Downloads · 30 days
8
21% of all-time downloads
banhchungtuongot/phowhisper-medium-vi-telephony-lora
phowhisper-medium-vi-telephony-lora is a automatic speech recognition model from banhchungtuongot. Use it when you need speech turned into text. It is set up for peft. The card lists the license as apache-2.0.
Encoder-only LoRA (decoder frozen) + encoder-consistency loss + low-SNR curriculum on vinai/PhoWhisper-medium, for Vietnamese telephony ASR (8 kHz + µ-law + babble + reverb) while keeping strong clean accuracy.
Downloads · 30 days
8
21% of all-time downloads
All-time downloads
38
Public
Repo size
56.7 MB
Likes
0
Public
Click a slice to open those files.
.safetensors56.7 MB · 100%
From the Hugging Face model README
Encoder-only LoRA (decoder frozen) + encoder-consistency loss + low-SNR curriculum on vinai/PhoWhisper-medium, for Vietnamese telephony ASR (8 kHz + µ-law + babble + reverb) while keeping strong clean accuracy.
| Metric | greedy | beam-5 |
|---|---|---|
| Clean WER | 0.0904 | 0.1018 |
| Telephony WER (all SNR) | 0.4301 | 0.4052 |
Per-SNR telephony (beam-5): 15 dB 0.271 · 10 dB 0.336 · 5 dB 0.380 · 0 dB 0.644. → SNR≥5 averages 0.329 (<0.35); the all-SNR 0.405 is dominated by the 0 dB multitalker-babble clips (near the acoustic floor for any ASR).
vs original Whisper Exp C base: clean 0.197→0.10, telephony 0.742→0.405.
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel
base = WhisperForConditionalGeneration.from_pretrained("vinai/PhoWhisper-medium")
model = PeftModel.from_pretrained(base, "banhchungtuongot/phowhisper-medium-vi-telephony-lora")
proc = WhisperProcessor.from_pretrained("vinai/PhoWhisper-medium")
proc.tokenizer.set_prefix_tokens(language="vi", task="transcribe")
Code: https://github.com/NguyenDucAnforwork/asr-h100-noise-robust