File size: 54,303 Bytes
6fa4bc9 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 |
{
"paper_id": "2020",
"header": {
"generated_with": "S2ORC 1.0.0",
"date_generated": "2023-01-19T12:14:03.823999Z"
},
"title": "Vietnamese Text-To-Speech Shared Task VLSP 2020: Remaining problems with state-of-the-art techniques",
"authors": [
{
"first": "Thi",
"middle": [
"Thu"
],
"last": "Nguyen",
"suffix": "",
"affiliation": {},
"email": ""
},
{
"first": "",
"middle": [],
"last": "Trang",
"suffix": "",
"affiliation": {
"laboratory": "",
"institution": "Hanoi University of Science and Technology Hanoi",
"location": {
"country": "Vietnam"
}
},
"email": "trangntt@soict.hust.edu.vn"
},
{
"first": "Nguyen",
"middle": [],
"last": "Hoang Ky",
"suffix": "",
"affiliation": {
"laboratory": "R&D Lab Vbee Services and Data Processing Solution Jsc. Hanoi",
"institution": "",
"location": {
"country": "Vietnam"
}
},
"email": ""
},
{
"first": "Pham",
"middle": [],
"last": "Quang",
"suffix": "",
"affiliation": {
"laboratory": "R&D Lab Vbee Services and Data Processing Solution Jsc. Hanoi",
"institution": "",
"location": {
"country": "Vietnam"
}
},
"email": ""
},
{
"first": "V",
"middle": [
"U"
],
"last": "Duy Manh",
"suffix": "",
"affiliation": {
"laboratory": "R&D Lab Vbee Services and Data Processing Solution Jsc. Hanoi",
"institution": "",
"location": {
"country": "Vietnam"
}
},
"email": ""
}
],
"year": "",
"venue": null,
"identifiers": {},
"abstract": "and Technology (HUST). This was the third time we organized the Text-To-Speech shared task. In order to better understand different speech synthesis techniques on a common Vietnamese dataset, we conducted a challenge that helps us better compare research techniques in building corpusbased speech synthesizers. Participants were provided with a single training dataset including utterances and their corresponding texts. There are 7,770 utterances of a female Southwest professional speaker (about 9.5 hours). There is a total of 59 teams registered to participate in this shared task, and finally, 7 participants were evaluated online with perceptual tests. The best synthetic voice with Tacotron 2 and Hifigan vocoder with Waveglow denoiser achieved 89.3% compared to the human voice in terms of naturalness, i.e. 3.77 over 4.22 points on a 5-point MOS scale). Some reasons for a quite-big gap between the best synthetic voice with state-of-the-art synthetic techniques and the human voice were: (i) improper prosodic phrasing for long sentences and (ii) wrong/bad pronunciation for loan words.",
"pdf_parse": {
"paper_id": "2020",
"_pdf_hash": "",
"abstract": [
{
"text": "and Technology (HUST). This was the third time we organized the Text-To-Speech shared task. In order to better understand different speech synthesis techniques on a common Vietnamese dataset, we conducted a challenge that helps us better compare research techniques in building corpusbased speech synthesizers. Participants were provided with a single training dataset including utterances and their corresponding texts. There are 7,770 utterances of a female Southwest professional speaker (about 9.5 hours). There is a total of 59 teams registered to participate in this shared task, and finally, 7 participants were evaluated online with perceptual tests. The best synthetic voice with Tacotron 2 and Hifigan vocoder with Waveglow denoiser achieved 89.3% compared to the human voice in terms of naturalness, i.e. 3.77 over 4.22 points on a 5-point MOS scale). Some reasons for a quite-big gap between the best synthetic voice with state-of-the-art synthetic techniques and the human voice were: (i) improper prosodic phrasing for long sentences and (ii) wrong/bad pronunciation for loan words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "Abstract",
"sec_num": null
}
],
"body_text": [
{
"text": "Processing Consortium. It is an initiative to establish a community working on speech and text processing for the Vietnamese language [2] . The VLSP 2020 was the sixth annual international workshop. The Text-To-Speech (TTS) shared task was a challenge in the VLSP Campaign 2020, which was organized at Hanoi University of Science and Technology. This was the third time we organized the challenge in speech synthesis.",
"cite_spans": [
{
"start": 134,
"end": 137,
"text": "[2]",
"ref_id": "BIBREF1"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "I. INTRODUCTION VLSP stands for Vietnamese Language and Speech",
"sec_num": null
},
{
