刺形状表现出探索偏好。总体而言，这些结果可能提示狗对尖刺形状的偏好强于圆形。这种偏好可能部分解释了探索阶段的结果，即在“Bouba”条件下狗未表现出形状偏好，但在听到“Kiki”时对尖刺形状的探索时间更长。这一结果与人类婴儿和雏鸡的研究发现形成对比，后者的效应在圆形-Bouba配对中更为明显。雏鸡的结果被解释为对圆形的先天知觉偏好的副产物，这种偏好在幼年动物中有利于最大化朝向同种个体（例如具有圆形轮廓的面部）的机会。我们未发现任何实证证据表明狗存在类似的形状偏向。由于对尖刺形状的总体注视偏好出现在研究1中而未出现在研究2中，可以假设狗受到了主人无意识线索的影响。然而，由于人类通常对圆形或曲线形状表现出更强的偏好而非尖刺形状[51]，这种可能性不大。另一种可能的解释是，就狗而言，动物的面部构型，尤其是考虑到古代犬和狼时，以更具棱角、“尖刺”特征为特点，这与人类婴儿典型的圆形面部形态形成对比，可能合理地导致相反的偏向。最后，由于我们测试的是成年个体，任何生命早期的圆形偏好可能已不再足够显著以影响其反应，其反应反而可能受经验因素塑造（例如咬合的可供性；食物、零食或玩具的典型形状）。还值得注意的是，在一项要求狗在自由探索期间基于质地或形状对新物体进行泛化的研究中，它们的首次接近基于形状，但随后主要选择质地相似的物体[52]。在先前关于跨模态对应的研究中也观察到了偏好的改变（例如空间-亮度[14]和音高-大小[13]），其中关联强度在重复试验中减弱，在约20次重复后达到随机水平。这与自发关联在引导生物体-环境初始互动中的生态相关性一致。在重复暴露或长时间（视觉）审视后，如本研究的情况，动物采取不同的探索策略以最大化找到相关事物的机会变得更为有利。这些假设并非相互排斥，其中一个或多个可能有助于解释所报告的效应；然而，需要进一步的实证研究来直接评估它们的贡献。
先前关于其他灵长类动物的数据未能在大型猿类（大猩猩和黑猩猩）[9]和倭黑猩猩[53]中报告Bouba-Kiki效应，提示该效应在非人灵长类中可能不存在。如果Bouba-Kiki效应代表更广泛跨模态对应的一个实例，正如来自雏鸡的证据以及基于我们的发现也来自狗的证据所支持的，那么灵长类动物似乎不太可能完全缺乏这种形式的跨模态映射。事实上，灵长类动物表现出若干其他有充分文献记载的跨模态对应[11,12]，表明其潜在机制存在于该分支中。因此，在这些研究中未能检测到Bouba-Kiki效应应谨慎解读，因为它可能反映了与所测试的极小且异质性样本相关的局限性，以及程序或情境因素，而非该现象的真正缺失。
如果跨模态对应关系的进化是为了让个体为其在环境中可能遇到的各种关联做好准备，我们可以推测这种能力在面临相似进化压力的物种之间是共享的。具体而言，在Bouba–Kiki效应的情况下，已有研究表明这种声音–形状对应关系反映了尖锐或圆形物体撞击表面时产生的声音的物理特性[20]。因此，可以合理预测非人类物种，包括其他哺乳动物和鸟类[10]，即使在缺乏语言的情况下也可能对这种关联作出反应。尽管如此，狗和人类构成一个特殊案例，因为自狗成为第一个被驯化的物种以来，它们已经在共享的人为生态位中共存了约32000年[54,55]。在这漫长的协同进化历史中，狗在神经、认知和交流领域经历了实质性的适应[21,56–62]。最近发表的一项研究[63]表明，声音象征可能促进人与狗之间有效的跨物种交流。作者证明，人类声音指令中的韵律声音象征可能增强跨物种的双向交流。具体而言，在与狗互动时，主人在发出指令时表现出自发的音高–动作一致性：“趴下”指令的基频显著低于“站立”指令[63]。目前尚不清楚这种共享的进化历史是否导致了人与狗之间更大的相似性，例如在交流中利用声音与形状匹配倾向的方式上。在人类中，声音–形状关联被认为为语言的出现和词汇的形成提供了基础；目前仍不清楚狗在处理人类语言或学习词语和指令时是否采用类似的机制。未来的研究应更深入地探讨这一课题，例如通过直接比较家养动物和野生动物来分离与人类协同进化的效应，或考察声音象征在狗感知和处理人类语言中的作用。
[内容请查看原文]
结论
我们的结果为成年家犬中存在Bouba–Kiki效应提供了证据，为近期文献增添了新内容，这些文献表明该效应可能代表一种广泛存在的声音–形状跨模态对应关系。未来研究应探索家犬对不同类型声音和形状的自发偏好、这种关联在生命早期阶段的存在情况，以及驯化和环境在促进声音–形状关联出现中的作用。
补充材料
[内容请查看原文]
致谢
我们衷心感谢所有参与本研究的狗主人及其爱犬。他们的参与和热情对我们研究的成功至关重要。我们还感谢Carlo Vitali和Tea Patri在数据收集方面的帮助。
伦理
所有程序均经帕尔马大学伦理委员会批准（批准号07/CESA/2025）。这些程序为非侵入性的，涉及自发注视和探索行为。在参与之前，主人被告知实验程序，并在实验前签署了知情同意书。
数据可及性
所有分析所依据的数据、统计代码（R脚本）和两个音频刺激材料均在开放科学框架（OSF）上公开可用：[64]。原始视频不公开提供以保护狗主人的隐私，但重现分析所需的所有衍生数据均可在该存储库中获取。
补充材料可在线获取[65]。
人工智能使用声明
人工智能工具（ChatGPT）仅用于支持语言编辑和提高文章清晰度。作者对文章内容承担全部责任。
作者贡献
E.B.：概念构思、形式分析、调查研究、方法学、可视化、撰写—初稿、撰写—审阅与编辑；G.P.：概念构思、调查研究、方法学、可视化、撰写—初稿、撰写—审阅与编辑；S.N.：概念构思、撰写—初稿、撰写—审阅与编辑；T.L.：方法学、软件、验证、撰写—初稿、撰写—审阅与编辑；M.L.：概念构思、方法学、撰写—初稿、撰写—审阅与编辑；P.V.：概念构思、方法学、监督、撰写—审阅与编辑。
所有作者对发表给予最终批准，并同意对本文中所进行的工作承担责任。
利益冲突声明
我们声明不存在竞争性利益。
资助
本工作得到PRIN 2022（国家重大利益研究项目）“声音象征的起源：一项比较研究”项目的支持，由欧盟–NextGenerationEU资助，VALS_P_22_PRIN_OSS_01，PNRR – M4C2 – I1.1，授予P.V.。本工作还得到欧盟–NextGenerationEU和帕多瓦大学2023年STARS Grants@Unipd计划的支持，项目CROSS – 声音象征的比较研究，授予M.L.。E.B.获得帕尔马大学博士奖学金资助。
参考文献
[内容请查看原文]
关联数据
