v5.0.0 — Local Models, Settings Panel, Enterprise Download Pipeline

What's New

Local Model Support (offline, no API key)

• 19 models from Parakeet, Whisper, Moonshine, SenseVoice, and GigaAM
• Accuracy/speed ratings calibrated against published WER benchmarks
• Models auto-download on first use, fully offline after that
• Device-aware recommendations based on your RAM, CPU, and GPU

Unified Settings Panel (/voice-settings)

• 4 tabs: General, Models, Downloaded, Device
• Overlay panel with keyboard navigation (arrow keys, fuzzy search)
• Toggle backend, change language, manage models — all in one place

Enterprise Download Pipeline

• Pre-checks: disk space, network, write permissions, sherpa-onnx runtime
• Live progress: percentage, MB/s speed, ETA countdown
• Post-verification: file integrity, size sanity check
• Concurrent download deduplication
• Partial downloads auto-resume on retry

Handy Integration

• Auto-detects models from Handy app
• Import via symlink — zero disk duplication

Bug Fixes

• Fixed local transcription text never appearing in editor
• Fixed voice state machine stuck in "finalizing" after session abort
• Fixed all models showing [may be slow] on 8-16GB machines

Install

pi install npm:@codexstar/pi-listen