AXERA-TECH/punc-ct-transformer

Sherpa Punctuation CT Transformer for AX650 NPU

Character-level punctuation restoration model for Chinese text, compiled as AXMODEL for the AX650 NPU. Based on sherpa-onnx punct CT Transformer architecture.

Model Description

• Architecture: CT Transformer (Connectionist Temporal Classification)
• Input: Raw Chinese text (may include English words)
• Output: Punctuation-annotated text with ， 。 ？ 、
• Vocab Size: 272,727 tokens
• Max Input Length: 64 tokens per window (long text uses sliding window with 4-token overlap)
• Output Classes: 6 — <unk>, _, ，, 。, ？, 、

Quick Start

git clone https://huggingface.co/AXERA-TECH/punc-ct-transformer
cd punc-ct-transformer

Python

# Install dependencies
pip install numpy
git clone https://github.com/AXERA-TECH/pyaxengine.git
cd pyaxengine && pip install . && cd ..

# Run
cd python && python3 example.py

C++

cd cpp && ./demo -m ../model.axmodel -t ../tokens.json

Python SDK API

import sys
sys.path.insert(0, "python")
from sherpa_punct_sdk import PunctuationPipeline

pipeline = PunctuationPipeline("model.axmodel", "tokens.json")
result = pipeline("你好吗how are you我很好谢谢")
print(result)  # 你好吗？how are you我很好，谢谢。

Example Output

Input	Output
你好吗how are you我很好谢谢	你好吗？how are you我很好，谢谢。
今天天气真不错我们出去走走吧	今天天气真不错，我们出去走走吧，
这个方案有三个优点第一成本低第二效率高第三维护简单	这个方案有三个优点，第一，成本低，第二效率高。第三，维护简单，

File Structure

.
├── model.axmodel              # Compiled AXMODEL for AX650 NPU
├── tokens.json                # Vocabulary (272,727 tokens)
├── python/
│   ├── example.py             # Demo script (run from python/ dir)
│   ├── requirements.txt       # Python dependencies
│   └── sherpa_punct_sdk/      # Python inference SDK
│       ├── pipeline.py        # End-to-end text → punctuation pipeline
│       ├── inference.py       # AX Engine inference wrapper
│       ├── preprocess.py      # Character-level tokenizer
│       └── postprocess.py     # Logits → punctuation decoder
└── cpp/
    ├── demo                   # C++ demo (aarch64, run from cpp/ dir)
    └── libsherpa_punct.a      # Static library

Requirements

• Hardware: AX650 NPU (or compatible chip)
• Python: 3.8+
• Python deps: numpy, pyaxengine
• C++: aarch64 Linux with libax_engine.so

License

MIT