Ashuza11/mashi-tts-bootstrap

Bootstrapping TTS for Mashi (Bashi/Shi): cross-lingual transfer from Kinyarwanda for an extremely low-resource Bantu language of eastern DRC — first Mashi speech dataset + dual-pipeline segmentation comparison. # mashi-tts-bootstrap **Bootstrapping TTS for Mashi: Exploring Cross-Lingual Transfer from Kinyarwanda for a Low-Resource Bantu Language** Mashi (also called Bashi or Shi, ISO 639-3: `shi`) is spoken by approximately 3 million people in South Kivu, eastern DRC. No TTS, ASR, or NLP tools exist for this language. This project builds the first Mashi speech dataset — 3 speakers, 5 domains, 723 curated clips — and fine-tunes a multilingual TTS model (Coqui XTTS v2) on it as a methodological contribution for low-resource languages. --- ## Research question Can fine-tuning a multilingual TTS model on roughly one hour of Mashi audio produce intelligible Mashi speech? > **Scope note (July 2026):** the originally planned Track A (automatic WhisperX > forced alignment) was dropped for time reasons. The project uses a single > curated dataset built by silence-based splitting at natural pauses, verified > and transcribed by the collector. --- ## Key contribution **Dataset contribution:** The first Mashi speech dataset — 3 speakers, 5 domains, **723 clips / 67 minutes** of segmented, transcribed, TTS-ready audio (from ~10 hours of raw Bible audio plus vocabulary, story, and clock-time recordings). **Method contribution:** a reproducible low-resource curation pipeline — silence-based segmentation at natural speech pauses (scripts 07/09), component-based synthesis of clock-time clips from web audio (script 08), and casing/apostrophe normalization (script 06) — all cutting only at silences so no word is ever split. --- ## The dataset ### Language - **Name:** Mashi (Bashi, Shi) — `shi` (ISO 639-3) - **Region:** South Kivu province, eastern DRC. Ngweshe, Nyangezi, Bunyakiri, Muganzo territories. - **Script:** Latin-based with diacritics (ê, â, î, û) - **Prior resources:** None. This is the first open Mashi speech dataset. ### Dataset collector **Muhigiri Ashuza** (age 28, Mushi from Bukavu, South Kivu, DRC) — native/heritage speaker of Mashi. He recorded, organized, and …