补充材料
[内容请查看原文]
数据可用性声明
所有分析所依据的数据、统计代码（R 脚本）以及两个音频刺激材料均在开放科学框架（OSF）上公开可用：[64]。为保护狗主人的隐私，原始视频不公开，但重现分析所需的所有衍生数据均可在该存储库中获取。
补充材料可在线获取 [65]。
查看英文原文▾
[内容请查看原文]
Introduction
The term sound symbolism refers to the association between sounds and specific perceptual elements. Interest in this topic has grown within psychology, with studies demonstrating that sound-symbolic effects broadly influence humans, not only in language development [1], but also in behaviour and cognition [2,3]. Perhaps the most well known example of sound symbolism is the Bouba–Kiki effect [4]. This effect describes the tendency to associate the pseudowords ‘Bouba’ with a rounded shape, and ‘Kiki’ with a spiky shape [5]. This spontaneous matching between shapes and sounds has been repeatedly demonstrated through explicit matching tasks and implicit association tests [6]. Moreover, it has been observed in adults from diverse linguistic and orthographic backgrounds [7,8], as well as in preverbal infants [1], suggesting that it is a widespread effect, present at the early stages of development. However, studies investigating the Bouba–Kiki effect in non-human animals are marked by mixed outcomes. Attempts to demonstrate the effect in non-human primates, such as chimpanzees and gorillas, were unsuccessful [9], reinforcing the view that this is in fact a phenomenon unique to our species. At the same time, there is evidence that 3-day-old chicks (Gallus gallus) exhibit a spontaneous association between rounded and spiky shapes and the sounds ‘Bouba’ and ‘Kiki’ [10]. This latter study points towards the idea that, rather than being a language-specific phenomenon related to sound symbolism, the Bouba–Kiki effect may be an instance of systematic mapping between sensory attributes across modalities, namely cross-modal correspondences [5]. Cross-modal correspondences are shared across different species, such as chimpanzees [11], monkeys [12], dogs [13,14], chickens [10,15] and tortoises [16,17]. They are thought to have ecological significance, as they could enable individuals to predict association based on the likelihood of two dimensions correlating in nature [8]. For instance, in the case of pitch–size correspondence [13,16], this probably reflects the fact that, owing to physical constraints of the vocal tract, small animals produce higher-pitched sounds than large animals, whereas larger animals produce lower-pitched sounds [18,19]. A similar explanation, based on statistical regularities and the co-occurrence of sensory information, may underlie the Bouba–Kiki effect. Round objects, when hitting or rolling on a surface, typically produce lower-frequency, more continuous sounds compared with spiky objects of the same size [20].
