---
pretty_name: WTFO Code-Switching Speech
license: other
task_categories:
- automatic-speech-recognition
language:
- zh
- en
tags:
- audio
- speech
- code-switching
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*.parquet
dataset_info:
  features:
  - name: audio
    dtype:
      audio:
        sampling_rate: 16000
  - name: text
    dtype: string
  - name: duration
    dtype: float64
  splits:
  - name: train
    num_bytes: 11550292741
    num_examples: 98662
  download_size: 11517079495
  dataset_size: 11550292741
---

# WTFO Code-Switching Speech

Code-switching speech dataset prepared for automatic speech recognition training.

## Dataset fields

- `audio`: self-contained 16 kHz mono PCM16 FLAC bytes using the Hugging Face `Audio` feature
- `text`: transcript
- `duration`: audio duration in seconds (`float64`)

## Split summary

- Split: `train`
- Examples: 98,662
- Total duration: 567422.698 seconds (157.62 hours)

The Parquet shards embed the audio bytes. They do not depend on the original local file paths.

## JSONL manifest

`metadata/text_audio_pairs.jsonl` contains `text`, `audio_filepath`, `duration`, and `lang`. All rows use `lang: "<|NAN|>"`.