CommonVoice 22 Sidon: Denoised 48 kHz Audio for 137 Languages

Mozilla Common Voice 22.0 audio restored using the Sidon denoising model (sarulab-speech/sidon-v0.1) at 48 kHz. Released by sarulab-speech in October 2025, this collection spans 137 languages processed into 21-second chunks. The data is formatted as WebDataset shards for efficient streaming and large-scale training.

Use Cases

• Training text-to-speech (TTS) systems using the 48 kHz reconstructed audio samples
• Multilingual automatic speech recognition (ASR) across 137 language folders
• Benchmarking speech enhancement models against Sidon-denoised outputs

Strengths

• High-resolution 48 kHz audio reconstruction
• Coverage of 137 distinct languages
• Standardized WebDataset shard format for high-throughput loading

Limitations

• Potential algorithmic artifacts introduced by the Sidon restoration model
• Fixed 21-second chunking may split natural speech segments or sentences
• Inherits any transcription errors present in the original Mozilla Common Voice 22.0 source

Provenance

Source

Mozilla Common Voice 22.0

Collection Method

Algorithmic restoration and denoising of crowdsourced audio using the Sidon-v0.1 model

Freshness

Last updated October 2025.

Geography

Global (137 languages)