ai-agents-timestamp-decoding-when-decoding-whisper-output-with-whispertokenizer-383d619e

THEN Fix the timestamp offset calculation in WhisperTokenizer.batch_decode when output_offsets=True. Instead of relying solely on cur_max_timestamp, use the actual segment timestamps predicted by the model to correctly offset consecutive chunks. This ensures that silence gaps are properly reflected in the decoded timestamps.

# Previously, decoding with output_offsets gave wrong timestamps for chunks after silence.
# Fix: ensure that offsets use segment timestamps from the model output, not computed from previous max.
result = processor.decode(token_ids, output_offsets=True)  # after fix, timestamps align with segments