Do not rely on whisper timestamps

[Jeronymous]

Implement a strategy to avoid relying at all on whisper timestamp prediction:

• do the alignment on the full 30 sec audio chunk
• recover the segment of words from that