"text": "To the best of our knowledge, Vietnamese TTS systems can be divided into three main types:(i) Hidden Markov Model (HMM) based systems, (ii) Deep Neural Network (DNN) based systems, and (iii) state-of-the-art end-to-end systems. HMM-based TTS systems [6] [10] and DNN-based TTS systems [4] [9] need to provide pause position and loanword pronunciation in the text pre-processing step. Some end-to-end TTS systems, such as Tacotron [3] [11] , could use a massive amount of text and audio data pairs to learn prosody and loanword modeling directly from the TTS training process. Nevertheless, corpora do not always design to support that purpose.",
"cite_spans": [
{
"start": 250,
"end": 253,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 254,
"end": 258,
"text": "[10]",
"ref_id": "BIBREF9"
},
{
"start": 285,
"end": 288,
"text": "[4]",
"ref_id": "BIBREF3"
},
{
"start": 289,
"end": 292,
"text": "[9]",
"ref_id": "BIBREF8"
},
{
"start": 430,
"end": 433,
"text": "[3]",
"ref_id": "BIBREF2"
},
{
"start": 434,
"end": 438,
"text": "[11]",
"ref_id": "BIBREF10"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "I. INTRODUCTION VLSP stands for Vietnamese Language and Speech",
"sec_num": null
},
{
"text": "This shared task has been designed for understanding and figuring out remaining problems in Vietnamese TTS with state-of-the-art speech synthesis techniques on the same dataset. Based on some subjective feedback from listeners of the last year's TTS shared task, three main problems have been raising for this year: prosodic phrasing (mainly focusing on pause detection) [5] , text normalization (mainly focusing on loanwords) [6] [8] , and removing noise for Internet datasets.",
"cite_spans": [
{
"start": 371,
"end": 374,
"text": "[5]",
"ref_id": "BIBREF4"
},
{
"start": 427,
"end": 430,
"text": "[6]",
"ref_id": "BIBREF5"
},
{
"start": 431,
"end": 434,
"text": "[8]",
"ref_id": "BIBREF7"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "I. INTRODUCTION VLSP stands for Vietnamese Language and Speech",
"sec_num": null
},
{
"text": "Participants took the released speech dataset, build a synthetic voice from the data and submit the TTS system. We then synthesized a prescribed set of test sentences using each submitted TTS system. The synthesized utterances were then imported to an online evaluation system. Some perception tests were carried out to rank the synthesizers focusing on evaluating the intelligibility and the naturalness of participants' synthetic utterances.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "I. INTRODUCTION VLSP stands for Vietnamese Language and Speech",
"sec_num": null
},
{
"text": "The rest of this paper is organized as follows. Section II presents the common dataset and its preparation. Section III introduces participants and a complete process of the TTS shared task in VLSP Campaign 2020. We then show the evaluation design and experimental results in Section IV. We finally conclude the task and give some possible ideas for the next challenge in Section V.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "I. INTRODUCTION VLSP stands for Vietnamese Language and Speech",
"sec_num": null
},
{
"text": "The topic of this shared task is to address remaining problems of TTS systems using state-of-the-art synthesis techniques. Based on some analyses on the previous task results, aforementioned, we raised the following issues for this shared task: (i) prosodic phrasing (focusing on pause detection for long input sentences), (ii) text normalization (focusing on expanding loanwords), and (iii) removing background noises (of Internet audios).",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "Due to the topic of this year's task, we decided to collect audiobooks from the Internet. Vbee Jsc supported to build the dataset for this task. The corpus was taken from a novel called \"Bell to Whom the Soul\" by Hemingway, a famous American novelist. Audio stories were downloaded manually, divided into 28 long audio files, each had 30 to 60 minutes in length. These files were then automatically split into smaller audio files that are less than 10 seconds in length (using Praat scripting tool). After this process, the number of sound files was up to nearly 20,000 sound files with different lengths.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "However, approximately 10,000 sound files that were too short in length (i.e. less than 750 ms) were discarded. Next, we used the ASR API of Vais Jsc to convert the remaining 10,000 audio files into text. These data were checked by the teams participating in the contest. Each team only had to check xxx files for participation. Finally, 7,770 best quality utterances and their corresponding texts were selected as the final dataset. Even though the speaker's voice was professional and pretty, the voice still contained some background noise due to the recording device's low quality.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "III. PARTICIPANTS For TTS shared task this year, participants had to follow a complete process (Fig. 1 ), which was managed in the website of the TTS shared task of VLSP Campain 2020 (https://tts.vlsp.org.vn).",
