Speech Dataset

How we gather dataset?

1. For semisupervised transcript, we use Google Speech to Text, after that verified / corrected by human.

2. We recorded using our own microphones.

License

Malay-Speech dataset is available to download for research purposes under a Creative Commons Attribution 4.0 International License.

This work is licensed under a Creative Commons Attribution 4.0 International License.

Dataset

Ambient

Simple ambients gathered from Youtube.

Audiobook

Gather Youtube urls for indonesian, english and low quality english audiobooks only.

Azure-TTS

Semisupervised Malay TTS dataset from Azure TTS cloud.

GCP-TTS

Semisupervised Malay TTS dataset from GCP TTS cloud.

Emotion

Speech emotion dataset used by Malaya-Speech for speech emotion detection.

IIUM

Read random sentences from IIUM Confession.

• voice by Husein Zolkepli and Shafiqah Idayu.

• Heavily speaking in Selangor dialect.

• Recorded using low-end tech microphone.

• 44100 sample rate, split by end of sentences.

• approximate 2.4 hours.

• Still on going recording.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Speech Dataset from IIUM Confession texts,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/iium}}
}

IIUM-Clear

Read random sentences from IIUM Confession, cleaner version.

• voice by Husein Zolkepli.

• Heavily speaking in Selangor dialect.

• Recorded using mid-end tech microphone.

• 44100 sample rate, random 7 - 11 words window.

• approximate 0.1 hours.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Speech Dataset from IIUM Confession texts,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/iium}}
}

IMDA

Mirror link for IMDA dataset, https://www.imda.gov.sg/programme-listing/digital-services-lab/national-speech-corpus, only downloaded PART 3 and SST dataset.

• 16000 sample rate.

• supervised approximate 2024 hours.

language

Gather youtube urls for hyperlocal language detection from speech {malay, indonesian, manglish, english, mandarin}.

Check hyperlocal language detection models at https://malaya-speech.readthedocs.io/en/latest/load-language-detection.html

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Hyperlocal languages for speech dataset,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/language}}
}

mixed-stt

Malay, Singlish and Mandarin STT dataset in TFRecord format. Included scripts how to load using torch.dataset.

news

Read random sentences from bahasa news.

• voice by Husein Zolkepli.

• Heavily speaking in Selangor dialect.

• Recorded using mid-end tech microphone, suitable for text to speech.

• 44100 sample rate, random 7 - 11 words window.

• approximate 3.01 hours.

• Still on going recording.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Speech Dataset from local news texts,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/news}}
}

noise

Simple noises gathered from Youtube.

Sebut perkataan

Read random words from malay dictionary started with ‘tolong sebut ‘.

• sebut-perkataan-man voice by Husein Zolkepli

• tolong-sebut voice by Khalil Nooh

• sebut-perkataan-woman voice by Mas Aisyah Ahmad

• Recorded using low-end tech microphones.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Short Speech Dataset,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/sebut-perkataan}}
}

Semisupervised audiobook

Semisupervised malay audiobooks from Nusantara Audiobook using Google Speech to Text.

• 44100 sample rate, super clean.

• semisupervised approximate 45.29 hours.

• windowed using Malaya-Speech VAD, each atleast 5 negative voice activities.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Semisupervised Speech Recognition from Audiobook,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/semisupervised-audiobook}}
}

Semisupervised malay

Semisupervised malay youtube videos using Google Speech to Text, after that corrected by human.

• 16000 sample rate.

• semisupervised approximate 1804 hours.

• random length between 2 - 20 seconds, windowed using google VAD.

• supervised 768 samples, approximate 1.3 hours.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Semisupervised Speech Recognition from Malay Youtube Videos,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/semisupervised-malay}}
}

Semisupervised manglish

Semisupervised manglish youtube videos using Google Speech to Text.

• 16000 sample rate.

• semisupervised approximate 107 hours.

• random length between 2 - 20 seconds, windowed using google VAD.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Semisupervised Speech Recognition from Manglish Youtube Videos,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/semisupervised-manglish}}
}

wattpad

Read random sentences from bahasa wattpad.

• voice by Husein Zolkepli.

• Heavily speaking in Selangor dialect.

• Recorded using mid-end tech microphone, suitable for text to speech.

• 44100 sample rate, random 7 - 11 words window.

• approximate 0.15 hours.

• Still on going recording.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Speech Dataset from Wattpad texts,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/wattpad}}
}

Wikipedia

Read random sentences from Bahasa Wikipedia.

• voice by Husein Zolkepli.

• Heavily speaking in Selangor dialect.

• Recorded using low-end tech microphone.

• 44100 sample rate, 4 words window.

• approximate 3.4 hours.

• Still on going recording.

@misc{Malay-Dataset, We gather Bahasa Malaysia corpus!, Speech Dataset from Wikipedia texts,
  author = {Husein, Zolkepli},
  title = {Malay-Dataset},
  year = {2018},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/huseinzol05/malaya-speech/tree/master/data/wikipedia}}
}

youtube

Semisupervised transcription and Unsupervised Speaker Diarization on 5k malay speakers youtube videos.

Contribution

Contact us at husein.zol05@gmail.com or husein@mesolitica.com if want to contribute to speech bahasa dataset.