buumba641/Zambia-MultiLingual-ASR-Dataset

🇿🇲 Zambia Multilingual ASR Dataset

A continuously growing and curated multilingual speech corpus for Zambian languages, designed to advance Automatic Speech Recognition (ASR) research through community-driven data collection and real-world evaluation.

---

Overview

The Zambia Multilingual ASR Dataset is an open, continuously evolving speech corpus developed as part of the ZamVoice project.

The dataset supports research and development of Automatic Speech Recognition (ASR) systems for under-resourced Zambian languages by combining benchmark speech recordings with continuously collected real-world speech.

Unlike traditional static datasets, this corpus grows over time through community participation. Speech recordings are collected through the ZamVoice platform, automatically transcribed using language-specific Whisper models, and may optionally be reviewed by contributors through ratings, corrected transcriptions, and written feedback.

---

Research Motivation

Existing speech corpora such as Zambezi Voice have significantly advanced speech technology research for Zambian languages. However, most publicly available datasets consist primarily of carefully recorded read speech collected under controlled conditions.

Speech encountered during real-world deployment differs considerably in terms of:

• Recording devices
• Background noise
• Speaking style
• Accents and dialects
• Speech rate
• Pronunciation
• Recording environments

These differences often reduce the performance of speech recognition systems outside their original training domain.

This dataset addresses this challenge by continuously collecting real-world speech while preserving benchmark-quality reference recordings, enabling research on:

• Domain adaptation
• Model robustness
• Continual learning
• Active learning
• Human-in-the-loop annotation

---

Relationship to Zambezi Voice

The Zambia Multilingual ASR Dataset builds upon the pioneering work of the University of Zambia Speech and Language Research Group through the Zambezi Voice project.

The language-specific Whisper models used by ZamVoice were fine-tuned using the Zambezi Voice corpus before deployment.

This dataset extends that work by introducing continuous community-driven data collection and curation.

The dataset may contain:

• Curated reference recordings originating from the Zambezi Voice corpus (where permitted and appropriately attributed)
• Community-contributed speech recordings collected through ZamVoice
• Automatically generated transcriptions
• Human quality ratings
• Corrected transcriptions
• User feedback

Combining these resources enables researchers to evaluate ASR systems on both controlled benchmark speech and naturally occurring real-world speech.

We gratefully acknowledge the University of Zambia Speech and Language Research Group and the Zambezi Voice project for providing the foundational speech resources that made this work possible.

---

Supported Languages

Current languages include:

• Bemba
• Nyanja
• Tonga

Future releases aim to expand coverage to additional Zambian languages.

---

Dataset Structure

Zambia-Multilingual-ASR-Dataset/

├── audio/
│   ├── *.wav
│   └── ...
│
├── meta_data.csv
│
└── README.md

---

Metadata

Each recording corresponds to one row in metadata.csv.

Column	Description
id	Unique sample identifier
timestamp	Date and time of submission
audio_file	Relative path to the audio recording
language	Spoken language
model	Whisper model used for transcription
prediction	Automatically generated transcription
rating	User quality rating (optional)
corrected_transcript	User-corrected transcription (optional)
comment	User feedback (optional)

If no feedback is provided, the corresponding fields remain empty until future annotation or review.

---

Data Collection Workflow

Every submission follows the workflow below.

Speech Recording
        │
        ▼
Automatic Transcription
        │
        ▼
Audio Saved
        │
        ▼
Metadata Recorded
        │
        ▼
(Optional)
Quality Rating
Corrected Transcript
Feedback Comment
        │
        ▼
Dataset Updated

Each recording contributes to the dataset immediately after transcription, while additional annotations can be added later through user feedback.

---

Dataset Curation

The dataset is periodically reviewed to improve overall quality.

Curation activities may include:

• Removing corrupted recordings
• Removing duplicate recordings
• Removing recordings containing no usable speech
• Correcting metadata
• Updating transcriptions using verified user corrections
• Verifying language labels
• Improving annotation quality

The objective is to maintain a high-quality research corpus while preserving the diversity of naturally occurring speech.

---

Intended Applications

This dataset supports research in:

• Automatic Speech Recognition (ASR)
• Speech Corpus Development
• Low-Resource Speech Processing
• Domain Adaptation
• Continual Learning
• Active Learning
• Model Benchmarking
• Transfer Learning
• African Language Technologies

Acknowledgements

This work builds upon the outstanding efforts of the University of Zambia Speech and Language Research Group and the Zambezi Voice project.

Special appreciation goes to the Zambezi Voice research team for creating the foundational multilingual speech corpus that enabled this work.

The ZamVoice project extends this vision by enabling continuous community participation, dataset curation, and real-world speech collection to support future speech technologies for Zambian languages.

---

License

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license unless otherwise specified for individual subsets or externally sourced recordings.

Users are responsible for ensuring compliance with the licensing terms of any referenced or incorporated third-party data.