In this study, we investigate the Bouba–Kiki effect in adult domestic dogs (Canis familiaris), aiming to provide novel evidence from non-human mammals and deepen our understanding of the function of sound–shape associations. As domestic dogs process and respond to human language using mechanisms analogous to humans, they constitute a promising comparative model for the study of sound symbolism. A minority of domestic dogs can learn to associate a new word with a new object after a single exposure, like human infants do (fast mapping) [21,22]. More recently, it has been shown that a small number of dogs (named ‘Gifted Word Learner’) are able to learn objects’ labels also from overheard speech, a socio-cognitive process that is present in both adult and infant humans [23]. Thus, demonstrating the Bouba–Kiki effect in domestic dogs would not only provide novel evidence for its presence in a non-human species, but also lay the groundwork for future studies aimed at deepening our understanding of the mechanisms underlying it. To this end, we presented dogs with visual shapes (either rounded or spiky) projected on two screens and paired with the sounds ‘Bouba’ or ‘Kiki’. We examined their spontaneous looking preference through an artificial intelligence (AI) computational model that analysed dogs’ angular vision and exploratory behaviour through manual coding. The preferential looking paradigm has been widely used to examine visual attention and the correspondence between visual and auditory stimuli not only in humans (e.g. infants [24–27]), but also in several non-human species, including chicks [10], rhesus monkeys [28,29], horses [30,31], tufted capuchin monkeys [32], bonobos [33] and domestic dogs [19,34–36]. This paradigm assumes that looking time is a valid measure of an individual's interest and perceptual or cognitive processing of a given stimulus [37]. We predicted that if dogs, like humans and chicks, respond to the Bouba–Kiki effect, they will (1) preferentially look at the round shape paired with the ‘Bouba’ sound, and the spiky shape paired with the ‘Kiki’ sound; and (2) engage in more exploratory behaviours towards the shape they spontaneously matched with the auditory stimulus.
We tested 51 adult domestic dogs (see supplementary material, table S1) with the owner able to see the shapes and listen to the sounds (Study 1). To rule out potential biases owing to the so-called Clever Hans effect, namely unintentional owner signals influencing the dogs’ responses, we tested an additional group of 19 subjects whose owners wore headphones playing music to prevent them from hearing the test sounds (see supplementary material, table S2) (Study 2). In both studies, each dog was tested in four consecutive trials, each lasting 30 s. The order of ‘Bouba’ and ‘Kiki’ trials was counterbalanced between subjects.
[内容请查看原文]
Material and methods
[内容请查看原文]
Subjects
[内容请查看原文]
Study 1
We tested 51 mesocephalic medium- to large-sized domestic dogs: 27 males (intact = 17 and neutered = 10) and 24 females (intact = 9 and neutered = 15) with an age between 1 and 12 years (mean = 6.3 years) (see supplementary material, table S1 for more details). Dogs were recruited through the University of Parma database and social media platforms.
[内容请查看原文]
Study 2
Based on a post hoc power analysis (power = 0.858), we tested 19 additional mesocephalic dogs: 8 males (intact = 6 and neutered = 2) and 11 females (intact = 5 and neutered = 6) with an age between 1 and 11 years (mean = 5.28 years) (see supplementary material, table S2 for more details). Subjects were recruited through dog training centres.
[内容请查看原文]
Experimental design and setup
Studies 1 and 2 were identical in all the experimental procedures, with the only exception being that in Study 2 the owners were unaware of the auditory stimuli presented to the dogs, as they were wearing headphones playing rock music.
The experiments were conducted in the Dog's Ethology Lab of the University of Parma, Italy. The experiment was conducted in an indoor room (3 × 4 m), which was divided into two sections by a wooden panel (figure 1). The wooden panel featured two openings, one on the right and one on the left, where two identical screens (27 inch E2723H, DELL, TX, USA) were positioned to display the visual stimuli (Bouba shape or Kiki shape [8], figure 1). Two speakers (JBL Flip 6 Bluetooth speakers) were placed behind each screen to deliver the auditory stimuli (‘Bouba’ sound or ‘Kiki’ sound [8]). Behind the apparatus, there was a designated area for the experimenter, who operated a computer connected to the screens, speakers and a GoPro 10, placed on top of the wooden panel for video recording. In front of the apparatus, at 3 m, a chair for the dog's owner was positioned.
[图：Setup of the experiment with the two screens, and steps of the computational analyses with dog's head detection and angu]
[内容请查看原文]
Experimental stimuli and procedure
The experiment was run with a dedicated Python program, which synchronized the auditory and visual stimuli presentation along with video recording.
The visual stimuli consisted of a black spiky shape and a black rounded shape (figure 2) based on previous studies [7,8,10], which demonstrated the cross-modal association Bouba–Kiki across various cultures and writing systems. The auditory stimuli comprised repeated pronunciations of ‘Bouba’ and ‘Kiki’; these sounds were originally spoken by the first author of a previous study [8], a trained female Polish phonetician. Each auditory stimulus was presented 15 times over a 30 s period per trial. Figure 3, in which the frequency (Hz) is plotted as a function of time (s) to represent energy distribution (dB), highlights the distinctive acoustic properties corresponding to the pseudowords ‘Kiki’ and ‘Bouba.’ The ‘Kiki’ spectrogram is characterized by a high-frequency spectral profile, whereas the ‘Bouba’ spectrogram exhibits a low-frequency spectral profile.
