Datasets:

1uckyan
/

code-switch_chunks

language:
  - zh
  - en
license: cc-by-nc-4.0
task_categories:
  - automatic-speech-recognition
tags:
  - code-switching
dataset_info:
  config_names:
    - SECoMiCSC
    - DevCECoMiCSC

Robust Code-Switching ASR Benchmark

Dataset Summary

This dataset is a processed and cleaned derivative of the open-source MagicData corpus, specifically optimized for our project Code-Switched ASR robustness (e.g., Whisper fine-tuning).

We addressed the "context fragmentation" issue in original long-form audio by applying a Smart-Merge Strategy (merging short segments into 5-15s chunks using ground-truth timestamps) and filtering out conversational fillers.

Original Data Sources

This dataset is derived from the following open-source datasets released by MagicData Technology:

Training Subset: Derived from ASR-SECoMiCSC
- Source: MagicData Open Source Community
Benchmark/Test Subset: Derived from ASR-DevCECoMiCSC
- Source: MagicData Open Source Community

> Note: This repository contains processed audio chunks and metadata only. Please refer to the original links for full datasets and license details.

Processing Pipeline (Why this version?)

Smart Segmentation: Instead of random VAD cutting, we merged short utterances into 5s - 15s segments based on speaker identity and time gaps. This provides better context for Transformer-based models.
Noise Filtering: Removed pure filler segments (e.g., "嗯", "啊", "[ENS]") to reduce hallucination during training.

Usage

from datasets import load_dataset

# 1. Load Training Data (SECoMiCSC)
dataset_train = load_dataset("1uckyan/code-switch_chunks", data_dir="SECoMiCSC", split="train")

# 2. Load Benchmark Test Set (DevCECoMiCSC)
dataset_test = load_dataset("1uckyan/code-switch_chunks", data_dir="DevCECoMiCSC", split="train")