Datasets:
AlienKevin
commited on
Commit
•
b71b2e5
1
Parent(s):
d979429
Update README.md
Browse files
README.md
CHANGED
@@ -13,4 +13,53 @@ tags:
|
|
13 |
pretty_name: Cantone
|
14 |
size_categories:
|
15 |
- 10K<n<100K
|
16 |
-
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
13 |
pretty_name: Cantone
|
14 |
size_categories:
|
15 |
- 10K<n<100K
|
16 |
+
---
|
17 |
+
|
18 |
+
# Cantone
|
19 |
+
|
20 |
+
A dataset of 34,489 recordings of Cantonese syllables by 10 speakers.
|
21 |
+
Those syllables are generated through the Cantonese speech synthesis engines of Amazon, Apple, Google, and Microsoft.
|
22 |
+
|
23 |
+
All recordings are stored as WAV files with the following format
|
24 |
+
* Channel: mono
|
25 |
+
* Sample rate: 16 kHz
|
26 |
+
* Bits per sample: 16
|
27 |
+
|
28 |
+
Here's a breakdown of the number of recordings under each speaker:
|
29 |
+
| Company | Speaker | # Syllables |
|
30 |
+
| --------|-------- | -------- |
|
31 |
+
| Amazon | Hiujin | 3,885 |
|
32 |
+
| Apple | Aasing | 2,977 |
|
33 |
+
| Apple | Sinji | 2,977 |
|
34 |
+
| Google | A | 3,653 |
|
35 |
+
| Google | B | 3,653 |
|
36 |
+
| Google | C | 3,653 |
|
37 |
+
| Google | D | 3,653 |
|
38 |
+
| Microsoft | Hiugaai | 3,349 |
|
39 |
+
| Microsoft | Hiumaan | 3,349 |
|
40 |
+
| Microsoft | Wanlung | 3,349 |
|
41 |
+
|
42 |
+
## Dataset Construction
|
43 |
+
|
44 |
+
1. Gathering
|
45 |
+
We first identified 3,904 common Cantonese syllables based on words.hk's syllable recordings.
|
46 |
+
The, we ask the speech synthesis APIs to pronounce each of the syllables.
|
47 |
+
The queries use SSML's phoneme attribute to precisely specify the syllable we want. Here's a sample SSML query that fetches the syllable jyut6:
|
48 |
+
```xml
|
49 |
+
<speak><phoneme alphabet='jyutping' ph='jyut6'></phoneme></speak>
|
50 |
+
```
|
51 |
+
|
52 |
+
Apple voices are gathered using jyutping text directly and a native Cantonese ASR system is used to filter out unsupported syllables.
|
53 |
+
|
54 |
+
2. Preprocessing
|
55 |
+
* All audios are converted to 16kHz WAV files
|
56 |
+
* Peak normalize all audios to -20 dBFS
|
57 |
+
* Clip silence at the beginning and end (sound below -50 dBFS are deemed silence)
|
58 |
+
|
59 |
+
3. Verification
|
60 |
+
Occassionally, some syllables are not synthesized correctly.
|
61 |
+
* Apple voices usually renders tone 5 syllables as tone 2: we remove all tone 5 syllables from apple voices
|
62 |
+
* Microsoft voices prepends consonants like ng, g, and b in front of isolate vowel syllables like aa: we remove all vowel syllables from microsoft voices
|
63 |
+
|
64 |
+
## License
|
65 |
+
MIT
|