CommonVoice22 Sidon Dacvae: Speech Audio Converted to VAE Latents

CommonVoice 22 speech data enhanced by Sidon and converted into DAC VAE latent representations. The dataset is provided by TTS-AGI and was last updated on March 22, 2026. Each sample includes original FLAC audio, a corresponding latent vector, and metadata.

Use Cases

• Training speech synthesis models based on pre-computed DAC VAE latent features.
• Fine-tuning voice conversion systems using the paired audio and latent representations.
• Benchmarking audio representation learning algorithms on a standardized speech dataset.
• Analyzing speech characteristics using the provided metadata and characters-per-second metrics.

Strengths

• Each sample includes three aligned files: original audio, latent representation, and metadata, ensuring data consistency.
• Latent vectors are stored as numpy arrays with a defined shape of [T_latent, 128], providing a structured feature format.
• Data is sharded into tar files of approximately 2GB each, which may facilitate distributed processing.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• The original CommonVoice data may reflect geographic or demographic biases inherent to its crowd-sourced collection.

Provenance

Source

sarulab-speech/commonvoice22_sidon on Hugging Face

Collection Method

Conversion of the CommonVoice 22 (Sidon-enhanced) dataset to DAC VAE latents.

Freshness

Last updated 2026-03-22 13:31:51; freshness should be verified.