Waxal Lug Clean: Luganda Text-to-Speech Audio with Artifact Removal

CraneAILabs provides a cleaned version of Luganda speech recordings from Google's WaxalNLP dataset, preprocessed for fine-tuning text-to-speech models. The dataset applies Silero VAD to remove click and pop artifacts from the start and end of audio clips, which are described as degrading model quality. This cleaned subset was last updated on March 15, 2026.

Use Cases

• Fine-tuning TTS models based on cleaned Luganda audio recordings.
• Improving speech synthesis quality by training on data with reduced audio artifacts.
• Developing speech technology for Luganda using a preprocessed audio corpus.

Strengths

• Audio has been processed with Silero VAD to remove specific click/pop artifacts described in the original recordings.
• Dataset is explicitly prepared for the task of fine-tuning text-to-speech models.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and license information are unknown, which may limit suitability assessment.

Provenance

Source

Google's WaxalNLP dataset, subset for Luganda TTS.

Collection Method

Preprocessed using Silero VAD for voice activity detection and artifact removal.

Freshness

Last updated 2026-03-15 17:54:53; freshness should be verified.