Qwen3-ASR 0.6B Vietnamese Money Slang VieNeu Prod3248 K-Variants
Private research checkpoint adapted for Vietnamese food-price and money-slang ASR.
Training started from runs/qwen3-asr-0.6b-vi-money-slang-vieneu-prod2064, then continued with 999 train examples covering every integer shorthand from 1k through 999k, plus held-out k-variant rows. Audio was generated with VieNeu TTS on CUDA.
Evaluation
| Evaluation set | Model | WER norm | CER norm | Money accuracy |
|---|---|---|---|---|
| kvariants1184 test | Qwen/Qwen3-ASR-0.6B | 66.02% | 60.00% | 7.53% |
| kvariants1184 test | previous prod2064 | 33.98% | 22.74% | 58.06% |
| kvariants1184 test | this checkpoint | 18.28% | 6.51% | 90.32% |
| prod2064 test | previous prod2064 | 9.79% | 6.54% | 87.50% |
| prod2064 test | this checkpoint | 7.72% | 5.65% | 88.13% |
VIVOS regression is pending because no local VIVOS audio/transcripts were mounted on this machine.
Intended Use
This is a private adaptation experiment for recognizing Vietnamese spoken food-price and money slang. It should be validated on real user recordings before production use.