Feature Request: Optional wav2vec2 CTC Forced Alignment

[obvirm]

Feature Request: Optional wav2vec2 CTC Forced Alignment

Problem:
Current timestamp accuracy relies on cross-attention weights, which can be
imprecise for word-level alignment, especially in noisy audio.

Proposed Solution:
Add optional wav2vec2 CTC forced alignment as a post-processing step.

Trade-offs to Consider:

• Significantly improved word-level timestamp accuracy
• Requires ONNX Runtime dependency
• Additional 300MB+ model download
• Increased inference time (2-pass)

Questions for Maintainers:

1. Would this violate the "minimal dependencies" philosophy?
2. Is there interest in a separate whisper-align binary?
3. Could this be implemented as an optional CMake flag?