复杂噪声场景构建

Example Image

复杂噪声场景下表现

示例：同一说话人无关语音

目标语音

ASR: whats the currency exchange rate for china

TTS无关语音

ASR: you should store the fruit in the fridge

场景一：不同位置插入无关语音

源语音
插入后语音
Oracle: resume the song from the audiobook by the beatles
ASR: resume the song from the audiobook by the beatles these tomatoes are less expensive
Ground Truth: 意图: play_audiobook, 实体: [player_setting=resume, media_type=audiobook, artist_name=beatles]
Prediction: 意图: play_audiobook, 实体: [player_setting=resume, media_type=audiobook, artist_name=beatles]
源语音
插入后语音
Oracle: remind me tonight to pick up my dry cleaning at eight pm
ASR: remind me tonight to pick up do you need some water clean at eight pm
Ground Truth: 意图: calendar_set, 实体: [event_name=pick up my dry cleaning, time=eight pm]
Prediction: 意图: calendar_set, 实体: [event_name=pick up, timeofday=tonight]
源语音
插入后语音
Oracle: play a game of contra as a second player
ASR: i will do my homework as soon as possible play a game of contra as a second player
Ground Truth: 意图: play_game, 实体: [game_name=contra]
Prediction: 意图: general_quirky, 实体: [ ]（归到异常类，即拒识）

场景二：插入无关语音 & 环境噪声

源语音
插入后语音
增强后语音
Oracle: resume the song from the audiobook by the beatles
ASR: resume the song from the audiobook sorry wait for me a second someone is calling me bye the beatles
Ground Truth: 意图: play_audiobook, 实体: [player_setting_resume, media_type=audiobook, artist_name=beatles]
Prediction: 意图: play_audiobook, 实体: [player_setting_resume, media_type=audiobook, artist_name=sauni]
源语音
插入后语音
增强后语音
Oracle: remind me tonight to pick up my dry cleaning at eight pm
ASR: remind me tonight to pick up a clean at eight pm excuse me may i have some water
Ground Truth: 意图: calendar_set, 实体: [event_name=pick up my dry cleaning, time=eight pm]
Prediction: 意图: calendar_set, 实体: [event_name=pick up cleaner, time=eight pm]
源语音
插入后语音
增强后语音
Oracle: whats the currency exchange rate for china
ASR: you should store the fruit in a fridge what is the currency exchange rate for china
Ground Truth: 意图: qa_currency, 实体: [place_name=china]
Prediction: 意图: general_quirky, 实体: [ ]（归到异常类，即拒识）

场景三：插入无关语音 & 环境噪声 & 其他人声

源语音
插入后语音
增强后语音
Oracle: resume the song from the audiobook by the beatles
ASR: that was just one night sorry wait for me a second someone is calling me dont do but buy the beatles
Ground Truth: 意图: play_audiobook, 实体: [player_setting_resume, media_type=audiobook, artist_name=beatles]
Prediction: 意图: play_audiobook, 实体: [player_setting_resume, media_type=audiobook]
源语音
插入后语音
增强后语音
Oracle: play a game of contra as a second player
ASR: play a game of contra as a step i will do my homework as soon as possible it was muchier
Ground Truth: 意图: play_game, 实体: [game_name=contra]
Prediction: 意图: play_game, 实体: [game_name=country a]
源语音
插入后语音
增强后语音
Oracle: whats the currency exchange rate for china
ASR: what is it for actually china
Ground Truth: 意图: qa_currency, 实体: [place_name=china]
Prediction: 意图: general_quirky, 实体: [ ]（归到异常类，即拒识）

场景四：多位置同时插入无关语音 & 环境噪声 & 其他人声

源语音
插入后语音
增强后语音
Oracle: play a game of contra as a second player
ASR: play a game or call me i will do my homework as soon as possible i will do my homework as soon as possible
Ground Truth: 意图: play_game, 实体: [game_name=contra]
Prediction: 意图: play_game, 实体: [game_name=cricket]
源语音
插入后语音
增强后语音
Oracle: whats the currency exchange rate for china
ASR: what is the grand switch i disagree with this thought you should store the fruit in the fridge fruit for china
Ground Truth: 意图: qa_currency, 实体: [place_name=china]
Prediction: 意图: general_quirky, 实体: [ ]（归到异常类，即拒识）

错例分析

源语音
插入后语音
增强后语音
Oracle: resume the song from the audiobook by the beatles
ASR: these tomatoes are less expensive resume the song from the audiobook by the beatles
Ground Truth: 意图: play_audiobook, 实体: [player_setting_resume, media_type=audiobook, artist_name=beatles]
Prediction: 意图: cooking_recipe, 实体: [food_type=tomato soup]
源语音
插入后语音
增强后语音
Oracle: whats the currency exchange rate for china
ASR: what if you should store the fruit in the fridge trust me its so good china
Ground Truth: 意图: qa_currency, 实体: [place_name=china]
Prediction: 意图: cooking_recipe, 实体: [ ]