Malay Speech Synthesis Dataset With Parliament And News Texts

Featuring approximately 227.7 hours of high-quality Malay speech audio synthesized by the ms-MY-OsmanNeural voice. The audio is sourced from two text corpora: Malay Wikipedia and News articles (94.5 hours) and transcripts from the Malaysian Parliament (133.2 hours). All audio has a 24,000 Hz sample rate and uses sentences ranging from 2 to 20 words.

Use Cases

• Train a text-to-speech model using the 24,000 Hz sample rate audio and corresponding text from Malay Wikipedia and News.
• Fine-tune a speech synthesis model for parliamentary speech styles using the 133.2 hours of Parliament-sourced audio.
• Analyze prosodic patterns in synthesized Malay speech across different text domains (News vs. Parliament).
• Develop a voice cloning model based on the ms-MY-OsmanNeural narrator's voice characteristics from 227.7 hours of data.

Strengths

• Large total duration of approximately 227.7 hours of synthesized speech.
• High-quality audio with a 24,000 Hz sample rate described as 'super clean'.
• Text sourced from two distinct and structured domains: Wikipedia/News and official Parliament transcripts.
• Sentences are curated to be between 2 and 20 words, providing controlled length variation.

Limitations

• All audio is synthesized from a single narrator voice (ms-MY-OsmanNeural), limiting speaker diversity.
• The dataset consists of synthetic speech, not natural human recordings, which may affect model performance for real-world applications.
• The temporal coverage of the source texts (Wikipedia, News, Parliament) is not specified, potentially limiting recency.

Provenance

Source

Texts sourced from Malay Wikipedia, News articles, and transcripts from the Malaysian Parliament.

Collection Method

Speech audio synthesized using Microsoft Azure's ms-MY-OsmanNeural text-to-speech voice.

Time Range

null

Freshness

null

Geography

Malaysia (Malay language focus).