"cite_spans": [],
"ref_spans": [
{
"start": 95,
"end": 102,
"text": "(Fig. 1",
"ref_id": "FIGREF0"
}
],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "First, each team registered to participate in the challenge. They were then provided with accounts to log into. On this site, all teams were asked to check the audio files to see if they match the corresponding text and edit if necessary. If they found that the text was exactly the content of the audio, they voted for that transcription. Each audio file needs to be checked by at least 3 teams. Audio files that had no vote after the validation process, we had to check them manually. The participants who completed the required task were asked to send their user license agreement with valid signatures. They were then able to download the training dataset. The dataset includes utterances and their corresponding texts in a text file.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "Participants were asked to build only one synthetic voice from the released database. All teams had 20 days for training and optimizing their voices. Each team then submitted the result with a TTS API following the announced specification requirement. We also supported teams that could not deploy their TTS systems to a public server by accepting their docker images that contain the TTS API. We then synthesized audio files from the text files in the test dataset using teams' TTS API. Synthesized files will be evaluated. After receiving evaluation results, the teams proceed to write and submit technical reports. Fig. 2 compares the number of participants of last year to this year. Fifty-nine teams registered for this year's challenge. Unlike last year, participants were asked to validate the provided dataset, and 19 joined the data validation process, and 15 teams obtained the data after sending the signed user agreement. Finally, nine teams, compared to four in 2019, submitted their TTS system. We synthesized testing audio through the TTS API of each team. Unfortunately, we could not use the TTS API of the two teams due to problems with their TTS system or their server. Table I gives the list of participants that had final submissions to the VLSP TTS shared task 2020. IV. EVALUATION Perceptual testing was chosen for evaluating synthetic voices. First, an intelligibility test was conducted to measure the understandability, then the MOS test, which allowed us to score and compare the global quality of TTS systems with respect to natural speech references. All subjects conducted the online evaluation via a web application. This online evaluation system was built by the School of Information and Communication Technology, Hanoi University of Science and Technology, and Vbee Jsc. This system was integrated into https://tts.vlsp.or.vn.",
"cite_spans": [],
"ref_spans": [
{
"start": 618,
"end": 624,
"text": "Fig. 2",
"ref_id": "FIGREF1"
},
{
"start": 1188,
"end": 1195,
"text": "Table I",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "II. COMMON DATASET",
"sec_num": null
},
{
"text": "They first registered on the website with necessary information including their hometowns, ages, genders, occupations. They were trained on how to use the website and how to conduct a good test. They were strictly asked to do the test in a controlled listening condition (i.e. headphones and in a quiet distraction-free environment). To ensure that the subjects focused on the test, we designed several sub-tests for each test due to a big number of testing voices (i.e. 8 voices including natural speech). As a result, each sub-test lasts from 25 to 30 minutes.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2019",
"sec_num": "2020"
},
{
"text": "On completion of any sub-test, or after logging in again, a progress page showed listeners how much they had completed. Detailed instructions for each sub-test were only shown on the page with the first part of each sub-test; subsequent parts had briefer instructions in order to achieve a simple layout and a focussed presentation of the task.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "2019",
