uniBruce
The correspondence of audio snippets between different emotions
Hi,
I downloaded part of the datasets and found that the correspondence of the audio snippets are not arranged as what I have expected.
I thought the number in the filename indicates the content of the audio. E.g. 001.mp4 in disgusted should be the same content as 001.mp4 in neutral. But unfortunately, they are not the same in M003. And I don't know why 30 snippets are provided for emotions other than neutral, but there are 40 snippets for neutral.
Could you explain to me why is it? And could you provide the correspondence relations of different snippets? It is really hard to use your dataset if the correspondence are not provided.
Thanks
Hi,
I downloaded part of the datasets and found that the correspondence of the audio snippets are not arranged as what I have expected.
I thought the number in the filename indicates the content of the audio. E.g. 001.mp4 in disgusted should be the same content as 001.mp4 in neutral. But unfortunately, they are not the same in M003. And I don't know why 30 snippets are provided for emotions other than neutral, but there are 40 snippets for neutral.
Could you explain to me why is it? And could you provide the correspondence relations of different snippets? It is really hard to use your dataset if the correspondence are not provided.
Thanks
Sorry about this problem. It seems that M003 has got one audio&video lost, which, I think, makes the correspondence a mess. I will contact the related staff to check with that. Normally, the correspondence you want is in the supplementary of the paper (i.e., the orders of the sentences). In terms of the number of emotions, you could try to read the supplementary as well. Actually, we capture emotional data (except neutral) in three intensities, which means the amount of emotional data will be much larger than that of the neutral class if we don't capture more sentences for neutral.
Thank you for the reply.
I have read the supp and realize why there are different numbers of snippets in neutral and other emotions. But there still exists other problems.
I have carefully listened to some of the audio snippets in the datasets. I found that not only the order of the sentences are wrong, some of the snippets actually contains more than one sentence. e.g. The actor read two sentences within one snippet. So I used asr to find the correspondence of different snippets........it is really painful.
If you have updated your datasets, would you please let me know? I am not sure if I have found the correspondence 100% correctly by asr.
Thank you for the reply.
I have read the supp and realize why there are different numbers of snippets in neutral and other emotions. But there still exists other problems.
I have carefully listened to some of the audio snippets in the datasets. I found that not only the order of the sentences are wrong, some of the snippets actually contains more than one sentence. e.g. The actor read two sentences within one snippet. So I used asr to find the correspondence of different snippets........it is really painful.
If you have updated your datasets, would you please let me know? I am not sure if I have found the correspondence 100% correctly by asr.
Ok, I will contact the annotation team asap. It seems that the mistakes are more likely from their work.
Hi, in my experience the correspondence is still broken, and some actors follow a completely different naming convention (the filenames go up to 60, skipping various numbers between them. I used MEAD in a research project of mine and annotated the transcriptions for all actors. You can find it here:
https://github.com/filby89/spectre/blob/master/data/list_full_mead_annotated.txt
