Add LFM Speech Dataset Studio community project

[Hisernberg]

Summary

This PR adds LFM Speech Dataset Studio to the Community Projects section.

The project is a local-first Bangla speech dataset workflow for Liquid-model ecosystem experiments. It audits the LFM-Audio GGUF local runner, evaluates a Bangla ASR baseline, measures Liquid text-model transcript repair potential, applies abstaining evidence routing, and generates schema-valid reviewable action cards.

Project link

https://github.com/Hisernberg/LFM-speech_dataset_studio

Why it fits the cookbook

• Uses Liquid models in a practical local-first workflow.
• Includes LFM-Audio GGUF runner capability auditing.
• Uses Liquid text models for transcript repair-potential analysis.
• Produces reproducible artifacts, examples, tests, and documentation.
• Focuses on low-resource Bangla speech workflows.
• Keeps strict claim discipline instead of reporting unsupported benchmark claims.

Claim discipline

This project does not claim LFM-Audio Bangla ASR quality unless the local runner produces valid transcripts. In the latest Kaggle run, the LFM-Audio runner built successfully but produced no valid transcripts, so the project reports that component only as a capability audit.

BanglaASR is reported only as a baseline speech front-end, not as a Liquid model.

Checklist

• Added one Community Project entry.
• Linked to a public GitHub repo.
• Did not add model weights, datasets, WAV files, or GGUF files.
• Kept the cookbook PR small and reviewable.