[图：Visual stimuli representing the rounded shape (on the left) and the spiky shape (on the right).]
[图：Comparative spectrogram, with the frequency (Hz) plotted as a function of time (s) to represent energy distribution (dB)]
Before the experiment, the owner filled out a consent form. Upon arrival at the laboratory, the dog was allowed to freely explore the room for 3 min, to get comfortable with the apparatus and the environment. Before the dog entered, a total of six pieces of food rewards were placed on the ground near the apparatus, two in front of each screen and two in the centre part of the apparatus. During the exploration, the experimenter explained the procedure to the dog's owner. Following the exploration phase, the owners sat and called the dog near them. The dog was held still by the owner with its shoulders aligned towards the middle of the apparatus, ensuring the dog was positioned straight, facing the centre of the apparatus. In Study 1, the owner was only instructed to look straight towards the middle of the apparatus; instead, in Study 2, the owner wore headphones in order not to hear the sound played (‘Bouba’ or ‘Kiki’). The study followed a within-subject design where each dog underwent four identical test trials in one session lasting about 15 min. The order of auditory stimuli (‘Bouba’ or ‘Kiki’) was counterbalanced and alternated across the four trials across dogs, and the spatial position (left or right) of the visual stimuli alternated.
The possible combinations of sound and visual stimuli are reported below:
[内容请查看原文]
BOUBA_bk: ‘Bouba’ sound with the Bouba shape on the right screen and the Kiki shape on the left screen;
[内容请查看原文]
BOUBA_kb: ‘Bouba’ sound with the Kiki shape on the right screen and the Bouba shape on the left screen;
[内容请查看原文]
KIKI_kb: ‘Kiki’ sound with the Kiki shape on the right screen and the Bouba shape on the left screen;
[内容请查看原文]
KIKI_bk: ‘Kiki’ sound with the Bouba shape on the right screen and the Kiki shape on the left screen.
Each test session started with either the ‘Bouba’ or ‘Kiki’ sound paired with one of the two possible image positionings (‘bk’ or ‘kb’). The subsequent trial featured the other auditory stimulus. The order of the eight combinations was counterbalanced across subjects (supplementary material, tables S1–S3).
During each trial, the auditory stimulus (either ‘Bouba’ or ‘Kiki’) was played for 30 s, with a total of 15 repetitions, while the visual stimuli were shown simultaneously.
The trial was divided into two phases:
[内容请查看原文]
Spontaneous looking phase (20 s): the shapes were outlined with a red frame, and the dog was held in position.
[内容请查看原文]
Exploratory phase (10 s): the frame turned green, signalling the owner to release the dog, allowing it to explore the stimuli freely.
A 1 min break followed each trial, during which the dog was allowed to explore the room. At the end of the break, the owner recalled the dog and positioned it for the next trial. The same procedure was repeated for all four trials.
[内容请查看原文]
Behavioural coding
The first look towards one of the two screens and its latency were calculated through the AI model. The behaviour of spontaneous looking of the dogs in the first 20 s of each trial session was analysed by an AI model (see §2e). The duration (s) of looking at the screens was calculated by the AI model (see §2e). The AI model was able to process and analyse over 91% of the recorded frames. This performance is comparable with other animal analysis studies; thus the AI model proved to be an effective and reliable method [38,39].
The exploratory behaviour of the last 10 s of each trial (‘exploring bouba’, ‘exploring kiki’, ‘no exploration’—see supplementary material, table S4) was coded manually through BORIS (Behavioral Observation Research Interactive Software [40]) by one coder (E.B.). An external behaviour coder also analysed 50 videos for the inter-rater reliability with the main coder; intra-class correlations (ICCs) (RStudio v. 2024.09.1) revealed good reliability—ICC between 0.883 and 0.937 (p-value < 0.001).