"sec_num": "2020"
},
{
"text": "In order to address the issue of duplicate contents of stimuli, we adopted the Latin square (nxn) [1] for all sub-tests, where n is a number of voices in the sub-test. To be more specific, each subject listened to one n th of the utterances per voice, without any duplicate content. With the Latin square design, the number of subjects should be at least twice more than the ones with the normal design. Stimuli were randomly and separately presented only once to subjects. Each stimulus was an output speech of a TTS system or a natural speech for a sentence. Details of the two tests are described in the following subsections.",
"cite_spans": [
{
"start": 98,
"end": 101,
"text": "[1]",
"ref_id": "BIBREF0"
}
],
"ref_spans": [],
"eq_spans": [],
"section": "2019",
"sec_num": "2020"
},
{
"text": "In the intelligibility test, subjects were asked to write down the text of the audio they heard (Fig. 3) . The subjects might listen again a second time if they do not hear clearly or have long sentences. They only listened to the utterances the third time when the subjects were distracting, or the sentence were very long. Twenty-seven subjects participated in this test. There were two main types of subjects who participated in the test: (i) 19 students (19-22 years-old, 10 females) from Hanoi University of Science and Technology, VNU University of Science; (ii) 8 speech experts (23-38 years-old, 4 female).",
"cite_spans": [],
"ref_spans": [
{
"start": 96,
"end": 104,
"text": "(Fig. 3)",
"ref_id": "FIGREF2"
}
],
"eq_spans": [],
"section": "A. Intelligibility Test",
"sec_num": null
},
{
"text": "The testing dataset included 36 sentences. Each subject needs to participate in at least two of the three sub-tests.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A. Intelligibility Test",
"sec_num": null
},
{
"text": "Subjects (i.e. listeners) were asked to assess by giving scores to the speech they had heard (Fig. 4) . When taking this test, subjects listen to the voice once, unless they do not hear it clearly, then listen for a second time.",
"cite_spans": [],
"ref_spans": [
{
"start": 93,
"end": 101,
"text": "(Fig. 4)",
"ref_id": "FIGREF3"
}
],
"eq_spans": [],
"section": "B. MOS Test",
"sec_num": null
},
{
"text": "Subjects randomly listened to utterances and then gave their scores for the naturalness of the utterances. The question presented to subjects was \"How do you rate the naturalness of the sound you have just heard?\". Subjects could choose one of the following five options (5-scale):",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B. MOS Test",
"sec_num": null
},
{
"text": "-5: Excellent, very natural (human speech) -4: Good, natural -3: Fair, rather natural -2: Poor, rather unnatural (rather robotic) -1: Bad, very unnatural (robotic). Testing text set includes 60 sentences. There are two subtests, including 60 random utterances each (taken from 480 utterances). Table III illustrates the design for the two MOS sub-tests. We put the natural speech (NATURAL) as a reference in both sub-tests. Due to an odd number of final participated teams, sub-test 1 included 3 teams (Team 2,4,5) while sub-test 2 had voices from the remaining 4 teams (Team 6,7,8,9). Subjects participated in two sub-tests for voices built from the common dataset. Due to a rather big number of voices in each sub-test (i.e. 5 including the natural reference), we let the subjects to heard randomly half of the utterances for each voice. The number of subjects who listened to each sub-test was 48 (20 females). Each subject needs to participate in all two sub-tests, estimated at 25 to 30 minutes.",
"cite_spans": [],
"ref_spans": [
{
"start": 294,
"end": 303,
"text": "Table III",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "B. MOS Test",
"sec_num": null
},
{
"text": "Due to a large number of loanwords in the test set, the intelligibility results were not good, at about 68-89% at both word and syllable levels, even with natural speech. The subjects might do not know how to write these loanwords or present different orthography from the original text. We should have a special design and more analyses for this type of test in the future.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "A. Intelligibility Score",
"sec_num": null
},
{
