Vietnamese voice-command ASR β fine-tuned VietASR Zipformer (private)
VietASR Zipformer transducer (68M, Apache-2.0, zzasdf/viet_iter3_pseudo_label) fine-tuned with former/finetune.py
(8 epochs, lr 1e-4, torchaudio RNN-T loss). Output: lowercase, no punctuation, numbers as words. The BPE model comes
from the base repo (data/Vietnam_bpe_2000_new/bpe.model). Decode: python -m former.decode --ckpt <folder>/last.pt --manifest M --out x.json --beam 4. Dev = 774 utterances of the 232 dev samples.
| folder | trained on | result | used by |
|---|---|---|---|
ft_tts100 |
100 % train + 6,000 TTS takes (vietnamese-voice-command-tts-audio, config asr) |
leaderboard 70.94 (v7_tts) | best submission |
ft_vasr100 |
100 % train | leaderboard 68 (v7) | v6, v7 |
ft_vasr85 |
85 % | dev WER 6.16 greedy / 5.98 beam 4 | dev experiments |
ft_tts85 |
85 % + TTS | dev 6.02 beam 4; unseen apps 7.26 vs 7.74 | TTS A/B |
ft_fold0..4 |
5-fold cross-fit of all train | 6.54 on held-out folds β train_kfold_zip.json (SFT inputs) |
agent SFT data |
ft_hold, ft_tts_hold, ft_hold_mix |
train minus 7 apps (+ TTS / + general data) | unseen-app WER 7.87 / TTS / mix | held-out tests |
ft_mix85 |
85 % + 12k general Vietnamese | dev 6.26 (no gain) | RESULTS.md |
ft_gip85 |
85 %, initialised from gipformer-68M | dev 6.57 (worse) | ASR checks |
Licence note: ft_tts100, ft_tts85, ft_tts_hold were trained partly on Drama-3 TTS takes, whose licence allows
use for ASR training only: never use these takes (or anything generated with Drama-3) to train the agent / an LLM.
Trained on competition data: keep private.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support
Model tree for baohuynhbk14/vietnamese-voice-command-zipformer
Base model
zzasdf/viet_iter3_pseudo_label