v0.3.0

🎯 WhisperForge 0.3.0: Quantization & Performance

🚀 Major Features

Phase C: INT8 Post-Training Quantization ✅

• 4× model size reduction: Tiny (150 MB → 37 MB), ideal for edge deployment
• --quantize int8 flag in model converter
• Transparent loading—no code changes required
• Full precision (FP32) and quantized (INT8) models interoperable

Phase B.5: GPU-Accelerated Mel Spectrogram ✅

• CubeCL DFT kernel for GPU mel filterbank matmul
• --features cubecl-stft enables GPU STFT pipeline
• Faster audio preprocessing on WGPU backend

Burn 0.21 & burn-flex Migration

• Latest Burn with improved numerical stability
• CPU fallback (burn-flex) seamlessly handles CPU inference
• Better WGPU runtime integration

📊 What's Changed

• Quantized models fully compatible with CLI and library API
• Streaming audio pipeline (Phase B) now fully integrated
• Fixed EOT suppression at step 0 for robustness
• Improved error handling across all crates

📦 All 5 Crates Published

• whisperforge-core v0.3.0 — library
• whisperforge-cli v0.3.0 — binary
• whisperforge-convert v0.3.0 — model converter
• whisperforge-align v0.3.0 — VAD + SRT
• whisperforge-diarize v0.3.0 — speaker diarization

🛠️ Dependency Updates

Simplified workspace dependency management: inter-crate deps now use workspace version automatically.

📖 Next Phase

Phase D: WASM Target—browser-native speech-to-text with wasm-bindgen.

Full Changelog: v0.2.0...v0.3.0