Spaces:

shaun3141
/

caribbean-voices-hackathon

Sleeping

App Files Files Community

caribbean-voices-hackathon / README.md

shaun3141

Switch to Docker SDK to use custom Dockerfile with multi-stage caching

d632650 20 days ago

preview code

raw

history blame contribute delete

2.54 kB

metadata

title: Caribbean Voices - OWSM v3.1 Platform
emoji: 🎤
colorFrom: purple
colorTo: pink
sdk: docker
app_port: 7860
pinned: false
hardware: gpu-a10g-large
license: mit

Caribbean Voices Hackathon - OWSM v3.1 Platform

Hugging Face Space for OWSM v3.1 training and inference with progress tracking.

✅ Dataset Upload Complete

Dataset: shaun3141/caribbean-voices-hackathon (Private)

The dataset has been uploaded and the app is configured to use it automatically.

Quick Start

HF Space Configuration

Environment Variable (optional - already set in code):

HF_DATASET_NAME=shaun3141/caribbean-voices-hackathon

For Private Dataset Access (if needed):

HF_TOKEN=your_hf_token_here

Set these in: Settings → Variables (or Settings → Secrets for token)

The Space Will Automatically:

✅ Load CSV files from dataset on startup
✅ Load audio files on-demand during inference
✅ Cache data locally for faster access
✅ No manual uploads needed!

Features

Data Loading: Auto-load from HF Dataset (no uploads needed!)
Entity Extraction: Extract Caribbean entities from training transcripts
Model Training: Fine-tune OWSM v3.1 with entity-weighted loss
Batch Inference: Generate transcriptions for test set with progress tracking
Single File Transcription: Quick transcription with multiple models
Status Monitoring: Real-time status of setup and training progress

Usage

Deploy Space - Data loads automatically from dataset
Check Status: View system status and setup progress
Extract Entities: Run entity extraction on training data
Train Model: Fine-tune OWSM (requires ESPnet recipes for full training)
Run Inference: Generate test set transcriptions
Download Results: Get submission CSV file

Requirements

ESPnet and espnet_model_zoo (for OWSM models)
PyTorch
Transformers
Gradio
Datasets (for HF Dataset loading)

Model Support

Wav2Vec2 Models: Fast baseline models
OWSM v3.1 Small: Open Whisper-style model with ESPnet

Dataset Information

Name: shaun3141/caribbean-voices-hackathon
Privacy: Private
Train: 19,856 samples
Test: 8,510 samples
URL: https://huggingface.co/datasets/shaun3141/caribbean-voices-hackathon

Documentation

DATASET_SETUP.md - Dataset upload instructions
DATA_LOADING.md - Data loading methods
SPACE_CONFIG.md - Space configuration guide
FILES.md - File organization