Critical bug: Fix --no-timestamps flag behavior

[OrelSokolov]

Fix --no-timestamps flag behavior

Summary

This PR fixes the --no-timestamps flag to only affect output formatting without changing transcription quality. Previously, the flag would alter the decoding process, resulting in different and lower quality transcription text.

Problem

When using --no-timestamps flag:

• ❌ Transcription text differed from the same audio without the flag
• ❌ Lower transcription quality
• ❌ Model would sometimes loop/repeat phrases infinitely
• ❌ Flag modified the decoding process by:
  • Adding <|notimestamps|> token to the prompt
  • Suppressing all timestamp tokens during decoding

Solution

The fix ensures --no-timestamps only controls output formatting:

• ✅ Model always uses timestamp logic during decoding (for better quality)
• ✅ Transcription text is identical regardless of the flag
• ✅ Added repetition detection to prevent infinite loops
• ✅ Improved segment handling to prevent early termination

Changes

Core Fixes (src/whisper.cpp)

1. Removed <|notimestamps|> token injection - The model no longer adds this token to prompts, allowing proper timestamp-based segmentation
2. Removed timestamp token suppression - Timestamp tokens are no longer suppressed from logits, enabling the model to segment properly
3. Added repetition detection - Detects and prevents infinite loops where the model repeats the same phrase
4. Improved error handling - Better buffer allocation error messages

Tests (tests/)

• Added test-no-timestamps.cpp - Automated test that verifies transcription quality is identical with/without the flag
• Added TEST_NO_TIMESTAMPS.md - Test documentation
• Updated tests/CMakeLists.txt - Test integration

Documentation

• Added NO_TIMESTAMPS_FIX.md - Detailed explanation of the problem and solution

Testing

Automated Test

cd build
ctest -R test-no-timestamps -V

Test verifies that:

• Transcription with timestamps enabled produces text: "And so my fellow Americans..."
• Transcription with --no-timestamps produces identical text
• ✅ Test passes (9.87s)

Manual Testing

# Both commands now produce identical transcription quality:
./whisper-cli -m model.bin -f audio.wav                    # With timestamps in output
./whisper-cli -m model.bin -f audio.wav --no-timestamps    # Without timestamps in output

Backward Compatibility

✅ Fully backward compatible

• All existing tests pass
• CLI interface unchanged
• API unchanged
• Only improvement: better transcription quality with --no-timestamps

Impact

• Users: Better transcription quality when using --no-timestamps
• Developers: Clear separation between output formatting and decoding logic
• Maintenance: Automated test prevents regression

Checklist

• Code follows project style guidelines
• Self-review completed
• Comments added for complex logic
• Documentation updated
• Tests added and passing
• No new warnings introduced
• Backward compatible