# DSP Final

## Old final exams
[96-2](https://www.ptt.cc/man/NTU-Exam/DE0A/D245/D298/DE42/D249/M.1406961291.A.5C4.html)
[98-1](https://www.ptt.cc/man/NTU-Exam/DE0A/D245/D298/DE42/D249/M.1406961291.A.FD7.html)
[100-1](https://www.ptt.cc/man/NTU-Exam/DE0A/D245/D298/DE42/D249/M.1406961291.A.8D3.html)
[100-2](https://www.ptt.cc/man/NTU-Exam/DE0A/D245/D298/DE42/D249/M.1406961291.A.951.html)

## 109
1.
(a) True. Hybrid system is an approach of integrating HMMs with Neural Networks in acoustic modeling. In this approach, the state posteriors are obtained with the output of Neural Networks.

(b) False. The state transition probabilities in a Hybrid System are also modeled by the output of Neural Networks.

(c) False. PCA is like an unsupervised training method, while LDA(Linear Discriminative Analysis) is more like a supervised training method.

(d) False. In a speaker dependent speech recognition system, the model is trained with different utterances spoken by many different speakers.

(e) True. In a transformer model, for an input sequence of words, the model itself can not obtain any information regarding the position/order for each word. Thus, we need positional encoding to solve this problem.

(f) True. The difference between contextualized word embeddings(ELMo, BERT) and traditional word embeddings is that for the former the same word in different context may have different vector representations.

(g) False. “Mismatch” in acoustic environment means that the speaker utterance and the corre- sponding transcript do not match, in other words, there are word errors in the tran- script.

(h) True. One of the possible approaches to cope with the acoustic environment mismatch prob- lems is Cepstral Moment Normalization. When there is convolution noise in the MFCC features, subtracting the MFCC features y ̄ with its own mean E[y ̄] yields CMS(Cepstral Mean Subtraction) features that are immune to convolutional noise.

(i) True. Following the previous question (h), histogram equalization usually performs better than Cepstral Mean Subtraction (CMS) and Cepstral Mean and Variance Normalization (CMVN).

(j) False. The EM(Expectation Maximization) algorithm is guaranteed to converge to a global optimal point.

2.
(a) $\alpha$ =1, 所有class都被拿來平均 $\alpha$ = $inf$，只考慮正確的以外最大的那個

5
(a) Please explain what Matrix Factorization is.
把一個大矩陣拆成兩個相乘
(b) Let A be an M × N matrix. Explain why matrix factorization can help in data compression(try to explain it with such parameters as M,N).
原本的data要M*N的空間，而做了matrix factorization，如果拆成兩個矩陣就只要 M*f + f*N (f < M && f < n) 所以可以做到data compression

(c) How can Matrix Factorization be used in a movie recommendation system?
原本的矩陣可以當作M個user N部電影，而每個user對其中某幾部電影有做評分，拆成的兩個矩陣M*f f*N 這裡可以想成電影有f個特徵，那M*f的矩陣就代表每個user對不同電影特徵的weight f*N的矩陣代表每部電影有哪些特徵，求出兩個矩陣的optimal之後，可以拿M*f矩陣的一個row當作一個user去跟f*N矩陣的每個col做內積，就能算出一個user對一部電影的相似度，因此這樣就能拿來當推薦系統

6
Please explain how the EM-algorithm works as detailed as possible. What are the “E step” and the “M step”? Why do we have to use EM-algorithm in training HMMs?

E-step: 利用目前（最佳）的model參數當作latent data的distribution去建構出一個目標函數。

M-step: 調整所有可能的參數來maximize目標函數

因為HMM的state sequence是一個latent data

7
There are three types of prosodic features mentioned in the lecture. List two of them and give two examples for each.

* 音調相關的：一個音平均的pitch、一個音最高的pitch和最低的pitch的差距
* 聲音長短相關的： 停頓的時間長短、平均每個音多長

8
Pseudo relevance feedback(PRF) can be used to improve the retrieval performance. What is PRF and the basic principles for it? Why don’t we simply use the first-pass results?
對First pass retrieval的結果再用前M筆和後N筆去對所有算相似度然後重新排序，新的結果才給user看。

9.
What is the difference between extractive and abstractive summarization?
Extractive-summarization 只會用文章中有的字或詞來做出summary而Abstractive-summarization 會根據文章的內容去生成summary，其中也可能包含文章裡沒有出現的字詞

## 106
1.
(a) (5) Describe how the MAP principle can be used for speaker adaptation.
(b) (10) Describe how Maximum Likelihood Linear Regression (MLLR) canbe used for speaker adaptation.

2. (a) (10) Explain the difference between Extractive and Abstractive Summarization.
(b) (10) Explain how Maximum Margin Relevance (MMR) can be used in summarizing a document. Use an example to describe it.

3. (10) Explain and discuss why words and subword units are both useful in voice-based information retrieval for Mandarin Chinese, but each with respective limitations? Use examples in Chinese language.

4. (15) What is the mismatch in acoustic environment between training/testing conditions for speech recognition? Explain what the model-based approaches, feature-based approaches are, including mentioning the names of two examples for each of them.

5. (20) For Llinear Discriminant Analysis (LDA), explain the meaning of within/between-class scatter matrices sw/sb, and the meaning of optimization criterion.

6. (15) What is the EM algorithm? Explain how it can be used to solve the Basic Problem 3 for a discrete HMM.

7. (10) Please explain what is the “lattice” of an utterance, and what is the “expected term frequency” for a term in the lattice.

8. (a) (5) Describe what Markov Decision Process (MDP) is including itsfundamental mathematical framework.
(b) (10) Describe how MDP can be used in user-content interaction and
spoken dialogue systems.

9.
(a) (10) What is a Deep Neural Network (DNN)? How can it be used in acoustic modeling?
(b) (10) Explain how a Recurrent Neural Network (RNN) can be used in language modeling.

10. (10) What is a Random Forest? How can it be used in Tone Recognition for Mandarin?

11. (10) Explain what the Pseudo-relevance Feedback is, its purpose and how to achieve the purpose.

12. (10) Explain what the Dialogue Acts are.