Ijwi-ry-Ikirundi-AI/Kirundi_Open_Speech_Dataset

🇧🇮 Kirundi Open Speech & Text Dataset

Building the first large-scale, open-source speech and text dataset for Kirundi

🚀 Get Started • 📊 Dataset • 🎯 Roadmap • 🫱🏿‍🫲🏾 Community

---

🌍 About This Project

Kirundi is spoken by over 12 million people, yet it remains a low-resource language largely ignored by modern AI systems. We're changing that.

This community-driven initiative aims to create the first comprehensive, open-source speech and text dataset for Kirundi, enabling the development of:

🎯 Target Applications

Application	Description	Status
🎙️ Speech-to-Text (ASR)	Transcribe spoken Kirundi into text	🔄 In Progress
🗣️ Text-to-Speech (TTS)	Generate natural-sounding Kirundi speech	📋 Planned
🌐 Machine Translation	Translate between Kirundi and other languages	📋 Planned
🎧 Speech Translation	Direct speech-to-speech translation	📋 Planned

Our Mission: Ikirundi cacu, Ijwi ryacu - Preserve and digitize our language for future generations.

🚀 How to Contribute

This project has two "homes," so your contribution flow depends on what you are adding:

• For Text, Translations, or Code: Use our GitHub Repository.
• For Audio: You must use our Hugging Face Repository.

Every contribution matters! Choose the option that works best for you:

📝 Option 1: Text / Translation Contribution (via GitHub)

🎯 Goal: Provide clean, high-quality Kirundi sentences and translations

✔️ What you are allowed to contribute:

• New Kirundi sentences
• Kirundi → French translations
• Kirundi → English translations
• (Optional) Domain and Source metadata

✔️ Required columns:

• Kirundi_Transcription → REQUIRED
• French_Translation or English_Translation → REQUIRED (at least one)
• Do NOT fill anything in the Audio or Admin columns.

---

🔧 Steps to Contribute

1. Fork & Clone

Fork and clone the dataset repository: https://github.com/Ijwi-ry-Ikirundi-AI/Kirundi_Dataset

2. Quality & Normalization Standards

To ensure high-quality ASR/TTS models, every Kirundi sentence must follow strict normalization rules.

---

Transcription Normalization (Kirundi_Transcription)

Rule	Description	Why It Matters
Full Spelling	No abbreviations, numbers, symbols. (e.g., “4” → “cane”, avoid &, @, etc.)	Models learn sounds, not numeric shortcuts.
Ending Punctuation	Every sentence ends with ., ?, or !	Helps TTS detect sentence boundaries and tone.
Initial Capitalization	Capitalize only the first word or proper nouns	Supports automatic capitalization training.
Diacritics Required	Keep tonal/long-vowel marks (â, ū, é, í, etc.) if the source uses them	Crucial for accurate pronunciation & prosody.
Length Limits	Ideal: 4–25 words (max 30)	Ensures easy recording and strong ASR alignment.

---

3. Add Your Data

Open metadata.csv and fill in only the following:

Column	Required	Notes
Kirundi_Transcription	✔️ Yes	The original Kirundi sentence
French_Translation	✔️ Yes (or English)	Provide high-quality translation
English_Translation	Optional	If omitted, it will be generated by AI
Domain	Optional	Grammar, Conjugation, Culture, etc.
Source	Optional	URL, book, document, etc.

❌ Do NOT modify these columns:

Open metadata.csv and fill in only the following:

Column	Required	Notes
Kirundi_Transcription	✔️ Yes	The original Kirundi sentence
French_Translation	✔️ Yes (or English)	Provide high-quality translation
English_Translation	Optional	If omitted, it will be generated by AI
Domain	Optional	Grammar, Conjugation, Culture, etc.
Source	Optional	URL, book, document, etc.

❌ Do NOT modify these columns:

• file_path
• duration
• speaker_id
• age
• gender
• Machine_Suggestion
• Kirundi_Length

These fields are automatically filled by scripts or by the admin.

---

🧩 Additional Tools Included

You may also use the built-in scraping pipeline:

4. Add Sentences to the Text File

Add new clean Kirundi sentences (one per line) to: kirundi_prompts_scraped.txt

5. Run the Appender Script

Automatically append your sentences to metadata.csv without duplicates:

python scripts/append_to_csv.py

This ensures consistency and keeps the dataset clean.

