How QnVoice synthesizes Filipino, Cebuano, and Urdu speech offline on iOS. The preprints document the models, measurements on iPhone, and what still needs testing.
Filipino
A quantized VoxCPM2 pipeline with a synthetic Tagalog reference. The paper separates text normalization, speech quality checks, and the changes that reduced playback delay on iPhone.
2.48 s median warm playback readiness on iPhone 15 Pro, down from 7.87 s across five requests.
Cebuano
A Cebuano voice profile that reuses the quantized VoxCPM2 weights. The study compares synthetic voice references and evaluates typed Cebuano synthesis separately from English-to-Cebuano translation.
3.54 s median time to playback for the five later requests on iPhone 15 Pro, out of six total.
Urdu
An Urdu Matcha model shortened from ten solver steps to three, with a learned correction to recover acoustic features. The study compares synthesis latency on iPhone and Mac.
1.30 s median warm synthesis on iPhone 15 Pro, down from 1.50 s across 72 runs per model.
The studies measure different stages of speech generation. Their timings are not a ranking across languages, and native-speaker listening evaluations remain open.