Cantone

A dataset of 34,489 recordings of Cantonese syllables by 10 speakers. Those syllables are generated through the Cantonese speech synthesis engines of Amazon, Apple, Google, and Microsoft.

All recordings are stored as WAV files with the following format

• Channel: mono
• Sample rate: 16 kHz
• Bits per sample: 16

Here's a breakdown of the number of recordings under each speaker:

Company	Speaker	# Syllables
Amazon	Hiujin	3,885
Apple	Aasing	2,977
Apple	Sinji	2,977
Google	A	3,653
Google	B	3,653
Google	C	3,653
Google	D	3,653
Microsoft	Hiugaai	3,349
Microsoft	Hiumaan	3,349
Microsoft	Wanlung	3,349

Dataset Construction

1. Gathering

We first identified 3,904 common Cantonese syllables based on words.hk's syllable recordings. The, we ask the speech synthesis APIs to pronounce each of the syllables. 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:

<speak><phoneme alphabet='jyutping' ph='jyut6'></phoneme></speak>

Apple voices are gathered using jyutping text directly and a native Cantonese ASR system is used to filter out unsupported syllables.

2. Preprocessing

• All audios are converted to 16kHz WAV files
• Peak normalize all audios to -20 dBFS
• Clip silence at the beginning and end (sound below -50 dBFS are deemed silence)

3. Verification

Occassionally, some syllables are not synthesized correctly.

• Apple voices usually renders tone 5 syllables as tone 2: we remove all tone 5 syllables from apple voices
• 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

License

MIT