5. Submit a Pull Request

Commit → push to your fork → open a Pull Request on GitHub.

🎤 Option 2: Record Audio (via Hugging Face - CRITICAL!)

🎯 Goal: Create high-quality Kirundi speech recordings

This is the most important step and MUST be done on Hugging Face.

📋 Step 0: First-Time Setup

1. Go to Hugging Face: Fork the Hugging Face Dataset.
2. Clone Your Fork:

   # Clone your fork from Hugging Face (replace "Your-HF-Username")
   git clone https://huggingface.co/datasets/Your-HF-Username/Kirundi_Open_Speech_Dataset
   cd Kirundi_Open_Speech_Dataset
   

3. Install Git LFS:

   git lfs install
   

   (This is a one-time setup on your computer. You can download Git LFS from git-lfs.github.com)

📋 Step 1: Record & Update

1. Find a Task: Open metadata.csv and find a row with an empty file_path.
2. Record: Record the audio for that kirundi_transcription following our Recording Guidelines.
3. Save: Save your audio file (e.g., rn_0001.wav) into the clips/ folder.
4. Update CSV: Fill in your row in metadata.csv (add the file_path, your speaker_id, age, and gender).

Audio Metadata Normalization

Column	Required Format	Explanation
Speaker_id	Anonymous format: S_01_M	Ensures privacy and dataset consistency.
Age	General ranges (e.g., 20s, 30s, 40s+)	Allows diversity without exposing real age.
Duration	Float number (e.g., 3.15)	Required for proper text–audio alignment.

📋 Step 2: Submit

1. Push to Hugging Face:

   git add .
   git commit -m "Added new audio clip clips/rn_0001.wav"
   git push
   

   Git LFS will automatically upload your audio to Hugging Face LFS storage.
2. Submit PR: Go to your Hugging Face fork and submit a Pull Request on Hugging Face.

⚠️ IMPORTANT :

Les fichiers audio NE doivent PAS être envoyés sur GitHub.
Uniquement sur Hugging Face.

---

🎧 Recording Guidelines

📏 Quality Standards for Audio Recordings

🎯 Recording Best Practices

Aspect	Requirement	Why It Matters
🔇 Environment	Quiet room, no background noise	Ensures clean training data
🎙️ Microphone	Headset mic or smartphone (close to mouth)	Clear audio capture
🗣️ Speaking Style	Natural, clear pronunciation	Realistic speech patterns
📝 Accuracy	Read exactly as written	Maintains text-audio alignment

⚙️ Technical Specifications

Audio Format:
  - Primary: WAV (uncompressed)
  - Alternative: MP3 (high quality)

Settings:
  - Sample Rate: 16kHz or 22.05kHz
  - Channels: Mono (1 channel)
  - Bit Depth: 16-bit
  - Duration: Natural sentence length

🛠️ Recommended Tools

• 🎵 Audacity (Free, cross-platform)
• 📱 ASR Voice Recorder (Android app - excellent for high-quality recordings)
• 📱 Smartphone voice recorder (built-in apps work great)
• 💻 Online recorders (for quick contributions)

---

🔄 Git Workflow (For Maintainers)

🎯 How to Sync Changes Across GitHub and Hugging Face

⚠️ Important: Never push directly to main on GitHub. Always use feature branches and Pull Requests.

📋 One-Time Setup

# Add Hugging Face as a remote (if not already added)
git remote add hf https://huggingface.co/datasets/Ijwi-ry-Ikirundi-AI/Kirundi_Open_Speech_Dataset.git

# Add GitHub as a remote (if not already added)
git remote add origin https://github.com/Ijwi-ry-Ikirundi-AI/Kirundi_Dataset

# Verify your remotes
git remote -v

# Install Git LFS (Mac with Homebrew)
brew install git-lfs

# Initialize LFS in your user account
git lfs install

📤 Standard Workflow: GitHub (via PR) → Hugging Face

# 1. Create a feature branch
git checkout -b feature/your-feature-name

# 2. Make your changes, then stage and commit
git add .
git commit -m "Your descriptive commit message"

# 3. Push to GitHub (creates branch on remote)
git push origin feature/your-feature-name

# 4. Go to GitHub and create a Pull Request → Admin will Merge to main

# 5. After PR is merged, pull main locally
git checkout main
git pull origin main

# 6. Push to Hugging Face
git push hf main

🔧 How Git LFS Works

Git LFS (Large File Storage) keeps your repo fast by storing pointer files (tiny text files) in Git, while the actual large files live on a separate server.

