Errors when transcribing MP3, but not M4A, WAV, or AIFF

[tinyapps]

When attempting to convert MP3 files, received this error:

% yap test.mp3                   
⨯ Failed to transcribe audio using locale: "en_US" [0.2s]
Error: nilError

After reading Issue #1 (Unsupported Locale), tried setting the locale manually:

% yap --locale en_US test.mp3
Error: The operation couldn’t be completed. (com.apple.coreaudio.avfaudio error 2003334207.)

Other tested audio formats (M4A, WAV, and AIFF) completed without error:

% yap test.m4a
✔︎ Audio transcribed using locale: "en_US" [0.3s]
<transcription appears here>

[atmosfar]

Your error code suggests the file does not exist, perhaps you are using the wrong path.

https://www.osstatus.com/search/results?platform=all&framework=all&search=2003334207

MP3 transcription works fine for me.

[atmosfar]

#15

[tinyapps]

Thank you for your reply, @atmosfar. The file does exist; I dragged it from Finder into Terminal to make sure. Other files with the same name (but different extensions) in the same directory work fine, as described above.

[atmosfar]

What does ffmpeg/ffprobe say about this MP3 file? Have you tried with a few different ones?

brew install ffmpeg
ffmpeg -i test.mp3

[tinyapps]

Duration: 00:01:41.83, start: 0.025057, bitrate: 76 kb/s
  Stream #0:0: Audio: mp3 (mp3float), 44100 Hz, mono, fltp, 76 kb/s
      Metadata:
        encoder         : LAME3.99r

whisper-cli doesn't have any trouble transcribing this or any other MP3 I've tried.

[atmosfar]

Is it failing on just this one file or every MP3 file?

Tried to recreate to see if it was to do with a non N^2 bitrate, but it seems fine.

  Duration: 00:00:03.45, start: 0.025057, bitrate: 57 kb/s
  Stream #0:0: Audio: mp3 (mp3float), 44100 Hz, mono, fltp, 56 kb/s, start 0.025057
    Metadata:
      encoder         : LAME3.100

yap --locale en_US test_clip.mp3
✔︎ Audio transcribed using locale: "en_US" [0.3s]
So you were hosting the podcast. Yeah. Nice.

[tinyapps]

Just tried a higher quality MP3 losslessly cut from a podcast episode, and yap transcribed it just fine. Here's how the file differed from all the other MP3 voice recordings which fail:

  Duration: 00:00:02.22, start: 0.000000, bitrate: 128 kb/s
  Stream #0:0: Audio: mp3 (mp3float), 44100 Hz, stereo, fltp, 128 kb/s

For a short while, I was using a one-liner to convert MP3 to M4A then run through yap, but whisper's large model produced more accurate results, so I've since gone back to it.

[atmosfar]

Yep, it's definitely more accurate. Hoping that the Speech framework improves, as it's early days for this pipeline.