Requested Feature

Add comprehensive MLX Whisper support to the docling ASR pipeline to provide 5x performance improvement on Apple Silicon devices through automatic hardware-aware model selection. The integration should be completely transparent to users - they simply use regular Whisper models and get MLX optimization automatically when beneficial.

User Need Addressed

• Performance: Apple Silicon users currently experience slower ASR performance compared to native MLX implementations
• Simplicity: Users shouldn't need to manually configure MLX-specific models or pipelines
• Compatibility: The solution should work seamlessly across all platforms with appropriate fallbacks
• Transparency: Existing code should work unchanged while gaining performance benefits

Key Requirements

1. Automatic Detection: Detect Apple Silicon (MPS) and MLX Whisper availability
2. Transparent Integration: Users use regular WHISPER_TURBO, WHISPER_BASE, etc.
3. Smart Fallback: Fall back to native Whisper on non-Apple Silicon systems
4. Complete Coverage: Support all Whisper model sizes (tiny, small, base, medium, large, turbo)
5. CLI Enhancement: Automatic pipeline detection for audio files
6. Type Safety: Proper type annotations and MyPy compliance

Alternatives

Alternative 1: Manual MLX Configuration

• Approach: Require users to explicitly configure MLX models
• Rejected: Adds complexity and breaks transparency principle
• Why: Users would need to know about MLX-specific models and configuration

Alternative 2: Separate MLX Pipeline

• Approach: Create a separate MLX-specific ASR pipeline
• Rejected: Fragments the user experience and requires manual pipeline selection
• Why: Users would need to choose between native and MLX pipelines

Alternative 3: Runtime Detection Only

• Approach: Only detect MLX at runtime without automatic model selection
• Rejected: Doesn't provide the seamless experience users expect
• Why: Users would still need to manually configure MLX models

Chosen Solution: Automatic Hardware-Aware Selection

• Approach: Embed automatic MLX/Native Whisper selection directly into model specs
• Benefits:
  • Completely transparent to users
  • Automatic performance optimization
  • Backward compatible
  • Works across all interfaces (Python API, CLI, examples)