Offline Filipino Speech on iOS with Quantized VoxCPM2
Year: 2026
Type: preprint
Abstract: QnVoice bundles quantized VoxCPM2 for Filipino text-to-speech on iOS. In a matched 20-prompt desktop comparison with the app’s earlier SpeechT5 system, Whisper’s character error rate fell from 15.40% to 8.54%, while word error stayed above 50% for both systems. On an iPhone 15 Pro, median time to playback readiness for five warm requests fell from 7.87 s to 2.48 s after changes to the vocoder and reference encoding. The optimized runtime took 13.63 s on the first request after the synthesis session was released. In a separate Mac comparison, the vocoder change preserved the sample count for each of six paired renders, with a maximum per-clip root mean square difference of 1.36 × 10⁻⁵ of full scale. Perceived naturalness remains a question for Filipino listeners.
Author: Asis, Ona Pacholo
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225419491