Clarifications on annotation

[MikeMpapa]

Hi there,
I am working on a building a new dataset in Spanish (polysyllabic language). I have gone though MakeDiffSinger but I still have some gaps. I would be grateful if you could sanity check me on my understanding and share any thoughts you might have

Questions for clarifications:

1. ph_seq: These are sequences of phonemes or syllables?
   Currently I using phonemes and their timestamps as provided by MFA. I am using a pre-trained Spanish model available by MFA. Would you recommend training a new one on my specific data?

2. note_dur: The midi notes should be estimated over phonemes, syllables, or words?
   Now I estimated one note for each phoneme and assumed ph_dur==note_dure

3. ph_num: The number of phonemes in each word or in each syllable?
   Now I assumed the number of phonemes in each word

4. note_seq: Do you think SOME would suffice to get a first shot at this ? I would speculate yes?

5. is_slur: how would you define slur in this context? I have not found plenty of resources on this topic
   Now I assumed no slurs at all

6. SPs and APs: Would you recommend doing that manually or using the enhance script might be OK for a first shot?

Thanks!

[Hikari-Tsai]

• ph_seq is a phoneme sequence.
• note_dur is the note duration. I am not sure how this works for Spanish, but for Japanese and Chinese, it is basically one vowel plus one consonant. So note_dur will be split into two ph_seq entries: one consonant and one vowel. Their sum equals note_dur.
• ph_num refers to each word, so your assumption is correct. However, there may be a SHIFT issue. This comes from common music production practice: the downbeat is usually aligned to the vowel, so the consonant comes in earlier. In other words, the consonant’s timing may be allocated from another note.
• SOME is very useful, so please definitely try it. However, it currently looks like the team has introduced a newer model.
• This is up to the creator. You can decide according to the song.
• For AP and SP, I recommend processing them with a model. Basically, if the dataset is large enough, around 2 hours or more, manual annotation is not necessary, since the model should have a reasonable level of accuracy.