Mnji4/DialectSeed

A community-built speech corpus for preserving local voices and building low-resource ASR/TTS. # DialectSeed **An open-source pipeline for collecting regional speech and building low-resource ASR/TTS systems.** Commercial speech systems tend to prioritize languages and accents with large user bases, abundant data, and clear deployment demand. Regional dialects often have none of those advantages. The result is not only cultural loss: people who rely on local speech can also be excluded from modern digital services. This is personal for me. My grandmother cannot type and does not speak standard Mandarin comfortably. A voice interface should be the most natural way for her to use a digital service, yet the speech technology around her is built for a language variety she does not normally use. DialectSeed started as a small project to record my hometown dialect, Yangjiang Cantonese. It is now being generalized into an open-source, end-to-end workflow that other communities can reuse for their own underrepresented local languages and speech varieties. The goal is simple: **make it cheap and reproducible to go from community recordings to a clean, consented dataset and then to a small dialect-adapted speech model.** > The training recipes are implemented, but the project does not yet claim a finished production-quality dialect ASR or TTS model. Model experiments are still in active development. ## The pipeline ```text Community contributors | v Browser recording + transcript correction | v Consent + anonymous speaker metadata | v Moderation and quality control | v D1 metadata + R2 audio storage | v Speaker-disjoint dataset export | +-----------------------+ | | v v Qwen3-ASR-0.6B VoxCPM1.5 ASR fine-tuning TTS LoRA | | +-----------+-----------+ | v Local speech technology ``` Yangjiang Cantonese remains the first seed variety, but it is only an example. The data model and collection workflow are designed to support many dialects, local languages, and speech varieties in the …