Sartify/Swahili-Challenge-Competition---Pan-African-Wide-Alignment-PAWA-ASR

Swahili Challenge Competition # Tiny-Lugha: Efficient Kiswahili AI Speech Stack Challenge 2025 - Starter Notebooks > **Tiny-Lugha: Efficient Kiswahili AI Speech Stack Challenge** > > Build lightweight STT + LLM pipeline optimized for NVIDIA T4 GPU (≤16GB) ## Challenge Overview This repository provides starter notebooks for the Sartify & ITU AI/ML Kiswahili EdgeVoice Challenge. The goal is to create a voice-to-text AI system that: 1. **Speech-to-Text (STT)**: Transcribe Kiswahili audio with low Word Error Rate (WER) 2. **Language Understanding**: Process text (translate from Swahili to English) through Pawa-Swahili-2B **All running on a single NVIDIA T4 GPU within Google Colab's free tier!** ``` ## Quick Start ``` ### 1. Open in Google Colab ### 2. Run the Notebooks in Order 1. **STT Training** (`asr_starter.ipynb`) 2. **TTS Training** (`tts_starter_notebook.ipynb`) ## Dataset Information ### Speech-to-Text - **Source**: Mozilla Common Voice 17.0 (Swahili) - **Size**: 100+ hours of labeled Kiswahili speech - **License**: CC-0 1.0 ### Language Model - **Model**: Pawa-Swahili-2B - **Parameters**: 2B - **License**: Apache 2.0 ## Model Architecture ### Speech-to-Text - **Base Model**: OpenAI Whisper-small - **Optimization**: Fine-tuned on Kiswahili data - **Target WER**: less the better ### Integration (STT + LLM) - **Memory Budget**: 30 for STT ## Evaluation Criteria | Component | Metric | Weight | Target | Direction | Per-Metric Scores | |-----------|---------|---------|---------| ---------| ---------| | **STT** | Word Error Rate (WER) | 85% | 30 (1 hour audio= 2min transcription)[huggingface.co] | higher is better | $s_{\text{IRTF}}=\bigl(\tfrac{\text{IRTF}-30}{2970}\bigr)$ | | **System** | Peak GPU Memory | 5% | float: """ Calculate Word Error Rate (WER) between reference and hypothesis strings. WER = (S + D + I) / N Where: - S = number of substitutions - D = number of deletions - I = number of insertions - N = number of …