Advanced AI agent, dataset alignment, and hyperparameter-optimized training with WER & CER analytics.
0AI Agent
1Collect
1.5Align & Review
2Review
3Train
Step 0: AI Expert Agent & Automation
Autonomous Workflow: This agent understands your goals. It collects data, dynamically selects between WER and CER based on language, finds optimal hyperparameters (including Dropout), and trains a single, fully optimized model with early stopping.
🤖 Agent is ready. Ask me to do something or explain a concept!
⚡ Quick Commands
📥 "Collect data from [URL] and train"
📚 "Explain WER/CER and freezing encoder"
🚀 "Run advanced auto-training"
🛡️ Safety & Cleanup
The agent automatically enforces save_total_limit=1, uses Early Stopping, and deletes all temporary search trials to prevent disk clutter.
Step 1: Select An Option To Load/Record Dataset
🎙️ Record Audio
📺 YouTube Collector
Falls back to Whisper v3 Turbo if subtitles unavailable.
New
📁 Excel + Audio Align
Upload an Excel/CSV and multiple WAVs. Audio named "clip_5.wav" maps to Row 5.
📄 Legacy CSV Upload
CSV columns: audio, transcription
Step 1.5: Interactive Alignment Preview
Review matches. Edit text, play audio, change language, or remove rows before saving.
Instruction: Audio files are mapped to Excel rows by number. Use the Language dropdown to enforce specific decoding languages per row.
Match ID
Audio
Language
Transcription (Editable)
Translation (Editable)
Action
Step 2: Review, Edit & Evaluate
Listen
Transcription
Split
Actions
Select a dataset
🧪 Test & Compare Models (WER & CER)
Note: Comparison uses ONLY rows labeled as 'test'.
Step 3: Train & Optimize
⚙️ Advanced Auto Hyperparameter Search (Optuna)
Searches over Learning Rate, Batch Size, Weight Decay, Dropout, and Scheduler.
Trains final model with extended epochs for maximum accuracy.
Running…
🛠️ Manual Training
🎯 Threshold Training Monitor
Current Round
0/0
Best WER
--%
Best CER
--%
Status
Running
⚡ Dynamic Parameter Adjustment
Changes will be applied at the start of the next training round.