┌─────────────────────────────────────────────────────────────┐
│                     YOUR LOCAL REPO                          │
│  clips/proverbs/krd_000089_proverbs.wav (actual 182KB file) │
└─────────────────────────────────────────────────────────────┘
                          │
            ┌─────────────┴─────────────┐
            ▼                           ▼
┌───────────────────────┐   ┌───────────────────────┐
│       GITHUB          │   │    HUGGING FACE       │
├───────────────────────┤   ├───────────────────────┤
│ Pointer file: 131B ✓  │   │ Pointer file: 131B ✓  │
│ LFS Object: 182KB ✓   │   │ LFS Object: 182KB ✓   │
└───────────────────────┘   └───────────────────────┘

When you push to either remote, Git LFS automatically uploads the actual audio files to that remote's LFS server. The Git repo stays small because it only contains tiny pointer files.

---

📊 Dataset Structure

Our dataset is organized around metadata.csv - the central hub for all project data.

📋 Schema Overview

Column	Type	Description	Example
ID	String	Unique identifier	krd_000001_jokes
File_Path	String	Relative path to audio file	clips/jokes/20260131_S01_M_jokes_krd_000001.wav
Kirundi_Transcription	String	Exact Kirundi sentence	Amahoro y'Imana abane nawe
French_Translation	String	High-quality French translation	Que la paix de Dieu soit avec toi
English_Translation	String	High-quality English translation	May the peace of God be with you
Domain	String	Topic or category	jokes, proverbs, grammar
Speaker_id	String	Anonymous speaker identifier	S01_M, S02_M
Age	String	Age group	20s, 30s, 40s+
Gender	String	Self-identified gender	male, female, other
Duration	Float	Audio duration in seconds	3.15
Audio_Status	String	Recording status	pending, recorded, validated

🎙️ Audio Status Tracking

Status	Description
pending	Not recorded yet
recorded	Audio file exists, awaiting peer-review
validated	Peer-review passed, ready for training
rejected	Quality issues, needs re-recording

📁 Directory Structure

Kirundi_Dataset/
├── 📄 README.md                    # This file
├── 📊 metadata.csv                 # Master dataset file
├── 📂 final_dataset_splits/        # Dataset split into 30 parts for recording
│   ├── final_dataset_part_001.csv  # ~100 sentences each
│   ├── final_dataset_part_002.csv
│   └── ... (30 files total)
├── 🎵 clips/                       # Audio recordings (Git LFS → Hugging Face)
│   ├── proverbs/                   # 826 sentences
│   ├── action-verbs/               # 489 sentences
│   ├── grammar/                    # 292 sentences
│   ├── vocabulary/                 # 202 sentences
│   ├── jokes/                      # 50 sentences
│   └── ... (19 domain folders)
└── 🔧 scripts/                     # Automation tools
    ├── append_to_csv.py            # Add new sentences to metadata.csv
    ├── process_audio.py            # Audio processing (VAD + normalize + denoise)
    ├── update_audio_status.py      # Sync audio files with CSV status
    ├── update_file_paths.py        # Update paths in CSVs
    ├── dataset_manager.ipynb       # Data management notebook
    └── kirundi_prompts_scraped.txt # Raw text input

🎙️ Audio Recording Workflow

Daily Recording Steps

1. Record: Record audio for sentences with Audio_Status: pending in final_dataset_splits/
2. Name & Place: Save files as krd_[ID]_[domain].wav in the correct clips/[domain]/ folder
3. Process: Run audio processing script <process_audio.py> (applies VAD, normalization, denoising)
4. Update CSV: Run status update script <update_audio_status.py> to sync audio files with CSV metadata
5. Commit: Create feature branch, commit changes, push to GitHub
6. PR & Merge: Create Pull Request on GitHub, get it merged
7. Sync HF: Pull main, push to Hugging Face

