MagicData-AccentedMandarin-ASR-Challenge

规则说明

1. 数据：训练数据只能使用180小时的MagicData-RAMC或SLR123和MagicData提供14小时的重口音普通话对话数据下载见邮件。 允许使用公开的噪声数据集 (如 MUSAN (openslr-17), RIRNoise (openslr-28)) 进行数据增广，但需要注明来源。禁止使用其他来源的数据(包括无监督数据)训练出的预训练模型。

2. 方法：使用ASR建模方法进行建模，允许包括模型融合，预训练-finetune，无监督自适应在内的所有方法，但需要符合1中的数据使用规范。

3. 测试：测试数据与MagicData提供14小时的重口音普通话对话数据同源，数据的发布请关注官方渠道，本次任务测试集会提供对应的时间标注信息，测试集中不存在噪音符号。

4. 打分：标点符号、非语言符不参与最终 WER(此处WER即指CER，字错误率) 计算。

DATA Preparation

# magicdata-ramc
extract_magicdata_ramc.py

# accented mandarin dev 
extract_magicdata_accented_dev.py

# accented mandarin test
extract_magicdata_accented_test_noref.py
# extract_magicdata_accented_test_ref.py

关于Accented data的数据格式

样例

{
    "dataset":"magicdata-accented-dev",
    "language":"ZH",
    "version":"0.1.0",
    "creation_time":"2022-05-20",
    "update_time":"2022-05-20",
    "description":"Chinese accent Mandarin",
    "copyright":"MagicData",
    "audios":[
        {
            "aid":"S0001_0_0_0_10000640003_2_1651589423000.wav",
            "path":"magicdata-accented-data/dev/audios/S0001_0_0_0_10000640003_2_1651589423000.wav",
            "total":1234.603,
            "valid":430.197,
            "topic":[
                "entertainment"
            ],
            "device":"iPhone",
            "scene":"indoor",
            "channel":"C0",
            "segments":[
                {
                    "uttid":"S0001_0_0_0_10000640003_2_1651589423000.wav-0001047-0002054",
                    "spkid":"",
                    "begin_time":1.047,
                    "end_time":2.054,
                    "text":"[ENS]"
                } ]
        }
    ],
    "speakes":[
        {
            "spkid":"G10000001",
            "gender":"Female",
            "age":16,
            "place_birth":"CHINA,Guangdong",
            "place_residence":"CHINA,Guangdong"
        }
    ]
}

Test data

• 下载后的压缩文件大小：

$du -l magicdata-accented-data-test.tar.gz
1478496 magicdata-accented-data-test.tar.gz

• 解压后测试集文件结构

../magicdata-accented-data/test
├── audios
├── magicdata_test_noref.json
└── magicdata_test_uttid_submit.csv  

• 测试集中语音有效时长约11小时。

• 最终提交文件：magicdata_test_uttid_submit.csv,确保uttid一致

uttid,hyp
S0001_0_0_0_10000640002_2_1651589422000.wav-0005758-0008260,北京 爱 数 智慧 语音 采集

• 测试集文件格式示例

{
    "dataset":"magicdata-accented-test",
    "language":"ZH",
    "version":"0.1.0",
    "creation_time":"2022-05-20",
    "update_time":"2022-05-20",
    "description":"Chinese accent Mandarin",
    "copyright":"MagicData",
    "audios":[
        {
            "aid":"S0001_0_0_0_10000640002_2_1651589422000.wav",
            "path":"magicdata-accented-data/test/audios/S0001_0_0_0_10000640002_2_1651589422000.wav",
            "total":1235.051,
            "valid":672.612,
            "topic":[
                "entertainment"
            ],
            "device":"iPhone",
            "scene":"indoor",
            "channel":"C0",
            "segments":[
                {
                    "uttid":"S0001_0_0_0_10000640002_2_1651589422000.wav-0005758-0008260",
                    "spkid":"",
                    "begin_time":5.758,
                    "end_time":8.26,
                    "text":""
                }
            ]
        },
        {
            "aid":"S0001_0_0_0_10000640963_2_1652196918000.wav",
            "path":"magicdata-accented-data/test/audios/S0001_0_0_0_10000640963_2_1652196918000.wav",
            "total":1810.901,
            "valid":729.072,
            "topic":[
                "food"
            ],
            "device":"iPhone",
            "scene":"indoor",
            "channel":"C0",
            "segments":[
                {
                    "uttid":"S0001_0_0_0_10000640963_2_1652196918000.wav-0001130-0002140",
                    "spkid":"",
                    "begin_time":1.13,
                    "end_time":2.14,
                    "text":""
                }
            ]
        }
    ]
}

Training

./run.sh

Submit

submit csv file

uttid,hyp
...

Baseline result

Model	Corr	Sub	Del	Ins	WER
CNN+TDNNF	76.596	18.35	5.049	0.875	24.28
CNN+TDNNF+finetune	81.876	15.053	3.07	1.12	19.24

Baseline model

采用传统的Hybrid的建模方式，基于Kaldi开源工具搭建了简易的重口音对话ASR 赛道的基线系统。首先用chain模型对北京爱数智慧提供的160小时中文对话数据训练了一个CNN+TDNN-F的基础模型，然后使用14小时的重口音普通话对话数据集进行了声学模型的自适应。

CNN+TDNNF+finetune