Tamil Speech Recognition Corpus With 1000 Hours

Approximately 1000 hours of Tamil audio paired with transcripts. The transcripts have been de-duplicated using exact match deduplication. The dataset was created by parambharat and last updated in December 2022.

Use Cases

• Train an ASR model to generate text transcripts from Tamil audio samples.
• Fine-tune a pre-trained speech model using the 1000 hours of Tamil audio data.
• Benchmark ASR model performance on de-duplicated Tamil transcript accuracy.
• Analyze phonetic patterns in Tamil speech using the audio and text alignment.
• Create a language model for Tamil using the cleaned transcript corpus.

Strengths

• 1000 hours of audio data provides substantial training material
• Transcripts have been processed to remove exact duplicates

Limitations

• Exact deduplication may remove linguistically valid repeated phrases
• No information on speaker count, demographics, or recording conditions

Provenance

Source

huggingface

Collection Method

Extended from OpenSLR sources, de-duplicated transcripts

Freshness

Last updated in December 2022.

Geography

Region listed as 'us', but specific coverage unknown