[内容请查看原文]
Computational analyses of angular vision
In the first 20 s, the preferential looking towards the screen was evaluated to define an association between the shape and the sound. An AI model was developed to analyse videos of dogs to determine their gaze direction, a critical metric for studying attention and decision-making processes. The proposed model's workflow began by standardizing video data, ensuring consistency in temporal resolution by sampling frames at 30 frames per second. This standardization was necessary to eliminate inconsistencies arising from varied video frame rates, which could introduce noise into behavioural analyses. Each frame was then processed to detect the dog's head orientation, as it serves as a proxy for gaze direction. For this purpose, we employed the YOLOv8 model, a state-of-the-art object detection algorithm optimized to locate the head efficiently, even under varying lighting conditions and head orientations. The YOLOv8 model [41] was fine-tuned on a custom dataset of 280 images, each manually annotated with bounding boxes around the dogs’ heads. The images were taken at random from the set of frames captured in our experiments. These annotations were carefully curated to provide a robust foundation for detecting head positions across diverse scenarios. After identifying the head's position, the model applied the Canny algorithm [42] to detect its edge. The edge delineated the border of the head, which was then used to generate a segmentation mask—representing the head's interior region. This segmentation was further analysed to fit an ellipsoid around the head. We used the ellipsoid, and not squares or circles, because it has the main and secondary axes (not present in squares and circles) useful to determine the direction of the dog's head. The ellipsoid fitting [43] process was driven by a two-dimensional Gaussian distribution applied to the segmented pixels, ensuring that the ellipsoid's axes aligned with the head's natural orientation. Simultaneously, the model identified screens within the frame, representing the visual stimuli of interest. These screens were detected by searching for blocks of pixels within a specified RGB (red, green and blue) colour range, defined between (220, 220, 220) and (255, 255, 255) of at least 500 pixels. This colour range and number of pixels were chosen to reliably isolate screens from other elements in the frame. The proposed model assumed the presence of two screens and extracted their bounding trapezoids to define their spatial extents. The final step involved computing the angle between the major axis of the ellipsoid (which reflects the head's orientation) and the line connecting the ellipsoid's centre to the centre of each screen's trapezoid. If the angle between these vectors was less than 30°, the frame was classified as the dog being oriented towards the respective screen. This angle threshold was selected based on empirical observations to balance sensitivity and specificity in gaze classification. The time spent looking at the right or left shape was assessed through the estimation of the direction of the head–-nose line and the assessment of the binocular angle of vision. In accordance with Maggs and colleagues [44], we decided to consider only the binocular vision of dogs to determine if they were looking at the screen or not. After computing the binocular angle, we defined the dog to be looking at a screen if the binocular vision angle included the screen. We noted the number of frames in which the dog had its binocular angle directed towards one of the screens. Based on these assumptions, we considered the frames in which the dog was looking at a screen to be those where there is an intersection between the dog's gaze and at least one of the screens.
[内容请查看原文]
Statistical analysis
From the computational analyses, we extracted the first look towards one of the two screens, latency from the start of the trial to the first look, and the number of frames in which the binocular angle was directed to one or the other screen. From the behavioural coding, we calculated the occurrence and duration of the exploratory behaviour towards the spiky or rounded shape in the 10 s of exploration.
To control for a potential general side bias, the mean looking proportion towards the right and left sides of the apparatus was analysed using a one-sample t-test [45]. To check for a preference towards one of the two shapes independently from the auditory stimuli, we analysed: (i) the direction of the first look with a χ2-test [46]; (ii) the first look latency using a generalized linear mixed-effects model (GLMM) with a gamma distribution and log link, including shape as a fixed effect and subject as random intercept, fitted with the ‘glmmTMB’ package [47]; (iii) the overall time spent looking at the two shapes using a GLMM with a beta distribution (function ‘glmmTMB’, package ‘glmmTMB’ [47]).
To investigate whether the dogs’ gaze, in the first 20 s, was oriented at the spiky or rounded shape when the ‘Kiki’ or the ‘Bouba’ sound was played, a GLMM with a beta distribution (function ‘glmmTMB’, package ‘glmmTMB’ [47]) was used. A beta regression was used since the response variable considered was the proportion of frames in which the dog was looking at the spiky or rounded shape divided by the total number of frames in which the dog was looking at the screens (proportion bounded between 0 and 1). The sound (two-level factor ‘Kiki’/‘Bouba’) was included as the main predictor. Trial, sex and age were included as control predictors, and the subject ID as a random effect to control for repeated observations of the same individual.
A GLMM was also used to find whether the number of trials (test predictor) influenced the duration of the looking time at the screen negatively, with sex and age as control predictors and subject ID as a random effect.
To analyse the exploration pattern of the dogs during the last 10 s of each trial, firstly exploration values greater than 10 s were set to 10 s, then a GLMM with binomial distribution was used (function ‘glmmTMB’, package ‘glmmTMB’ [47]). Whether the dogs explored the spiky or rounded shape (occurrence, binomial response) was used as a response variable, and the sound was included as the main test predictor, whereas trial, sex and age were included as control predictors. Finally, a GLMM with a gamma distribution was used (function ‘glmmTMB,’ package ‘glmmTMB’ [47]), to evaluate the influence of sound stimuli on the exploratory behaviour towards the screens. As an overall test of the main predictors and to avoid ‘cryptic multiple testing’ [48], we compared the full models as described above with null models lacking the test predictor but otherwise identical. We tested the effect of individual fixed effects using likelihood ratio tests comparing the full models with reduced models lacking the fixed effects one at a time [49]. For these tests as well as the full–null model comparison, we used a likelihood ratio test [50]. Confidence intervals (CIs) for each comparison were calculated with the generic R function ‘confint’. Model stability was estimated, dropping the individuals one at a time from the data and comparing the estimates derived for models fitted to these subsets with those obtained for the full dataset for each model, with the R function ‘glmmTMB.stab’. These revealed all the models to be of good stability (for details see the supplementary material). All models were tested for zero- and one-inflation. No evidence of inflations was detected, which suggests an appropriate chosen model distribution. All statistical analyses were performed in R v. 4.5.1. Full package version information is provided in the the Open Science Framework (OSF) repository. Results were considered statistically significant at p ≤ 0.05.
