[Manishram-ai]

_detect_mime_type only treats bytes as MP3 when they start with an ID3 tag or exactly FF FB, which is MPEG-1 Layer III without CRC. Many MP3s without an ID3 tag start differently: 16 to 24 kHz files (FF F3), 8 to 12 kHz files (FF E3) and files with CRC (FF FA). AAC in ADTS starts with FF F1. All of these fall through to the "image" default, so automatic modality detection treats audio columns as images.

This uses the same check as the frontend's detectAudioMimeType in packages/utils/src/media.ts: the 11-bit sync word, then the layer bits to tell AAC from MPEG audio. JPEG is still checked first, so FF D8 isn't affected.

Testing:

• 5 new cases in tests/test_modality_detection.py fail on main and pass with the fix (54 passed).
• I made a 22.05 kHz MP3 and an AAC file with ffmpeg. Both were detected as "image" before and "audio" after.
• ruff format --check and ruff check are clean.

[donghaoren]

Thanks! Could you add a few more non-audio tests like this:

b"\xff\xd8\xff\xe0" + b"\x00" * 20,  # JPEG must win over the sync check
b"\xff\xff\xff" + b"\x00" * 20,  # 0xFF padding
b"\xff\xe1\x00" + b"\x00" * 20,  # not valid MPEG or ADTS
b"\xff",  # too short

These should all return "image".

[Manishram-ai]

Thanks, added all four in f1083fc. Two of them (0xFF padding and FF E1 00) came back as audio, so the check now also looks at the rest of the frame header before calling it audio. The frontend check in media.ts has the same gap, and I'm happy to fix that in a separate PR.