---
task_categories:
  - automatic-speech-recognition
language:
  - en
  - pcm
  - tl
  - yo
  - sw
tags:
  - code-switching
  - low-resource
  - conversational
  - speaker-diarization
  - multilingual
pretty_name: Code-Switching Low-Resource Language Speech Dataset
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: train
        path: samples/*
    default: true
dataset_info:
  config_name: default
  features:
    - name: audio
      dtype: audio
    - name: language
      dtype: string
    - name: transcript
      dtype: string
  splits:
    - name: train
      num_examples: 16
---

# Code-Switching Low-Resource Language Speech Dataset

## Overview

This dataset is a collection of spontaneous, conversational speech focused on code-switching across several low-resource languages. The recordings are sourced from the [Yapdo corpus](https://huggingface.co/datasets/liva-ai/yapdo-convo) and include full human-generated transcripts. It is designed for Automatic Speech Recognition (ASR) training and benchmarking, with a particular emphasis on capturing natural, real-world multilingual speech as it actually occurs in everyday conversation.

## Language Pairs

The dataset captures code-switching across the following language combinations:

- **English-Nigerian Pidgin**
- **English-Tagalog**
- **English-Yoruba**
- **English-Swahili**

## Dataset Size

| Language                | Hours    |
| ----------------------- | -------- |
| English-Swahili         | ~616     |
| English-Nigerian Pidgin | ~10      |
| English-Tagalog         | ~10      |
| English-Yoruba          | ~10      |
| **Total**               | **~646** |

## Transcription Process

All transcripts go through a minimum of two people. A native-speaker transcriber works through the audio from scratch using a custom-built transcript editor, producing a full verbatim transcript with precise timestamps down to the millisecond, labeled speaker turns for diarization, and audio event tags for non-speech sounds (e.g. [phone buzzing], [laughing], [door closing]). Native fluency allows transcribers to accurately capture code-switching points and conversational phenomena (overlaps, disfluencies, colloquialisms) that are difficult to transcribe without deep familiarity with each group of speakers.

Once the initial transcription is complete, the transcript is run through an automated format and spelling checker. The transcriber reviews and fixes any detected errors before submitting. This version is then passed to a senior reviewer, someone with a proven track record of high-quality transcripts, who listens to the audio in its entirety and manually corrects any remaining spelling errors or inconsistencies by hand.

Additional hours of transcription are available upon request. All transcription is handled in-house by our native-speaker teams.

## Key Features

- **Spontaneous conversational data**: natural, unscripted dialogue capturing code-switching as it occurs in real conversation
- **Dual-channel audio**: source audio captured on a per-speaker level
- **Human transcription by native speakers**: full verbatim transcripts with millisecond timestamps, speaker diarization, and audio event tags

### Transcript Format Example

```
Speaker 1 [00:00:00.000 - 00:00:01.030]: You don see your balance?

Speaker 1 [00:00:01.630 - 00:00:02.430]: Wetin you get?

Speaker 2 [00:00:15.850 - 00:00:21.660]: I dey see recent, recent, re-recent-- which one be the eh, the particular balance?
```

## Technical Analysis

| Property                 | Value                                      |
| ------------------------ | ------------------------------------------ |
| **Sample rate**          | 48 kHz                                     |
| **Bit depth**            | 16-bit PCM                                 |
| **File format**          | WAV                                        |
| **Mean SNR**             | ~33 dB                                     |
| **Median RMS**           | -26 dBFS                                   |
| **Average speech ratio** | 0.35                                       |
| **Spectral centroid**    | ~0.66 kHz                                  |
| **Frequency content**    | 3.3 kHz (averaged over 10-30 second clips) |