For Studies 1 and 2, the stability of the models was assessed using a function provided by Roger Mundry (Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany), and all of them revealed acceptable stability.
[内容请查看原文]
Results
[内容请查看原文]
Study 1
[内容请查看原文]
General side bias towards the right or left
One-sample t-test revealed no significant side bias towards the right or left screen (for the right side: mean = 0.499, 95% CI [0.490, 0.508], t50 = −0.24, p = 0.81; for the left side: mean = 0.501, 95% CI [0.492, 0.510], t50 = 0.24, p = 0.81).
[内容请查看原文]
First look, its latency and overall looking preference regardless of the sound
The dogs’ first look was directed equally towards one of the two shapes (
[内容请查看原文]
= 1.96, p = 0.161). A likelihood ratio test revealed a significant effect of shape on the latency responses (full–null model comparison:[内容请查看原文]
= 4.647, p = 0.031), with dogs showing shorter latencies when looking at the spiky shape than the rounded shape (estimate (Es.) −0.18 ± 0.08, p = 0.03, 95% CI [−0.35; −0.02]). Dogs looked more towards the spiky shape than the rounded shape (full–null model comparison: χ² = 33.764, p < 0.001; looking at the spiky shape: Es. 0.27 ± 0.05, p < 0.001, 95% CI [0.18; 0.36]).[内容请查看原文]
Looking at the screen
Dogs looked at the screens for an average of 5.8 s (5.8 ± 2.1 s) out of 20 s of trial. No effect of the trial on time spent looking at the screens was found (p = 0.24; supplementary material, table S5).
[内容请查看原文]
Preferential-looking task
The AI model computed dogs’ binocular vision angle and the proportion of frames in which that angle included the spiky or rounded shape. The proportion of looking frames was influenced by the auditory stimuli played (full–null model comparison: χ² = 215.69, p < 0.001). Specifically, dogs looked longer at the rounded shape when the ‘Bouba’ sound was played (p < 0.001, Es. −0.76 ± 0.04, 95% CI [−0.84, −0.69]), and at the spiky shape when the ‘Kiki’ sound was played (p < 0.001, Es. 0.76 ± 0.04, 95% CI [0.69, 0.84]) (figures 4 and 5). No effects of trial, sex, age and order were found.
[图：Looking rate of dogs’ binocular vision towards the screen depicting the rounded or spiky shape when the ‘Bouba’ (blue) o]
[图：Dog looks at the rounded shape when hearing the ‘Bouba’ sound (Study 1—the owner was able to see the shapes and listen t]
[内容请查看原文]
Exploratory behaviours
During the exploratory phase of the trials, when the ‘Kiki’ sound was played, dogs investigated the spiky shape more frequently and for a longer duration compared with the ‘Bouba’ condition (frequency: full–null model comparison: χ² = 7.8, p = 0.005; Es. 1.07 ± 0.4, p-value = 0.007, 95% CI [0.31, 1.89]; duration: full–null model comparison: χ² = 14.41, p = 0.0001, Es. 1.9 ± 0.6, p-value = 0.001, 95% 95% CI [0.89, 3.289]) (figure 6). On the contrary, neither the occurrence nor the duration of exploration of the rounded shape was influenced by the ‘Bouba’ sound played. No effects of trial, sex, age and order were found.
[图：Dog explores the spiky shape when hearing the ‘Kiki’ sound (Study 1—the owner was able to see the shapes and listen to t]
[内容请查看原文]
Study 2
[内容请查看原文]
General side bias towards the right or left
One-sample t-test revealed no significant side bias towards the right or left screen (right mean = 0.498, 95% CI [0.485, 0.511]) and (left mean = 0.502, 95% CI [0.489, 0.515]) (right: t(18) = –0.32, p = 0.75; left: t(18) = 0.32, p = 0.75).
[内容请查看原文]
First look, its latency and overall looking preference regardless of the sound
The dogs’ first look was directed equally towards one of the two shapes (
[内容请查看原文]
= 0.47, p = 0.49). A likelihood ratio test showed no significant difference in the latency responses towards one of the two shapes (full–null model comparison:[内容请查看原文]
= 1.82, p = 0.17). The dogs, in general, did not show an overall looking preference towards one of the two shapes (full–null model comparison: χ² = 0.39, p = 0.53).[内容请查看原文]
Looking at screens
Dogs looked at the screen for an average of 8.8 s (8.8 ± 1.7 s). No effect of the number of trials on time looking at the screen was found (p = 0.4; supplementary material, table S6).
