Audio Loading Uses MP3-Specific Decoder and May Include Priming/Trailing Frames

[Vaida12345]

Hi @Yujia-Yan, thanks for the great work!

I’ve been experimenting with your model, and it appears that the current model can only handle mp3 files correctly.

In transcribe.py:10, the readAudio function uses pydub.AudioSegment.from_mp3(path) to obtain the audio file. However, it seems that this function returns the untrimmed full package mp3 frames, including priming frames and remainder frames. This discrepancy might also affect training, as my audio is transcribed most accurately when first converted to mp3. On the other hand, using the original wav file, a few sustain events are missing, which are quite noticeable to the ear.

[Vaida12345]

Additionally, it appears that the model currently treats all input as 16-bit audio, which is why the y = np.float32(y)/2**15 line is present. However, loading an audio file with a 24-bit depth seems to work fine.

[AIGLE25]

you just need to change the line by "audio = pydub.AudioSegment.from_file(path)" or something like this