JONAHKYAGABA/-Parakeet-Kinyarwanda-Fine-Tuning-Pipeline
Domaine:
natural language processingType de record:
softwareproject
Créateur:
JON
Hôte:
# 🇷🇼 Parakeet Kinyarwanda TTS Fine-Tuning Pipeline
# 🇷🇼 Parakeet Kinyarwanda TTS Fine-Tuning Pipeline
This project implements a complete pipeline for fine-tuning the Parakeet CTC model (v0.6b) on the **Kinyarwanda TTS dataset** using NVIDIA NeMo and Hugging Face's ecosystem.
## 📌 Project Overview
- 🔉 Loads and parses 5 hours of Kinyarwanda TTS data (CSV and audio)
- 📊 Computes dataset statistics (duration, WER, vocabulary, SNR)
- 📁 Generates train/val/test JSON manifests
- 🔤 Builds a Byte Pair Encoding (BPE) tokenizer
- 🎙️ Loads and adapts a pre-trained `parakeet-ctc-0.6b` ASR model
- 🧠 Fine-tunes the model using subword tokenization and frozen encoder
- 📈 Logs metrics and training progress to Weights & Biases (wandb)
- ☁️ Optionally pushes final `.nemo` model to the Hugging Face Hub
## 📂 File Structure
Similaires
JONAHKYAGABA/-Whisper-ASR-Fine-Tuning-for-AmharicJONAHKYAGABA/-Bemba-ASR-Evaluation-PipelineJONAHKYAGABA/-Kinyarwanda-ASR-Qualitative-EvaluationJONAHKYAGABA/Kinyarwanda-Speech-Translation-Evaluation-FrameworkJONAHKYAGABA/oromo-ASR-evaluation-error-analysis-pipelineFine-Tuning Whisper for Kinyarwanda: A Practical Approach to Low-Resource ASR Development