Brazilian Portuguese Speech Recognition Corpus with 290 Hours

CORAA v1.1 contains 290.77 hours of Brazilian Portuguese audio with transcriptions, segmented into over 400,000 audio files. The dataset is compiled from five distinct speech projects, including academic recordings and TEDx talks, and is validated for automatic speech recognition research.

Use Cases

• Train an automatic speech recognition model on 290.77 hours of Brazilian Portuguese audio with transcriptions.
• Fine-tune a speech-to-text model using segmented audio files from diverse sources like TEDx talks and academic projects.
• Analyze linguistic patterns or dialectal variations across the five constituent speech projects within the corpus.
• Benchmark ASR system performance on validated Brazilian Portuguese audio segments.

Strengths

• Substantial volume of 290.77 hours of validated speech data.
• Diverse audio sources from five distinct Brazilian Portuguese speech projects.
• Over 400,000 segmented audio files for granular analysis.

Limitations

• Limited to Brazilian Portuguese, not generalizable to other Portuguese dialects or languages.
• Potential for audio quality and recording condition variance across the five source projects.
• Dataset composition details and validation specifics require external reference to the full description.

Provenance

Source

Compilation of five projects: ALIP, C-ORAL Brazil, NURC-Recife, SP-2010, and TEDx talks in Portuguese.

Collection Method

Audio collection and transcription from constituent speech projects, with validation.

Freshness

Last updated in December 2022.

Geography

Brazil