soundscape-bench
收藏资源简介:
SoundScape-Bench 是一个专为评估通用音频标注任务而设计的多语言声景基准数据集。该任务要求模型全面描述音频片段中的所有可听内容,包括语音(谁在何时、以何种语言、如何说话)、声音效果、音乐和人声爆发,并将这些信息组织成一个结构化的 JSON 列表。本数据集是 LAION 通用音频标注流程(UAAP)的配套基准。数据集包含 200 个精心构建的声景音频片段(格式为 16 kHz 单声道 MP3),每个片段时长在 10 至 60 秒之间(平均 35 秒)。其核心特点是所有片段均由已知且已理解的音频元素(来自特定源数据集的语音、音效、音乐、人声爆发片段)在时间线上组合而成。因此,每个片段都拥有一个精确的、由构建过程决定的完美答案,使得评分可以完全自动化,无需人工或大语言模型评判。数据内容统计如下:总计 796 个标注事件,包括 342 个语音事件、199 个声音事件、140 个人声爆发事件和 115 个音乐事件。语音覆盖六种语言:英语、德语、法语、中文(各 69 个片段)、西班牙语(54 个)和荷兰语(12 个)。约 18% 的片段包含时间上重叠的语音轮次,增加了任务复杂性。每个音频片段都配有一个结构化的 JSON 答案文件。答案遵循统一的 UAAP 模式,包含片段 ID、时长、说话人数量、语言列表以及事件列表。事件分为四种类型(speech, vocal_burst, sound_event, music),每种类型都包含精确的时间戳(start_time, end_time)和丰富的属性描述。例如,语音事件包含转录文本、说话人 ID、语言、情感、年龄、性别、音色和说话风格;声音事件和音乐事件包含自然语言描述和响度;人声爆发事件包含爆发类型和情感。所有片段的答案也整合在一个 `answers.jsonl` 文件中。数据集的评估采用Reward指标,该指标通过匈牙利算法在预测事件和答案事件之间进行按类型匹配。每个匹配对的得分是时间重叠度(IoU)与内容相似度的乘积。对于语音,内容相似度是情感/风格描述余弦相似度与转录词错误率(WER)得分的加权组合;对于其他类型,则使用描述文本的余弦相似度。该指标旨在全面衡量模型在识别、定位和描述复杂音频场景中各类事件的能力。数据集音频内容来源于多个公开数据集:语音来自 EmoIA-HQ、Multilingual LibriSpeech 和 Emilia-Dataset;声音效果来自 AudioSet with grounded captions;音乐来自 Captioned AI Music Snippets;人声爆发来自 Improved Synthetic Vocal Bursts。数据集整体的包装和答案密钥以 CC-BY-4.0 许可发布,但提醒用户注意底层源数据集的许可,并强调该基准主要用于研究和评估目的。
SoundScape-Bench is a multilingual soundscape benchmark dataset specifically designed for evaluating general audio captioning tasks. This task requires models to comprehensively describe all audible content in an audio clip, including speech (who spoke when, in what language and how they spoke), sound effects, music, and vocal bursts, and organize this information into a structured JSON list. This dataset is a companion benchmark to the LAION Universal Audio Annotation Pipeline (UAAP). The dataset contains 200 meticulously constructed soundscape audio clips formatted as 16 kHz mono MP3, with each clip lasting between 10 and 60 seconds (averaging 35 seconds). The core characteristic of this dataset is that all clips are temporally assembled from known and well-characterized audio elements: speech, sound effects, music, and vocal burst segments sourced from specific upstream datasets. Thus, each clip has a precise ground truth answer determined by the construction process, enabling fully automated evaluation without requiring human or large language model (LLM) judgment. The dataset's content statistics are as follows: a total of 796 annotated events, including 342 speech events, 199 sound events, 140 vocal burst events, and 115 music events. Speech events cover six languages: English, German, French, and Chinese (69 clips each), Spanish (54 clips), and Dutch (12 clips). Approximately 18% of the clips contain temporally overlapping speech turns, increasing the task's complexity. Each audio clip is paired with a structured JSON answer file. The answers follow a unified UAAP schema, including clip ID, duration, number of speakers, language list, and event list. Events are categorized into four types: speech, vocal_burst, sound_event, and music, with each type containing precise timestamps (start_time, end_time) and rich attribute descriptions. For example, speech events include transcript text, speaker ID, language, emotion, age, gender, timbre, and speaking style; sound and music events include natural language descriptions and loudness; vocal burst events include burst type and emotion. The answers for all clips are also consolidated into a single `answers.jsonl` file. The dataset uses the Reward metric for evaluation, which performs type-wise matching between predicted events and ground truth events via the Hungarian algorithm. The score for each matched pair is the product of temporal intersection over union (IoU) and content similarity. For speech events, content similarity is a weighted combination of the cosine similarity of emotion/style descriptions and the word error rate (WER) score of the transcript; for other event types, cosine similarity of the descriptive text is used instead. This metric is designed to comprehensively evaluate a model's ability to identify, localize, and describe various events in complex audio scenes. The audio content of the dataset is sourced from multiple public datasets: speech segments are taken from EmoIA-HQ, Multilingual LibriSpeech, and Emilia-Dataset; sound effects from AudioSet with grounded captions; music from Captioned AI Music Snippets; and vocal bursts from Improved Synthetic Vocal Bursts. The overall packaging and answer keys of the dataset are released under the CC-BY-4.0 license, but users are reminded to adhere to the licenses of the underlying source datasets, and the benchmark is primarily intended for research and evaluation purposes.
