Live Demo

資訊四     吳肇中、鍾毓安

2015.10.22

Flow of Design

Section 1. Exist in Training set ...

- 供學術機構做進一步研究

 

- 也許是因為留學巴黎的緣故

 

- 平均年齡二十七歲的年輕男性

供學術機構做進一步研究

base model

our model

也許是因為留學巴黎的緣故

base model

our model

平均年齡二十七歲的年輕男性

base model

our model

- 不要欺騙百姓

 

- 一個人一天最多能種一分地

 

- 雙雙獲得男女的團體冠軍

Section 2. Exist only in Testing set ...

不要欺騙百姓

base model

our model

一個人一天最多能種一分地

base model

our model

雙雙獲得男女組的團體冠軍

base model

our model

- 今天天氣真好

 

- 這個怪怪的

 

- 我是台灣人,今天要出國去上海

Section 3. Our brainstorming ...

今天天氣真好

base model

our model

這個怪怪的

base model

our model

我是台灣人,今天要出國去上海

base model

our model

Conclusion

測試時間、背景雜訊的不同 ...

 

最公平的比較方法,應該是同時對兩個 live demo servers (uploaded with different models) 錄音 ...

Research Topic

Variable Length Word Encoder

Our dreamed encoder should meet two conditions, which are:

    (1) Can accept variable lengths of input vectors, but the lengths of

          the encoded vectors should be the same.

    (2) An efficient representatives of original inputs (smaller but

          sufficient).

Current idea: Auto-encoder for (2), RNN for (1)

    Design a new unsupervised learning model that embeds the

      learning concept of auto-encoder into RNN.

To be continued ...

Speech-Lab-Present-2

By Yu-An, Chung

Speech-Lab-Present-2

  • 503