Downloads · 30 days
17
53% of all-time downloads
ArikaisAllie/yueyu-whisper-small-lora-v1
yueyu-whisper-small-lora-v1 is a automatic speech recognition model from ArikaisAllie. Use it when you need speech turned into text. It is set up for peft. The card lists the license as other.
Sole author: Alice Chen (Chen Jiachen / Jiachen Chen)
Downloads · 30 days
17
53% of all-time downloads
All-time downloads
32
Public
Repo size
85.6 MB
Likes
1
Public
Click a slice to open those files.
.pt57.1 MB · 59%
From the Hugging Face model README
Sole author: Alice Chen (Chen Jiachen / Jiachen Chen)
LoRA adapter for Yueyu / Yue opera (越语·越剧) speech recognition, fine-tuned from openai/whisper-small.
Part of the public project Yueyu Detecting (越语侦听), built and maintained solely by Alice Chen.
Held-out test set: 27 clips, whole-piece split (祥林嫂, 何文秀).
| Model | Mean raw CER |
|---|---|
openai/whisper-tiny | 7.3 |
openai/whisper-small (no LoRA) | 10.7 |
longer local LoRA v2 on whisper-small | 1.1 |
| Δ (LoRA − plain small) | −9.6 |
JSON: repo data/corpus/asr/runs/cer-v2b.json. Exact match still 0% on this holdout.
| Model | Mean CER |
|---|---|
openai/whisper-tiny (site baseline) | 42.9 |
this adapter on whisper-small | 18.7 |
On the later size-matched decode, the same v1 adapter scores raw mean CER 5.55 (cer-v2.json).
Caveats: small gold set (~18.4 min total); Speak on the live site still uses a generic baseline; not production Yueyu ASR. See repo docs/TECHNICAL_REPORT.md.
from transformers import WhisperForConditionalGeneration, WhisperProcessor
from peft import PeftModel
base = "openai/whisper-small"
adapter = "ArikaisAllie/yueyu-whisper-small-lora-v1"
processor = WhisperProcessor.from_pretrained(base, language="chinese", task="transcribe")
model = WhisperForConditionalGeneration.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)
Generative-AI tools were used only as assistants for repetitive, mechanical work (organizing/reformatting materials, wording polish); all data, training decisions, evaluation, and writing are the author's own and were verified against primary sources.
r=16, alpha=32, targets q_proj/v_proj)Author: Alice Chen (Chen Jiachen). Training clips derive from archive starter materials for academic/demo use. Do not claim commercial rights to source performances. Adapter weights released for research and educational reuse with attribution to Alice Chen / Yueyu Detecting.