[内容请查看原文]
Preferential-looking task
The results of the AI model showed that dogs looked longer at the rounded shape when the ‘Bouba’ sound was played (full–null model comparison: χ² = 137.42, p < 0.001, Es. −0.97 ± 0.05, 95% CI [−1.07, −0.88]), and at the spiky shape when the ‘Kiki’ sound was played (full–null model comparison: χ² = 137.42, p < 0.001, Es. 0.97 ± 0.05, 95% CI [0.88, 1.07]) (figure 7).
[图：Looking rate of dogs’ binocular vision towards the screen depicting the rounded or spiky shape when the ‘Bouba’ (blue) o]
[内容请查看原文]
Exploratory behaviour
Neither the occurrence nor the duration of exploration of the two shapes was influenced by the sounds played (‘Bouba’ sound—occurrence: full–null model comparison: χ² = 1.06, p = 0.30; duration: full–null model comparison: χ² = 1.62, p = 0.20; ‘Kiki’ sound—occurrence: full–null model comparison: χ² = 0.23, p = 0.63; duration: full–null model comparison: χ² = 0.6, p = 0.45). No effects of sex, age, order and trial were found.
[内容请查看原文]
Discussion
In both Study 1 and Study 2, dogs showed no lateralization towards either the right or the left screen and the first gaze was directed equally towards the two shapes. They also showed longer looking times towards the rounded shape when hearing the ‘Bouba’ sound, and towards the spiky shape when hearing the ‘Kiki’ sound, thus displaying a spontaneous sound–shape association consistent with the Bouba–Kiki effect previously reported in humans and baby chicks. This strengthens the view that the Bouba–Kiki effect is part of a broader set of cross-modal correspondences shared among different species, instead of it being unique to humans and linked to language. In line with this idea, substantial evidence indicates that humans and dogs share several cross-modal correspondences, including pitch–size, pitch–space and pitch–luminance associations [13,14,18]. They can infer another dog's body size from vocal cues—linking high-pitched growls to small dogs and low-pitched growls to large ones [18,19]. Similarly, they pay more attention to audio–visual combinations that follow this pattern, such as high-pitch sounds paired with stimuli located higher [13]. Korzeniowska and colleagues [14] demonstrated that dogs respond faster to congruent pitch–size pairings, showing quicker orientation towards small stimuli emitting high-pitched sounds than towards incongruent combinations.
Interestingly, in Study 1, dogs also displayed: (i) a shorter latency when first looking at the spiky shape, independent of the auditory stimulus; (i) an overall general looking preference for the spiky shape, regardless of the auditory stimuli; (iii) an exploratory preference towards the spiky shape when hearing the ‘Kiki’ sound. Overall, these results may suggest a stronger preference for spikiness over roundness in dogs. This preference may partially explain the results from the exploratory phase, where dogs did not show a shape preference in the ‘Bouba’ condition, but spent longer exploring the spiky shape when hearing ‘Kiki’. This result is in contrast with findings in human infants and in chicks, in which the effect appeared stronger for the rounded–Bouba pairing. The result in baby chicks was interpreted as a byproduct of an innate perceptual preference for roundness, which is advantageous in young animals for maximizing the chances of orienting towards a conspecific (e.g. a face, which has a round outline). We are unaware of any empirical evidence for a comparable shape bias in dogs. As the overall looking preference for spiky shapes emerged in Study 1 but not in Study 2, it could be hypothesized that the dogs were influenced by the owners’ unintentional cues. However, since humans typically show a stronger preference for rounded or curved shapes rather than spiky ones [51], this is unlikely. Another possible explanation could be that, in the case of dogs, animals’ facial configuration, especially when considering ancient dogs and wolves, is characterized by more angular, ‘spiky’ features, which contrasts with the round facial morphology typical of human infants and could plausibly contribute to an opposite bias. Finally, as we tested adult individuals, any early-life preference for roundness may no longer be sufficiently salient to influence their responses, which could instead be shaped by experiential factors (e.g. affordances for biting; the typical shape of food, treats or toys). It is also worth noting that in a study where dogs were required to generalize novel objects based on texture or shape during free exploration, their first approach was based on shape, but they later predominantly chose objects with similar texture [52]. Modifications in preference have also been observed in previous studies on cross-modal correspondences (e.g. space–luminance [14] and pitch–size [13]), where the strength of the association weakened over repeated trials, reaching chance level after about 20 repetitions. This is consistent with the ecological relevance of spontaneous associations in guiding initial organism–environment interactions. After repeated exposure or prolonged (visual) inspection, as in the case of the present study, it becomes more advantageous for the animal to adopt a different exploratory strategy to maximize the chances of finding something relevant. These hypotheses are not mutually exclusive, and one or several of them may help explain the reported effect; however, further empirical investigation is needed to directly assess their contribution.
Previous data on other primates failed to report the Bouba–Kiki effect on both great apes (gorillas and chimpanzees) [9] and bonobo [53], suggesting that it may be absent in non-human primates. If the Bouba–Kiki effect represents one instance of broader cross-modal correspondences, as supported by evidence from chicks and, based on our findings, also from dogs, it seems unlikely that primates would completely lack this form of cross-modal mapping. Indeed, primates show several other well documented cross-modal correspondences [11,12], indicating that the underlying mechanisms are present in this clade. Therefore, the failure to detect the Bouba–Kiki effect in these studies should be interpreted with caution, as it may reflect limitations related to the very small and heterogeneous samples tested, as well as procedural or contextual factors, rather than a true absence of the phenomenon.
