Interspeech 2016 - Experiment results for Sheffield Wargame Corpora (SWC1, SWC2, SWC3)

The files in the dataset correspond to results that have been generated for Interspeech 2016 paper: "The Sheffield Wargames Corpus - Day Two and Day Three" (DOI: 10.21437/Interspeech.2016-98). This paper details a natural English speech corpora recorded in natural environment with multi-media and multi-microphones, reports baseline speech recognition performance based on standalone training and adaptation, and it also releases a Kaldi recipe for standalone training. The files in the zip file are of three types: - .ctm, which correspond to the output of the automatic speech recognition system and the columns include segment information as well as transcripts of the recognition. - .ctm.filt.sys, which correspond to scoring of the automatic speech recognition system and includes the overall word error rate as well as the number of insertions, deletions and substitutions of the overall system. - .ctm.filt.lur, which provides a more detailed decomposition of the word error rate across multiple genres. The three file types are repeated for all the results described in Table 4 and Table 5 of the paper. The following is a description about the naming convention of the files (already explained in the paper): "ihm" refers to "Individual Headset Microphone". "sdm" refers to "Single Distant Microphone". "mdm8" refers to "Multiple Distant Microphone - 8 channels". "LDA" refers to "Linear Discriminant Analysis". "MLLT" refers to "Maximum Likelihood Linear Transform". "SAT" refers to "Speaker Adaptive Training". "MMI" refers to "Maximum Mutual Information". "DNN" refers to "Deep Neural Network". "sMBR" refers to "state-level Minimum Bayes Risk". "fMLLR" refers to "feature-level Maximum Likelihood Linear Regression". "o4" refers to "maximally 4 overlapping speakers in scoring". All three file types are standard outputs that are recognized by the automatic speech recognition community and can be opened using any text editor.本数据集包含的文件对应Interspeech 2016会议论文《The Sheffield Wargames Corpus - Day Two and Day Three》（DOI: 10.21437/Interspeech.2016-98）所生成的实验结果。该论文详细介绍了一套在自然环境下录制的多模态多麦克风自然英语语音语料库，报告了基于独立训练与自适应策略的基准语音识别性能，同时发布了用于独立训练的Kaldi训练脚本。 该压缩包内的文件共分为三类： 1. .ctm文件：对应自动语音识别系统的输出结果，其列信息包含分段信息与识别转录文本； 2. .ctm.filt.sys文件：对应自动语音识别系统的性能评分结果，包含整体词错误率以及系统的插入错误、删除错误与替换错误次数； 3. .ctm.filt.lur文件：针对多种语音场景类型的词错误率提供了更细致的分解结果。 上述三类文件均覆盖了论文中表4与表5所描述的全部实验结果。 以下为论文中已阐释的文件命名规范： - ihm（Individual Headset Microphone）：头戴式单个麦克风； - sdm（Single Distant Microphone）：远场单麦克风； - mdm8（Multiple Distant Microphone - 8 channels）：8通道远场多麦克风； - LDA（Linear Discriminant Analysis）：线性判别分析； - MLLT（Maximum Likelihood Linear Transform）：最大似然线性变换； - SAT（Speaker Adaptive Training）：说话人自适应训练； - MMI（Maximum Mutual Information）：最大互信息； - DNN（Deep Neural Network）：深度神经网络； - sMBR（state-level Minimum Bayes Risk）：状态级最小贝叶斯风险； - fMLLR（feature-level Maximum Likelihood Linear Regression）：特征级最大似然线性回归； - o4（maximally 4 overlapping speakers in scoring）：评分场景下最多存在4位重叠说话人。 上述三类文件均为自动语音识别领域公认的标准输出格式，可通过任意文本编辑器打开。