"text": "The perceptual evaluation of the general naturalness was carried out on different voices of participants and a natural speech reference (NATURAL) of the same speaker as the training corpus. Fig. 5 and Table IV show the final MOS test results. Only three teams submitted technical reports, i.e. Team2, Team6, and Team7.",
"cite_spans": [],
"ref_spans": [
{
"start": 190,
"end": 196,
"text": "Fig. 5",
"ref_id": "FIGREF4"
},
{
"start": 201,
"end": 209,
"text": "Table IV",
"ref_id": "TABREF0"
}
],
"eq_spans": [],
"section": "B. MOS Score",
"sec_num": null
},
{
"text": "We can see that Team2 was the best team (i.e. 3.769)about 89.3% compared to the natural speech (i.e. 4.220/5). This team adopted Tacotron-2 as the acoustic model, and HiFi-GAN as a real-time vocoder, and Waveglow as a denoiser. Team7 was the second place with a 3.698 score (only less than the first place 0.07 point). This team used FastSpeech and PostNet, which could be considered as a faster acoustic model, compared to Tacotron-2 or only FastSpeech. Team6 was the fifth place with a 3.313 score. Their acoustic model was Tacotron2, and their vocoder was Waveglow. Although using state-of-the-art synthesis techniques that lead to a high-quality synthetic voice, there were still some remaining problems in the results of participants. Some reasons were found for a quite-big gap between the best synthetic voice with state-of-the-art synthetic techniques and the human voice: (i) improper prosodic phrasing for long sentences and (ii) wrong/bad pronunciation for loan words.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "B. MOS Score",
"sec_num": null
},
{
"text": "Several two-factorial ANOVAs were run on the MOS results, illustrated in Table V . The two factors were the TTS system (8 levels) and the Sentence (60 levels) or the Subject (48 levels). All factors and their interactions in both ANOVAs had significant effect (p<0.0001).",
"cite_spans": [],
"ref_spans": [
{
"start": 73,
"end": 80,
"text": "Table V",
"ref_id": "TABREF5"
}
],
"eq_spans": [],
"section": "C. Analysis and Discussion",
"sec_num": null
},
{
"text": "The TTS system factor alone explained an important part of the variance over levels of both Sentence (29%) and Subject factors (30%). The Sentence factor explained only about 8% of the variance (partial 2 = 0.08) while the Subject did 19% (partial 2 = 0.19). The interaction between the System and Sentence or Subject explained a quite important part of the variance, i.e. 21% and 14% respectively. We did observe the sentences with bad scores and found that they were long sentences or had a number of loanwords. Synthetic utterances having consecutive loanwords are extremely bad intelligible. These problems led to bad scores for both Intelligibility and MOS Test.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "C. Analysis and Discussion",
"sec_num": null
},
{
"text": "We did some valuable experiments on TTS systems from different participants using a common dataset in the TTS shared task in the VLSP Campaign 2020. Participants had to validate a piece of training data before receiving the common dataset. There are 7,770 utterances of a female Southwest professional speaker (about 9.5 hours) in the released training dataset. Although using state-of-the-art synthesis techniques that lead to a high-quality synthetic voice, there were still some remaining problems in the results of participants. The best synthetic voice with Tacotron 2 and Hifigan vocoder with Waveglow denoiser achieved 89.3% compared to the human voice, i.e. 3.77 over 4.22 point on a 5-point MOS scale). Some reasons were found for a quite-big gap between the best synthetic voice with state-of-the-art synthetic techniques and the human voice: (i) improper prosodic phrasing for long sentences and (ii) wrong/bad pronunciation for loan words. For the next speech synthesis task of the VLSP Campaign in 2021, we may have more advanced topics for Vietnamese speech synthesis, such as speaker adaptation or expressive speech synthesis.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "VI. CONCLUSIONS",
"sec_num": null
}
],
"back_matter": [
{
"text": "The VLSP 2020 TTS shared task was mainly supported by the R&D Lab, Vbee Services and Data Processing Solution Jsc, and School of Information and Communication Technology. They supported this shared task in developing, deploying, and conducting the online evaluation, based on perception tests as well as building the dataset for the challenge. This task was funded by the Vingroup Innovation Foundation (VINIF) under the project code DA116_14062019 / year 2019. We would like to thank Vais Jsc. for their ASR in building the dataset, and last but not least, the subjects who gave time and effort for the experiments.",