If cross-modal correspondences evolved to prepare the individual for the kinds of associations they are probably to encounter in their environment, we can speculate that this ability would be shared across species that faced similar evolutionary pressures. Specifically, in the case of the Bouba–Kiki effect, it has been shown that this sound–shape correspondence mirrors the physical properties of sounds produced when spiky or round objects hit a surface [20]. As such, it is reasonable to predict that non-human species, including other mammals and birds [10], may respond to this association even in the absence of language. Nonetheless, dogs and humans constitute a special case because they have occupied a shared anthropogenic niche for approximately 32 000 years since dogs became the first domesticated species [54,55]. During this long co-evolutionary history, dogs have undergone substantial adaptations in neural, cognitive and communicative domains [21,56–62]. A study recently published [63] suggests that sound symbolism may facilitate effective interspecies communication between humans and dogs. The authors demonstrated that prosodic sound symbolism in human vocal commands may enhance mutual cross-species communication. Specifically, when interacting with dogs, owners show spontaneous congruent pitch–action associations when giving commands: the ‘down’ command has a significantly lower fundamental frequency than the ‘stand’ command [63]. It remains unclear whether this shared evolutionary history has led to greater similarities between humans and dogs, for example in the way predispositions for matching sounds to shape are exploited in communication. In humans, sound–shape associations are thought to provide a foundation for language emergence and vocabulary formation; it is still unknown whether dogs employ similar mechanisms when processing human speech or learning words and commands. Future research should investigate this topic more thoroughly, for example by conducting a direct comparison between domestic and wild animals to isolate the effect of co-evolution with humans, or by examining the role of sound symbolism in dogs’ perception and processing of human speech.
[内容请查看原文]
Conclusion
Our results provide evidence of the Bouba–Kiki effect in adult domestic dogs, contributing to recent literature that suggests this effect may represent a widespread case of sound–shape cross-modal correspondence. Further research should explore domestic dogs’ spontaneous preferences for different kinds of sounds and shapes, the presence of this association in early stages of life and the role of domestication and environment in favouring the emergence of sound–shape association.
Supplementary Material
[内容请查看原文]
Acknowledgements
We sincerely thank all the dog owners who contributed to this study together with their dogs. Their participation and enthusiasm were invaluable to the success of our research. We also thank Carlo Vitali and Tea Patri for their help in data collection.
Ethics
All the procedures were approved by the ethical committee of the University of Parma (approval numbers 07/CESA/2025). The procedures were non-invasive and involved spontaneous looking and exploratory behaviour. Prior to participation, the owners were informed about the experimental procedure and they signed a consent form before the experiment.
Data accessibility
All data underlying the analyses, statistical code (R scripts) and the two audio stimuli are publicly available on the Open Science Framework (OSF): [64]. Raw videos are not publicly available to protect the privacy of dog owners, but all derived data necessary to reproduce the analyses are available in the repository.
Supplementary material is available online [65].
Declaration of AI use
An artificial Intelligence tool (ChatGPT) was used only to support language editing and improve the clarity of the article. The authors take full responsibility for the content of the article.
Authors' contributions
E.B.: conceptualization, formal analysis, investigation, methodology, visualization, writing—original draft, writing—review and editing; G.P.: conceptualization, investigation, methodology, visualization, writing—original draft, writing—review and editing; S.N.: conceptualization, writing—original draft, writing—review and editing; T.L.: methodology, software, validation, writing—original draft, writing—review and editing; M.L.: conceptualization, methodology, writing—original draft, writing—review and editing; P.V.: conceptualization, methodology, supervision, writing—review and editing.
All authors gave final approval for publication and agreed to be held accountable for the work performed herein.
Conflict of interest declaration
We declare we have no competing interests.
Funding
This work was supported by PRIN 2022 (Progetti di Rilevante Interesse Nazionale), project ‘At the origins of sound symbolism. A comparative investigation’, funded by the European Union – NextGenerationEU, VALS_P_22_PRIN_OSS_01, PNRR – M4C2 – I1.1, awarded to P.V. This work was also supported by the European Union – NextGenerationEU and by the University of Padua under the 2023 STARS Grants@Unipd programme, project CROSS – Comparative Research of Sound Symbolism, awarded to M.L. E.B. was supported by a PhD scholarship from the University of Parma.
References
[内容请查看原文]
Associated Data
Supplementary Materials
[内容请查看原文]
Data Availability Statement
All data underlying the analyses, statistical code (R scripts) and the two audio stimuli are publicly available on the Open Science Framework (OSF): [64]. Raw videos are not publicly available to protect the privacy of dog owners, but all derived data necessary to reproduce the analyses are available in the repository.
Supplementary material is available online [65].