Expresso Mimi Codes: Discrete Tokens for Speech Disentanglement Research

Pre-extracted Kyutai Mimi tokens for the Expresso speech dataset. The dataset contains tokens from all 32 codebooks for both read and conversational subsets, intended for training Mimi-based speech models. It was created by shangeth and published on Hugging Face, with a last update timestamp of 2026-05-03 12:54:40.

Use Cases

• Train Mimi-based speech models based on the pre-extracted discrete tokens.
• Research speech disentanglement based on the dataset's relevance for that task.
• Experiment with discrete token representations for speech synthesis or analysis.

Strengths

• Tokens are pre-extracted for all 32 codebooks, saving computational effort.
• Includes both read and conversational subsets of the Expresso dataset.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• License is CC-BY-NC-4.0, restricting commercial use.

Provenance

Source

shangeth

Collection Method

Pre-extracted from the Expresso dataset (shangeth/expresso) using Kyutai Mimi.

Freshness

Last updated 2026-05-03 12:54:40; freshness should be verified.