國語語音屬性偵測器 之初步經驗 交通大學電信系 王逸如 2005/12/17.

Presentation on theme: "國語語音屬性偵測器 之初步經驗 交通大學電信系 王逸如 2005/12/17."— Presentation transcript:

1 國語語音屬性偵測器 之初步經驗 交通大學電信系 王逸如 2005/12/17

2 Outline 前言 用 TIMIT 製作之英語語音屬性偵測器 使用國語語音屬性偵測器來偵測國語/英語語音屬性
國語數字串適合做國語語音屬性偵測器效能評估語料嗎? 2005/12/17

3 前言 New generation ASR 2005/12/17 語音事件及相關知識之整合 語音屬性與事件之偵測 語音證據之確認
語音訊號 語音屬性與事件之偵測 語音事件及相關知識之整合 語音證據之確認 供決策用之證據序列 知識、模型、資料庫、以及工具設計 2005/12/17

4 Detectors in New-generation ASR
Issues of detectors in new-generation ASR What kinds of attributes, events can/need to detect? What kinds of acoustic features can be used in the detectors? The architectures of detectors. Detectors using Statistical methods Labeled training data were needed. 2005/12/17

5 Labeled speech data in Mandarin?
Auto-labeling Mandarin speech data using HMM in order to get training data for detectors The labeling accuracy of phones with short duration such as stops, are poor. Are detectors cross-language? The attributes and events in speech are language independent? 2005/12/17

6 用 TIMIT 製作之英語語音屬性偵測器 TIMIT database Train : 3.8 hrs, 140,000 phones
Test : 1.4 hrs, 50,000 phones Manner: Vowel, Fricative, Stop, Nasal, Glide, Affricate Position: Bilabial, Lab-dent, Dental, Alveolar, Velar, Glottal, Rhotic, Front, Central, Back 2005/12/17

7 Some statistics of TIMIT
TIMIT Training Data TIMIT Testing Data total frames : 1,416,713 total frames : 513,526 Manner count Frame number Min (10ms/ frame) Average frame Vowel 57463 549896 <1 9.57 20911 202289 1 9.67 Fricative 21424 195416 9.12 7724 71036 9.20 Stop 25871 106575 4.12 9176 37755 4.11 Nasal 14157 80454 5.68 5104 29043 5.69 Glide 20257 129666 6.40 7822 51199 6.55 Silence 35877 340525 9.48 12777 117734 Affricate 2031 14181 2 6.98 631 4470 7.08 2005/12/17

8 Architectures of base detector
GMM based Bayesian detector 2005/12/17

9 Segment-based detector
Performance of pronunciation manner detections Frame-based detector Segment-based detector EER(%) Bayesian ANN HMM SEG_MCE Vowel 12.3 9.0 1.7 1.8 Fricative 10.0 11.3 6.4 3.6 Stop 16.7 14.5 9.9 5.4 Nasal 8.7 12.2 11.2 Glide 16.3 15.9 8.0 6.1 Silence 9.7 3.7 2.1 0.8 Affricate 7.2 2005/12/17

10 Performance of pronunciation position detectors
EER(%) GMM-based Bayesian detector Bilabial 12.2 Lab-dent 11.0 Dental 12.7 Alveolar 12.0 Velar 12.4 Glottal 18.3 Rhotic 9.4 Front 13.5 Central 17.7 Back 17.8 2005/12/17

11 Do we need Manner-position joint detectors?
Combine the results of manner and position detectors /n/, /en/, /nx/ /m/, /em/ 2005/12/17

12 使用國語語音屬性偵測器 來偵測國語/英語語音屬性
Without labeled Mandarin speech database Use phone-level auto-alignment result to train the Mandarin manner detectors The performance of Mandarin manner detectors for English speech data The performance of Mandarin manner detectors for Mandarin speech data 2005/12/17

13 Force aligned the training data using 3-state CI phone-level HMMs
Mandarin training set TCC-300 Mandarin speech database Train : 23.9 hrs, 300,000 syllables Test : 2.4 hrs, 34,000 syllables Force aligned the training data using 3-state CI phone-level HMMs Train the GMM-based Bayesian Mandarin manner detectors 2005/12/17

14 Performance of pronunciation manner detections of Mandarin speech
Frame-based Bayesian detector EER(%) English Mandarin Vowel 12.3 10.70 Fricative 10.0 15.7 Stop 16.7 11.5 Nasal 8.7 Glide/Liquid 16.3 9.2 Silence 9.7 8.0 Affricate 7.2 2005/12/17

15 Compare the detecting results of TIMIT speech data using detectors trained from English/Mandarin
Labeling errors in Mandarin training data environment miss-match Test data : TIMIT Frame-based detector EER(%) detector trained from English from Mandarin Vowel 12.3 21.3 Fricative 10.0 26.1 Stop 16.7 31.0 Nasal 8.7 15.6 Glide (Liquid) 16.3 /l/ Silence 9.7 24.0 Affricate 7.2 18.5 2005/12/17

16 HMM force-alignment result is poor
Examples of the detection results of TIMIT-trained and TCC-trained manner detectors. HMM force-alignment result is poor Could not find Inter-syllable silence The training data of Stop, fricative, affricate, silence were poor 2005/12/17

17 Treat the GMM models in manner detectors as a 1-state HMM, they can used to force align the TCC-300 database Manner-based 1-state HMM HMM manners count min frame Average Frame Vowel 418337 1 8.80 3 9.77 Fricative 74276 8.71 11.17 Stop 76291 4.31 8.30 Nasal 119535 7.26 5.80 Liquid 14653 8.18 6.83 Silence 350316 7.53 4.16 Affricate 75889 3.88 10.30 2005/12/17

18 Segmentation position difference of stops, liquid, affricates
2005/12/17

19 國語數字串適合做國語語音屬性偵測器效能評估語料嗎?
Evaluation and Test set To test the performance of new generation ASR? Attribute-dependent test sets are needed Labeled and attribute-rich database 2005/12/17

20 The manner/position attributes of Mandarin digits
Bilabial Lab-dent Dental Alveolar Velar Palatal Front Central back Vowel yi, a_n a, er, e_ng, e_n wu, ou Fricative s Stop b Nasal n_n ng Affricate q, j Liquid l g, k, h 2005/12/17