MaprangSTT: On-Device Thai STT model by Cactus Whistle

· November 8, 2026

MaprangSTT (มะปราง) is a fine-tuned Cactus Whistle automatic speech recognition (ASR) model for Standard Thai and Regional Thai Dialects (Central, Isan / Northeastern, Northern / Khummuang, Korat, Southern / Pattani, and Elderly speech).

MaprangSTT model compiles into a standalone 32.49 MB archive ready for fast on-device CPU/GPU/mobile inference.

Key Highlights

  • 6.74% CER / 18.57% WER on Mozilla Common Voice v27 (th) test set (down from 14.01% CER / 31.65% WER on the Common Voice-only model).
  • 13.58% CER / 31.05% WER on Google FLEURS (th_th) test set (1,021 clips) — nearly a 3× reduction in Character Error Rate compared to training on Common Voice alone (37.57% CER).
  • Spoken Thai Number Transcription: Trained with automatic digit/number-to-Thai-words normalization (1967 $\rightarrow$ หนึ่ง พัน เก้า ร้อย หกสิบ เจ็ด).
  • Compact & Edge-Ready: Standalone 32.49 MB W4A8 quantized .cact archive (maprang.cact / MaprangSTT.cact) plus 29.0 MB LoRA + embedding adapter (adapter.safetensors).

Twitter, Facebook