数据集概述:SoundScape-Bench
SoundScape-Bench 是一个用于评估“通用音频标注”(universal audio annotation)任务的多语言基准数据集。该任务要求模型将音频片段中所有可听见的内容(包括语音、说话人信息、时间、语言、表达方式、音效、音乐和声音爆发)描述为一个结构化的JSON列表。
核心特性
- 设计目标:为自动化的音频标注系统提供精确、可自动评分的参考答案,避免人工评分或LLM评判的偏差。
- 构建方式:每个音频片段由已知的语音、音效、音乐和声音爆发片段拼接而成,因此其“完美答案”在构建时即可确定,支持100%自动化评分。
- 许可证:CC-BY-4.0(数据集打包和答案键);内源数据集保留各自原始许可,适用于研究/评估用途。
数据集内容
- 音频数量:200个测试片段(
clips/clip_NNN.mp3,16 kHz单声道)。 - 持续时间:10-60秒(平均35秒),时长分布均匀。
- 事件数量:包含342个语音事件、199个音效事件、140个声音爆发事件、115个音乐事件。
- 语音语言:英语、德语、法语、中文各69个,西班牙语54个,荷兰语12个。
- 特殊性质:约18%的片段包含时间上重叠的语音。
文件结构
clips/clip_NNN.mp3:音频文件。clips/clip_NNN.json:每个片段的单独答案键。answers.jsonl:合并后的答案文件,每行一个JSON对象,包含audio字段指向对应mp3。PROTOCOL.md:构建日志。
答案键格式(UAAP模式)
每个剪辑的JSON包含clip_id、duration、num_speakers、languages和events字段。事件分为四种类型,示例如下:
- 语音(speech):包含开始/结束时间、转录文本、说话人ID、语言、情绪、年龄、性别、音色、说话风格。
- 声音爆发(vocal_burst):包含开始/结束时间、说话人ID、爆发类型、情绪。
- 音效(sound_event):包含开始/结束时间、描述、响度。
- 音乐(music):包含开始/结束时间、描述、响度。
评分机制(Reward指标)
- 过程:对每种事件类型,在预测事件与答案键事件之间进行匈牙利匹配。
- 分数计算:每对匹配事件的得分为
IoU(时间) × 内容,其中:- 语音的内容分 = 0.5 × 情绪/风格标题的余弦相似度 + 0.5 × (1 - 转录的WER,中文用字符级WER)
- 音乐、音效、声音爆发的内容分 = 标题的余弦相似度(使用
google/embedding-gemma-300m) - WER计算使用
jiwer
- 最终分数:在所有答案键事件上的平均分(遗漏事件得0分)。
排行榜(Reward分数,完整200个测试集)
| 排名 | 系统 | Reward |
|---|---|---|
| 1 | Gemini 3.1 Pro (omni) | 0.297 |
| 2 | Gemini 3.5 Flash (omni) | 0.256 |
| 3 | UAAP Gemma-12B + DiCoW(pipeline默认) | 0.253 |
| 4 | UAAP Nemotron+VibeVoice → Gemma (MOSS-free) | 0.236 |
| 5 | UAAP Gemma + DiCoW | 0.233 |
| 6 | Gemini 3 Flash (omni) | 0.212 |
| 7 | UAAP legacy triple-ASR ensemble | 0.196 |
| 8 | GPT-Audio 1.5 (omni) | 0.097 |
数据来源与许可
- 语音:来自
laion/emolia-hq(英语/德语/法语)、facebook/multilingual_librispeech(西班牙语/荷兰语)、amphion/Emilia-Dataset(中文);语音“表达方式”标题由laion/BUD-E-Whisper_V1.21重新生成。 - 音效:来自
mitermix/audioset-with-grounded-captions(源自YouTube,AudioSet)。 - 音乐:来自
laion/captioned-ai-music-snippets。 - 声音爆发:来自
laion/improved_synthetic_vocal_burts。
使用方式(Python)
python from datasets import load_dataset ds = load_dataset("laion/soundscape-bench", split="test") # 读取answers.jsonl ex = ds[0]; print(ex["clip_id"], ex["duration"], ex["languages"], len(ex["events"]))
音频文件:ex["audio"] -> clips/clip_000.mp3(需同时下载clips/文件夹)




