STEAK
收藏资源简介:
STEAK(Speech-to-Text for Error of Atc readbacK)是一个专为空中交通管制读回错误任务设计的合成语音数据集。它完全通过合成方式生成,包含管制员与飞行员之间无线电通话交换的文本和对应音频。文本生成基于形式化规则和ATC通信本体,涵盖82种命令类型(如爬升/下降、航向、速度、频率转换等),这些命令通过加权抽样(权重源自真实ATC语料库ATCO2和UWB-ATCC)组合成语句,并遵循命令间的不兼容规则以确保逻辑一致性。具体数值(如航路点、机场、频率、航空公司呼号)来自OpenNav、VATEUD、Wikipedia等公开资源。音频通过TTS合成、基于Speech Accent Archive参考语音的音色/口音转换,以及添加真实ATCO2录音的噪声配置文件生成,以模拟真实ATC通信环境。数据集总规模为2,519,694条音频,每条音频对应一个角色(管制员或飞行员)的单次话语,并通过`pair_id`字段关联形成完整的对话对。数据根据音频的RMS值分为三个难度子集:STEAK-rare(简单/清晰)、STEAK-medium和STEAK-tough(困难/嘈杂)。训练/测试分割以对话对为单位,确保同一对话对的两个角色始终处于同一分割中,防止数据泄漏。该数据集适用于自动语音识别任务,特别是针对空中交通管制领域的语音理解和错误检测研究。
STEAK (Speech-to-Text for Error of ATC Readback) is a synthetic speech dataset specifically designed for the air traffic control (ATC) readback error task. It is entirely generated via synthetic methods, containing both the text and corresponding audio of radio communication exchanges between air traffic controllers and pilots. The text generation is based on formalized rules and the ATC communication ontology, covering 82 types of commands such as climb/descend, heading, speed, frequency change, etc. These commands are combined into utterances via weighted sampling, where the sampling weights are derived from real ATC corpora ATCO2 and UWB-ATCC, and the generation complies with incompatibility rules between commands to ensure logical consistency. Specific values including waypoints, airports, communication frequencies, and airline call signs are sourced from public resources such as OpenNav, VATEUD, and Wikipedia. The audio is generated through three steps: TTS synthesis, timbre and accent conversion based on reference voices from the Speech Accent Archive, and adding noise profiles extracted from real ATCO2 recordings, to simulate realistic ATC communication environments. The dataset contains a total of 2,519,694 audio clips, each corresponding to a single utterance from one role, either air traffic controller or pilot, and they are linked into complete dialogue pairs via the `pair_id` field. The dataset is partitioned into three difficulty subsets based on the RMS value of the audio clips: STEAK-rare (simple/clear), STEAK-medium, and STEAK-tough (difficult/noisy). The training/test split is conducted at the dialogue pair level, ensuring that the two utterances from the same dialogue pair are always assigned to the same split to prevent data leakage. This dataset is applicable to automatic speech recognition tasks, particularly for research on speech understanding and error detection in the air traffic control domain.
数据集名称
STEAK (Speech-to-Text for Error of ATC readbacK)
数据集简介
STEAK 是一个合成生成的数据集,模拟空中交通管制员(ATCO)与飞行员之间的无线电通话。数据集中的文本和音频均为合成生成,包含约 252 万条音频样本。
任务类型
- 自动语音识别(Automatic Speech Recognition, ASR)
语言
- 英语(en)
标签
- 空中交通管制(air-traffic-control, atc)
- 合成语音(synthetic-speech)
数据规模
- 约 250 万至 1000 万条样本(size_categories: 1M<n<10M)
许可证
- MIT(license: mit)
数据集配置
数据集包含三个子集,按难度(基于音频 RMS 值)划分:
| 子集名称 | 难度 | RMS 范围 (dBFS) | 训练集样本数 | 测试集样本数 | 总计 |
|---|---|---|---|---|---|
| STEAK-rare | 简单(清晰) | ≤ -13.08 | 419,716 | 420,218 | 839,934 |
| STEAK-medium | 中等 | (-13.08, -11.55] | 420,434 | 419,479 | 839,913 |
| STEAK-tough | 困难(嘈杂) | > -11.55 | 419,864 | 419,983 | 839,847 |
数据生成方法
- 文本生成:基于正式规则和空中交通管制本体论,可生成 82 种指令。单条话语最多包含 7 条指令(平均 2.2 条),指令权重来自真实 ATC 语料库(ATCO2 和 UWB-ATCC)的探索性数据分析。每条指令有多种管制员侧措辞和飞行员侧复诵方式,随机抽取。同时包含不兼容规则(如垂直指令互斥、地面/飞行状态冲突等)。具体数值(航路点、机场、频率、航空公司等)来源于公开数据源。
- 音频渲染:
- 声音/音色:使用 Kokoro TTS 合成,随机选取音色。
- 口音:通过 seed-vc 进行语音转换,参考音频来自 Speech Accent Archive(约 3000 名不同母语者),按性别分配(ATCO 角色男性约 70%,飞行员角色男性约 90%)。
- 噪音:添加来自真实 ATCO2 录音的噪声配置文件。
- 难度划分:基于渲染音频的 RMS 值(dBFS)的三分位数。
训练/测试集划分
- 划分基于
pair_id,确保每一对 ATCO-飞行员通话完整地位于训练集或测试集中,避免角色间信息泄漏。 - 每个子集中,训练集与测试集样本数基本平衡(约各占 50%)。
数据文件格式
- 文件格式:Parquet
- 数据集加载方式:使用 Hugging Face
datasets库,通过load_dataset("DEEL-AI/STEAK", subset_name, split="train/test")加载。
数据列说明(部分关键列)
audio: 音频文件text: 转录文本pair_id: 通话对唯一标识,同一对 ATCO 和飞行员样本共享此 IDrole: 角色(atco / pilot)voice: 使用的合成语音accent_ref: 口音参考音频来源atco2_noise_profile: 添加的 ATCO2 噪声配置文件rms: 音频 RMS 值(dBFS),用于划分难度commandes: 包含指令和条件信息(atco_condition、pilot_condition_readback等)