"cite_spans": [],
"ref_spans": [],
"eq_spans": [],
"section": "ACKNOWLEDGMENT",
"sec_num": null
}
],
"bib_entries": {
"BIBREF0": {
"ref_id": "b0",
"title": "Experimental Designs",
"authors": [
{
"first": "Cochran",
"middle": [],
"last": "William",
"suffix": ""
},
{
"first": "G",
"middle": [],
"last": "Cox Gertrude",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "",
"suffix": ""
}
],
"year": 1992,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Cochran William G. and Cox Gertrude M. \"Experimental Designs, 2nd Edition\". Wiley, 2 edition, April 1992. ISBN 0471545678.",
"links": null
},
"BIBREF1": {
"ref_id": "b1",
"title": "Special Issue in VLSP",
"authors": [
{
"first": "Mai",
"middle": [],
"last": "Luong Chi",
"suffix": ""
}
],
"year": 2018,
"venue": "Journal of Computer Science and Cybernetics",
"volume": "",
"issue": "4",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Luong Chi Mai. \"Special Issue in VLSP 2018\". Journal of Computer Science and Cybernetics, V.34, N.4 (2018).",
"links": null
},
"BIBREF2": {
"ref_id": "b2",
"title": "2018 IEEE International Conference on Acoustics, Speech and Signal Processing",
"authors": [
{
"first": "J",
"middle": [],
"last": "Shen",
"suffix": ""
},
{
"first": "R",
"middle": [],
"last": "Pang",
"suffix": ""
},
{
"first": "R",
"middle": [
"J"
],
"last": "Weiss",
"suffix": ""
}
],
"year": 2017,
"venue": "",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Shen, J., Pang, R., Weiss, R.J., et al. 2017. \"Natural TTS Synthesis by Conditioning WaveNet on Mel Spectrogram Predictions\". 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2018.",
"links": null
},
"BIBREF3": {
"ref_id": "b3",
"title": "Statistical Parametric Speech Synthesis using Deep Neural Networks\" on 2013 IEEE International Conference on Acoustics, Speech and Signal Processing",
"authors": [
{
"first": "H",
"middle": [],
"last": "Ze",
"suffix": ""
},
{
"first": "A",
"middle": [],
"last": "Senior",
"suffix": ""
},
{
"first": "M",
"middle": [],
"last": "Schuster",
"suffix": ""
}
],
"year": 2013,
"venue": "",
"volume": "",
"issue": "",
"pages": "7962--7966",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "H. Ze, A. Senior, and M. Schuster, \"Statistical Parametric Speech Synthesis using Deep Neural Networks\" on 2013 IEEE International Conference on Acoustics, Speech and Signal Processing 2013, pp. 7962-7966. IEEE.",
"links": null
},
"BIBREF4": {
"ref_id": "b4",
"title": "Prosodic Phrasing Modeling for Vietnamese TTS using Syntactic Information",
"authors": [
{
"first": "Thu",
"middle": [],
"last": "Nguyen Thi",
"suffix": ""
},
{
"first": "Albert",
"middle": [],
"last": "Trang",
"suffix": ""
},
{
"first": "Tran",
"middle": [
"Do"
],
"last": "Rilliard",
"suffix": ""
},
{
"first": "Christophe",
"middle": [],
"last": "Dat",
"suffix": ""
}
],
"year": 2014,
"venue": "15th Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nguyen Thi Thu Trang, Albert Rilliard, Tran Do Dat, and Christophe d'Alessandro, \"Prosodic Phrasing Modeling for Vietnamese TTS using Syntactic Information\" in 15th Annual Conference of the International Speech Communication Association. Singapore. 2014.",
"links": null
},
"BIBREF5": {
"ref_id": "b5",
"title": "HMM-based TTS for Hanoi Vietnamese: Issues in Design and Evaluation",
"authors": [
{
"first": "Thu",
"middle": [],
"last": "Nguyen Thi",
"suffix": ""
},
{
"first": "Alessandro",
"middle": [],
"last": "Trang",
"suffix": ""
},
{
"first": "Rilliard",
"middle": [],
"last": "Christophe",
"suffix": ""
},
{
"first": "Tran Do",
"middle": [],
"last": "Albert",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Dat",
"suffix": ""
}
],
"year": 2013,
"venue": "14th Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "2311--2315",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nguyen Thi Thu Trang, Alessandro Christophe, Rilliard Albert, and Tran Do Dat. \"HMM-based TTS for Hanoi Vietnamese: Issues in Design and Evaluation\" in 14th Annual Conference of the International Speech Communication Association (Interspeech 2013), pages 2311- 2315. Lyon, France, August 2013b. ISCA.",
"links": null
},
"BIBREF6": {
"ref_id": "b6",
"title": "A method for Vietnamese Text Normalization to Improve the Quality of Speech Synthesis",
"authors": [
{
"first": "Thu",
"middle": [],
