Downloads · 30 days
28
8% of all-time downloads
xCloudinfo/Whisper-Large-v3-xVITA-zhTW
Whisper-Large-v3-xVITA-zhTW is a automatic speech recognition model from xCloudinfo. Use it when you need speech turned into text. The card lists the license as apache-2.0.
云碩科技(xCloudinfo)自研繁體中文(台灣華語)語音辨識模型。以openai/whisper-large-v3為基座微調,原生輸出繁體中文(台灣用字),不需任何繁簡轉換後處理。
Downloads · 30 days
28
8% of all-time downloads
All-time downloads
346
Public
Parameters
1.5B
24 GB on disk
Likes
1
Public
Click a slice to open those files.
.bin8.9 GB · 74%
From the Hugging Face model README
云碩科技(xCloudinfo)自研繁體中文(台灣華語)語音辨識模型。以openai/whisper-large-v3為基座微調,原生輸出繁體中文(台灣用字),不需任何繁簡轉換後處理。
xVITA = Verified Intelligence for Taiwan(可驗證的台灣智慧),為云碩自有模型家族品牌。
不用寫程式,開瀏覽器就能測:xVITA線上試用Space(麥克風錄音或上傳音檔,免費CPU推論)。
generation_config.json 內建純簡體詞元封鎖清單當第二道保險(transformers載入即生效)。ct2/ 目錄,faster-whisper直接載入)。| 模型 | 內容CER(繁簡統一後) | 實際輸出CER(vs繁體標準答案) |
|---|---|---|
| openai/whisper-large-v3(base) | 5.94% | 33.26%(輸出為簡體) |
| Whisper-Large-v3-xVITA-zhTW | 5.57% | 8.60%(原生繁體) |
正規化:僅比對CJK字元、阿拉伯數字轉中文數字後計算。
from transformers import pipeline
pipe = pipeline("automatic-speech-recognition",
model="xCloudinfo/Whisper-Large-v3-xVITA-zhTW",
device=0)
out = pipe("audio.wav", generate_kwargs={"language": "zh", "task": "transcribe"})
print(out["text"]) # 繁體中文
from faster_whisper import WhisperModel
model = WhisperModel("xCloudinfo/Whisper-Large-v3-xVITA-zhTW",
device="cuda", compute_type="float16")
# 或下載後指向 ct2/ 子目錄
segments, info = model.transcribe("audio.wav", language="zh", beam_size=5)
text = "".join(s.text for s in segments)
# 長音檔建議加這行保險層(台灣正體 s2tw,非 s2t):
from opencc import OpenCC
print(OpenCC("s2tw").convert(text))
gguf/ 目錄提供三個量化等級:
| 檔案 | 大小 | 適用 |
|---|---|---|
ggml-xvita-zhtw-q5_0.bin | 1.1 GB | 邊緣裝置、記憶體受限環境 |
ggml-xvita-zhtw-q8_0.bin | 1.7 GB | 品質與大小平衡 |
ggml-xvita-zhtw-f16.bin | 3.1 GB | 最高品質 |
./build/bin/whisper-cli -m ggml-xvita-zhtw-q5_0.bin -l zh -f audio.wav
本模型的核心價值在於云碩自有、自產、可溯源的台灣華語語料,非直接取用網路現成資料集訓練:
云碩科技(xCloudinfo Corp. Limited)專注於地端私有化人工智慧解決方案:地端大型語言模型微調與蒸餾、繁體中文語音辨識、文件辨識與企業知識檢索。xVITA是云碩自研模型的品牌名稱。
其他模型請見 https://hf.co/xCloudinfo。
An xCloudinfo (云碩科技) in-house Traditional Chinese (Taiwanese Mandarin) speech-recognition model. Fine-tuned on openai/whisper-large-v3, it natively outputs Traditional Chinese (Taiwanese usage), with no Simplified-to-Traditional post-processing needed.
xVITA = Verified Intelligence for Taiwan, the brand of xCloudinfo's own model family.
No coding needed, just open a browser to test: xVITA online-trial Space (record from microphone or upload an audio file, free CPU inference).
generation_config.json as a second safeguard (effective as soon as transformers loads it).ct2/ directory, loaded directly by faster-whisper).| Model | Content CER (after unifying Traditional/Simplified) | Actual-output CER (vs. Traditional-Chinese reference) |
|---|---|---|
| openai/whisper-large-v3 (base) | 5.94% | 33.26% (output is Simplified) |
| Whisper-Large-v3-xVITA-zhTW | 5.57% | 8.60% (native Traditional) |
Normalization: computed comparing only CJK characters, after converting Arabic numerals to Chinese numerals.
from transformers import pipeline
pipe = pipeline("automatic-speech-recognition",
model="xCloudinfo/Whisper-Large-v3-xVITA-zhTW",
device=0)
out = pipe("audio.wav", generate_kwargs={"language": "zh", "task": "transcribe"})
print(out["text"]) # 繁體中文
from faster_whisper import WhisperModel
model = WhisperModel("xCloudinfo/Whisper-Large-v3-xVITA-zhTW",
device="cuda", compute_type="float16")
# 或下載後指向 ct2/ 子目錄
segments, info = model.transcribe("audio.wav", language="zh", beam_size=5)
text = "".join(s.text for s in segments)
# 長音檔建議加這行保險層(台灣正體 s2tw,非 s2t):
from opencc import OpenCC
print(OpenCC("s2tw").convert(text))
The gguf/ directory provides three quantization tiers:
| File | Size | Suited to |
|---|---|---|
ggml-xvita-zhtw-q5_0.bin | 1.1 GB | edge devices, memory-constrained environments |
ggml-xvita-zhtw-q8_0.bin | 1.7 GB | balance of quality and size |
ggml-xvita-zhtw-f16.bin | 3.1 GB | highest quality |
./build/bin/whisper-cli -m ggml-xvita-zhtw-q5_0.bin -l zh -f audio.wav
The core value of this model is in xCloudinfo's own, self-produced, traceable Taiwanese-Mandarin corpus, not training that directly uses ready-made datasets from the web:
xCloudinfo Corp. Limited (云碩科技) builds on-premises, private AI solutions: on-premises large language model fine-tuning and distillation, Traditional Chinese speech recognition, document recognition and enterprise knowledge retrieval. xVITA is xCloudinfo's in-house model brand.
More models: https://hf.co/xCloudinfo