jsut-lab

The repository provides HTK/HTS-style alignment files with additional full-context labels for JSUT (Japanese speech corpus of Saruwatari-lab., University of Tokyo) corpus (v1.1). All alignment files (.lab) were extracted by forced-alignment using Julius and full-contexts are generated by OpenJTalk.

The label files are expected to be used for speech reseach; e.g., text-to-speech and voice conversion.

Directory structure is exactly same as the JSUT. You can put the label files to the JSUT data directory if you want:

tree ~/data/jsut_ver1.1/ -d -L 2
/home/ryuichi/data/jsut_ver1.1/
├── basic5000
│   ├── lab
│   └── wav
├── countersuffix26
│   ├── lab
│   └── wav
├── loanword128
│   ├── lab
│   └── wav
├── onomatopee300
│   ├── lab
│   └── wav
├── precedent130
│   ├── lab
│   └── wav
├── repeat500
│   ├── lab
│   └── wav
├── travel1000
│   ├── lab
│   └── wav
├── utparaphrase512
│   ├── lab
│   └── wav
└── voiceactress100
    ├── lab
    └── wav

Label format

Fields: <begin_time> <end_time> <full-context-label>. Time are in 100ns units as same as HTK labels.

$ cat basic5000/lab/BASIC5000_0773.lab | head
 
0 2525000 xx^xx-sil+s=a/A:xx+xx+xx/B:xx-xx_xx/C:xx_xx+xx/D:18+xx_xx/E:xx_xx!xx_xx-xx/F:xx_xx#[email protected]_xx|xx_xx/G:6_3%0_xx_xx/H:xx_xx/I:[email protected]+xx&xx-xx|xx+xx/J:1_6/K:3+6-32
2525000 3825000 xx^sil-s+a=N/A:-2+1+6/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
3825000 4825000 sil^s-a+N=g/A:-2+1+6/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
4825000 5825000 s^a-N+g=i/A:-1+2+5/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
5825000 6125000 a^N-g+i=i/A:0+3+4/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
6125000 7524999 N^g-i+i=N/A:0+3+4/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
7524999 8125000 g^i-i+N=w/A:1+4+3/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
8125000 8425000 i^i-N+w=a/A:2+5+2/B:xx-xx_xx/C:18_xx+xx/D:24+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
8425000 10125000 i^N-w+a=pau/A:3+6+1/B:18-xx_xx/C:24_xx+xx/D:07+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32
10125000 11325000 N^w-a+pau=d/A:3+6+1/B:18-xx_xx/C:24_xx+xx/D:07+xx_xx/E:xx_xx!xx_xx-xx/F:6_3#[email protected]_1|1_6/G:3_1%0_xx_0/H:xx_xx/I:[email protected]+3&1-6|1+32/J:2_10/K:3+6-32

For details, please refer to HTS documents: http://hts.sp.nitech.ac.jp

What can I do with this?

If you want to make traditional DNN-based TTS systems, please check out the tutorials at https://r9y9.github.io/nnmnkwii/latest/. You can use alignment and full-context labels to generate linguistic features.

If you are intersted in end-to-end approaches, please have a look at https://github.com/espnet/espnet. The labels are used at the preprocessing stage for the JSUT recipe (see also https://r9y9.github.io/blog/2017/11/12/jsut_ver1/ to know why we need alignments for end-to-end TTS).

Happy speech hacking!

Source code to generate labels

https://github.com/r9y9/segmentation-kit/tree/jsut2

Notice

• Alignments are likely to have mistakes because they were automatically generated by Julius. Note that they are not hand-annotated labels.

References

• JSUT (Japanese speech corpus of Saruwatari-lab., University of Tokyo)
• HTS
• Julius
• OpenJTalk
• 日本語 End-to-end 音声合成に使えるコーパス JSUT の前処理 [arXiv:1711.00354]
• pyopenjtalk
• nnmnkwii
• sarulab-speech/jsut-label Hand-annotated phonetic and prosodic information from Saruwatari-lab.