How does VideoARM ingest subtitles for benchmarks without embedded audio (e.g. LongVideoBench)?

[SoftEgLi]

Hi, thanks for the great work and for open-sourcing the implementation!

I'm trying to reproduce your LongVideoBench numbers and ran into a question about how the subtitle/audio channel is intended to work on this particular dataset.

[qiankemeng]

Hi, thanks for the question.

The current open-source version does include the audio_transcriber tool, which extracts audio segments from the input video and transcribes them with Whisper when an audio stream is available.

I may not have fully understood the exact problem you ran into. Could you clarify your setup a bit?

• Are you running VideoARM directly on the LongVideoBench MP4 files?
• Do those videos contain an audio stream in your local copy?
• Are you expecting VideoARM to use LongVideoBench’s separate subtitle files instead of transcribing audio?
• What exactly happened in your run: no_audio, ffmpeg failure, empty transcript, or a mismatch with the provided subtitles?

If you can share the command and a short log snippet, we can better identify whether this is an audio-transcription issue or a dataset-subtitle integration issue.