"last": "Nguyen Thi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Trang",
"suffix": ""
},
{
"first": "Thi",
"middle": [],
"last": "Pham",
"suffix": ""
},
{
"first": "Tran Do",
"middle": [],
"last": "Thanh",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Dat",
"suffix": ""
}
],
"year": 2010,
"venue": "Proceedings of the 2010 Symposium on Information and Communication",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nguyen Thi Thu Trang, Pham Thi Thanh, and Tran Do Dat. \"A method for Vietnamese Text Normalization to Improve the Quality of Speech Synthesis\" in Proceedings of the 2010 Symposium on Information and Communication (SoICT 2010), Hanoi, Vietnam. 2010.",
"links": null
},
"BIBREF7": {
"ref_id": "b7",
"title": "A Hybrid Method for Vietnamese Text Normalization",
"authors": [
{
"first": "Thu",
"middle": [],
"last": "Nguyen Thi",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Trang",
"suffix": ""
},
{
"first": "Xuan",
"middle": [],
"last": "Dang",
"suffix": ""
},
{
"first": "Nguyen Xuan",
"middle": [],
"last": "Bach",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Tung",
"suffix": ""
}
],
"year": 2019,
"venue": "Proceedings of the 2019 3rd International Conference on Natural Language Processing and Information Retrieval",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nguyen Thi Thu Trang, Dang Xuan Bach, and Nguyen Xuan Tung. \"A Hybrid Method for Vietnamese Text Normalization\" in Proceedings of the 2019 3rd International Conference on Natural Language Processing and Information Retrieval (NLPIR 2019). Japan. 2019.",
"links": null
},
"BIBREF8": {
"ref_id": "b8",
"title": "Development of Vietnamese Speech Synthesis System using Deep Neural Networks",
"authors": [
{
"first": "Nguyen Quoc",
"middle": [],
"last": "Nguyen Van Thinh",
"suffix": ""
},
{
"first": "Phan",
"middle": [],
"last": "Bao",
"suffix": ""
},
{
"first": "Do",
"middle": [],
"last": "Huy Kinh",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Van Hai",
"suffix": ""
}
],
"year": 2018,
"venue": "Journal of Computer Science and Cybernetics",
"volume": "",
"issue": "4",
"pages": "349--363",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Nguyen Van Thinh, Nguyen Quoc Bao, Phan Huy Kinh, Do Van Hai, Development of Vietnamese Speech Synthesis System using Deep Neural Networks, Journal of Computer Science and Cybernetics, V.34, N.4 (2018), 349-363.",
"links": null
},
"BIBREF9": {
"ref_id": "b9",
"title": "An HMM-based Vietnamese Speech Synthesis System",
"authors": [
{
"first": "Luong",
"middle": [],
"last": "Vu Thang Tat",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Mai Chi",
"suffix": ""
},
{
"first": "S",
"middle": [],
"last": "Nakamura",
"suffix": ""
}
],
"year": 2009,
"venue": "Proceedings of the Oriental COCOSDA International Conference on Speech Database and Assessments",
"volume": "",
"issue": "",
"pages": "116--121",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Vu Thang Tat, Luong Mai Chi, and Nakamura S. \"An HMM-based Vietnamese Speech Synthesis System\" in Proceedings of the Oriental COCOSDA International Conference on Speech Database and Assessments, pages 116-121, Beijing, China, 2009.",
"links": null
},
"BIBREF10": {
"ref_id": "b10",
"title": "Tacotron:Towards End-to-end Speech Synthesis",
"authors": [
{
"first": "Yuxuan",
"middle": [],
"last": "Wang",
"suffix": ""
},
{
"first": "Daisy",
"middle": [],
"last": "Skerry-Ryan",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Stanton",
"suffix": ""
},
{
"first": "Ron",
"middle": [
"J"
],
"last": "Yonghuiwu",
"suffix": ""
},
{
"first": "Navdeep",
"middle": [],
"last": "Weiss",
"suffix": ""
},
{
"first": "Zongheng",
"middle": [],
"last": "Jaitly",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Yang",
"suffix": ""
},
{
"first": "Zhifeng",
"middle": [],
"last": "Yingxiao",
"suffix": ""
},
{
"first": "Samy",
"middle": [],
"last": "Chen",
"suffix": ""
},
{
"first": "",
"middle": [],
"last": "Bengio",
"suffix": ""
}
],
"year": 2017,
"venue": "18th Annual Conference of the International Speech Communication Association",
"volume": "",
"issue": "",
"pages": "",
"other_ids": {},
"num": null,
"urls": [],
"raw_text": "Yuxuan Wang, RJ Skerry-Ryan, Daisy Stanton, YonghuiWu, Ron J Weiss, Navdeep Jaitly, Zongheng Yang, YingXiao, Zhifeng Chen, Samy Bengio, et al. \"Tacotron:Towards End-to-end Speech Synthesis\" in 18th Annual Conference of the International Speech Communication Association. Sweden. 2017.",
"links": null
}
},
"ref_entries": {
"FIGREF0": {
