gladia-normalization open-sources fairer STT evals

This is a small library with outsized practical value: most STT eval pipelines quietly depend on normalization, but few teams make those rules explicit or reproducible. Turning that hidden glue code into a versioned open-source package makes benchmark claims easier to trust.

• –The core pitch is solid because WER really does over-penalize surface-form differences, so normalization often matters as much as the recognizer when teams compare engines.
• –YAML-defined stages and immutable published presets are the right design choice for eval work, where reproducibility matters more than clever heuristics.
• –The three-stage pipeline and CLI make it useful beyond Gladia's own stack; teams can standardize transcript cleanup without rewriting one-off scripts per project.
• –The multilingual angle is promising, but the maintainers themselves flag non-English presets as still needing refinement, so cross-language benchmark users should treat current behavior as a starting point, not ground truth.
• –This also doubles as quiet marketing for Gladia's STT platform: open-sourcing the eval layer helps position the company as credible on speech benchmarking, not just API delivery.