Quick Reference Commands

# After recording audio files (e.g., 50 proverbs)
source .venv/bin/activate

# Step 1: Process the audio (in place)
python scripts/process_audio.py clips/proverbs/

# Step 2: Update CSV metadata
python scripts/update_audio_status.py

# Step 3: Create branch and commit
git checkout -b audio/proverbs-batch-1
git add .
git commit -m "Add 50 proverbs recordings - Speaker S01_M"

# Step 4: Push to GitHub (for PR)
git push origin audio/proverbs-batch-1

# Step 5: After PR is merged on GitHub
git checkout main
git pull origin main

# Step 6: Push to Hugging Face
git push hf main

🔧 Audio Processing Scripts

# Process single audio file (overwrites in place)
python scripts/process_audio.py clips/proverbs/krd_000090_proverbs.wav

# Process entire domain folder
python scripts/process_audio.py clips/proverbs/

# Process ALL clips at once
python scripts/process_audio.py clips/

# Update CSV status from clips/ folder
python scripts/update_audio_status.py

🛠️ Useful Commands

Task	Command
Check what's changed	git status
See pending recordings	Check CSV for Audio_Status = pending
Verify LFS tracking	git lfs ls-files
Check LFS file sizes	git lfs status
List all remotes	git remote -v

📏 Audio Naming Convention

Path format: clips/[DOMAIN]/[DATE]_[SPEAKER]_[DOMAIN]_[SENTENCE_ID].wav
Example: clips/jokes/20260131_S01_M_jokes_krd_000001.wav

Speaker IDs:
- S01_M: César (Male)
- S02_M: Arsène (Male)

Domain folders (19 total):
proverbs, action-verbs, grammar, vocabulary, general, emotions,
jokes, adjectives, idioms, language, geography, food, greetings,
politeness, location, apologies, advice, time, adverbs

🎯 Project Roadmap

🚀 Building the Future of Kirundi AI, One Step at a Time

📈 Development Phases

graph LR
    A[📝 Text Collection] --> B[🌐 Translation]
    B --> C[🎤 Audio Recording]
    C --> D[🤖 Model Training]
    D --> E[🚀 Deployment]

Phase	Goal	Target	Status	Progress
Phase 1	📝 Text Collection	10,000+ sentences	🔄 In Progress
Phase 2	🌐 Translation	Complete dataset	✅ Mostly Done
Phase 3	🎤 Audio Recording	20+ hours	🔄 In Progress
Phase 4	🤖 Model Training	ASR/TTS/MT models	📋 Planned
Phase 5	🚀 Public Release	Open-source models	🎯 Future

🎯 Current Milestones

• ✅ Repository Setup: Project structure and automation scripts
• ✅ Data Pipeline: Automated text processing and validation
• ✅ PDF Cleanup: Removed binary files from git history
• ✅ Dataset Split: 2,903 sentences split into 30 files for recording
• ✅ Audio Infrastructure: Processing scripts (VAD, normalization, denoising)
• ✅ Status Tracking: Audio_Status column for peer-review workflow
• 🔄 Audio Recording: Recording Kirundi speech (101/2,903 complete)
• 📋 Model Training: Fine-tuning Whisper for Kirundi ASR

---

🫱🏿‍🫲🏾 Community

Join our mission to preserve and digitize Kirundi!

🌟 Contributors

We thank all our amazing contributors who are helping build this dataset:

📊 Current Stats

Metric	Count
📝 Total Sentences	2,903
🎤 Audio Recordings	101
⏳ Pending Recording	2,802
👥 Contributors	3

---

⚖️ License

Component	License	Details
🔧 Code & Scripts		Free to use, modify, and distribute
📊 Dataset		Open access with attribution

📜 What This Means

• ✅ Free to use for research, commercial, and educational purposes
• ✅ Attribution required when using the dataset
• ✅ No restrictions on derivative works
• ✅ Open source forever

---

💝 Support This Project

🇧🇮 Ikirundi cacu, Ijwi ryacu 🇧🇮

Our language, our voice - Thank you for helping build the future of Kirundi AI

---

⭐ Star this repository • 🔄 Share with friends • 🫱🏿‍🫲🏾Contribute today