"text": "A complete process for participating TTS shared task VLSP 2020.",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF1": {
"text": "Participants in VLSP TTS 2020 and 2019.",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF2": {
"text": "Online Tool for Intelligibility Test.",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF3": {
"text": "Online Tool for MOS Test.",
"uris": null,
"type_str": "figure",
"num": null
},
"FIGREF4": {
"text": "MOS Test Final Results.",
"uris": null,
"type_str": "figure",
"num": null
},
"TABREF0": {
"content": "<table><tr><td>No</td><td>Team ID</td><td>Affiliation</td><td>Submission</td></tr><tr><td>1</td><td>Team1</td><td>Unaffiliated</td><td>API (error)</td></tr><tr><td>2</td><td>Team2</td><td>Smartcall</td><td>API</td></tr><tr><td>3</td><td>Team3</td><td>Unaffiliated</td><td>Docker image (error)</td></tr><tr><td>4</td><td>Team4</td><td>Viettel Telecom</td><td>API</td></tr><tr><td>5</td><td>Team5</td><td>IC</td><td>IC</td></tr><tr><td>6</td><td>Team6</td><td>VAIS</td><td>API</td></tr><tr><td>7</td><td>Team7</td><td>Falcon</td><td>API</td></tr><tr><td>8</td><td>Team8</td><td>Sun Asterisk Inc.</td><td>Docker Image</td></tr><tr><td>9</td><td>Team9</td><td>UET</td><td>Docker Image</td></tr></table>",
"num": null,
"text": "LIST OF TEAMS PARTICIPATING IN VSLP TTS 2020",
"type_str": "table",
"html": null
},
"TABREF2": {
"content": "<table><tr><td>.</td><td colspan=\"2\">DESIGN FOR INTELLIGIBILITY SUB-TESTS</td></tr><tr><td>Sub-test 1</td><td>Sub-test 2</td><td>Sub-test 3</td></tr><tr><td>IntelligibilityTest-1</td><td>IntelligibilityTest-2</td><td>IntelligibilityTest-3</td></tr><tr><td>Team 7</td><td>NATURAL</td><td>NATURAL</td></tr><tr><td>Team 8</td><td>Team 5</td><td>Team 2</td></tr><tr><td>Team 9</td><td>Team 6</td><td>Team 4</td></tr><tr><td colspan=\"3\">There are three sub-tests in the intelligibility test,</td></tr><tr><td colspan=\"3\">following the Latin Square design aforementioned. In each</td></tr><tr><td colspan=\"3\">sub-test, there were 3 voices of 3 different teams with or</td></tr><tr><td colspan=\"3\">without the natural speech reference (NATURAL). Details</td></tr><tr><td colspan=\"3\">for each sub-test is presented in Table II. Each sub-test</td></tr><tr><td colspan=\"3\">included voices of two (sub-test 2 and sub-test 3) or three</td></tr><tr><td colspan=\"3\">teams (sub-test 1). The natural speech was put in both sub-</td></tr><tr><td colspan=\"3\">test 2 and sub-test 3 for more reference. As a result, each sub-</td></tr><tr><td colspan=\"2\">test had a total of 3 voices.</td><td/></tr></table>",
"num": null,
"text": "",
"type_str": "table",
"html": null
},
"TABREF3": {
"content": "<table><tr><td>Sub-test 1</td><td>Sub-test 2</td></tr><tr><td>MOS Test 1</td><td>MOS Test 2</td></tr><tr><td>NATURAL</td><td>NATURAL</td></tr><tr><td>Team 2</td><td>Team 6</td></tr><tr><td>Team 4</td><td>Team 7</td></tr><tr><td>Team 5</td><td>Team 8</td></tr><tr><td/><td>Team 9</td></tr></table>",
"num": null,
"text": "DESIGN FOR MOS TEST SUB-TESTS",
"type_str": "table",
"html": null
},
"TABREF4": {
"content": "<table><tr><td>Testing voice</td><td>MOS Score (5-scale)</td><td>Synthesis Techniques</td></tr><tr><td>NATURAL</td><td>4.220</td><td/></tr><tr><td/><td/><td>\u2022 Acoustic model: Tacotron 2;</td></tr><tr><td>Team2</td><td>3.769</td><td>Vocoder: HiFi-GAN;</td></tr><tr><td/><td/><td>\u2022 Denoiser: Waveglow</td></tr><tr><td/><td/><td>\u2022 Acoustic model: FastSpeech</td></tr><tr><td>Team7</td><td>3.698</td><td>+ PostNet;</td></tr><tr><td/><td/><td>\u2022 Vocoder: Waveglow</td></tr><tr><td>Team6</td><td>3.313</td><td>\u2022 Acoustic model: Tacotron 2; \u2022 Vocoder: Waveglow</td></tr></table>",
"num": null,
"text": "MOS TEST RESULTS WITH SYNTHESIS TECHNIQUES",
"type_str": "table",
"html": null
},
"TABREF5": {
"content": "<table><tr><td>Factor</td><td>df</td><td>df error</td><td>F</td><td>p</td><td/></tr><tr><td>System</td><td>7</td><td>5,688</td><td>335.38</td><td>0.0000</td><td>0.29</td></tr><tr><td>Sentence</td><td>59</td><td>5,688</td><td>8.71</td><td>0.0000</td><td>0.08</td></tr><tr><td>System:Sentence</td><td>412</td><td>5,688</td><td>3.57</td><td>0.0000</td><td>0.21</td></tr><tr><td>System</td><td>7</td><td>5,798</td><td>353.37</td><td>0.0000</td><td>0.30</td></tr><tr><td>Subject</td><td>47</td><td>5,798</td><td>29.29</td><td>0.0000</td><td>0.19</td></tr><tr><td>System:Subject</td><td>314</td><td>5,798</td><td>2.89</td><td>0.0000</td><td>0.14</td></tr></table>",
"num": null,
"text": "ANOVA RESULTS OF MOS TEST",
"type_str": "table",
"html": null
}
}
}
} |