Offline English-to-Cebuano Translation for QnVoice with a Compact Index-Translate Model
Year: 2026
Type: preprint
Abstract: We adapt a quantized Index-Translate-2B model for offline English-to-Cebuano text translation. Vocabulary pruning from 248,320 to 55,000 entries and a rank-16 conversation adapter yield a 978.6 MB GGUF release. On an Apple M1 Pro Mac, the candidate improves chrF++ over the quantized Helsinki-NLP OPUS-MT baseline on FLORES-200 devtest (52.70 to 55.37, 1,012 pairs) and on 40 QnVoice conversation examples (48.96 to 55.42). It scores lower on 424 OPUS Tatoeba pairs (55.02 to 41.00). The preprint reports paired bootstrap intervals, real translation examples, and reproducibility details. The candidate remains a research release separate from the translator in QnVoice for iOS.
Author: QnVoice
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225558431