Datasets:

Modalities:
Audio
Size:
< 1K
ArXiv:
DOI:
Libraries:
Datasets

Annotation quality is very low, not usable for training

#8
by da1sypetals-iota - opened

title describes itself

I disagree with that characterization. GTSinger has already been widely used in both academia and industry for singing and music model training, so it is simply not accurate to call it "not usable for training." At least for the Chinese and English portions, the lyric annotations, score annotations, and MFA alignments have generally been regarded as reasonably reliable. For other languages, only the Korean and Italian subsets do not have the same level of fine-grained annotation, and we have never tried to hide that limitation.

More importantly, this dataset was not assembled casually. We involved dozens of university students in a careful manual verification and correction process, and a substantial amount of human effort went into improving the annotation quality.

If you believe there are serious annotation problems, please point them out with concrete evidence and specific metrics: which subset, which type of annotation, what error rate, and what reproducible impact on training results. Simply saying that the annotation quality is "very low" and "not usable for training" is not a serious or professional assessment.

To be precise, the MIDI does not align well with the audio. Audio itself is pretty good, but as a nobody-cares amateur singer, I manually listened to around 1% audio wav and their accompanying MIDI, it's just not right. Drag some of the audio and MIDI to any DAW and you will know what I am saying.
Perhaps I am expecting too high (ROFL).

Thanks for clarifying. I understand the issue better now.

You are right that the MIDI / MusicXML in GTSinger should not be interpreted as a perfectly time-aligned DAW arrangement. For singing voice, this is much harder than instrument MIDI transcription: singers often do not strictly follow the written score, and expressive singing naturally includes timing deviations, pitch bends, vibrato, slides, and phoneme-level duration changes. So a MIDI track derived from vocal performance will not always look like a cleanly arranged instrumental score when dragged into a DAW.

Our annotation strategy was therefore not to create frame-perfect MIDI that exactly follows every F0 fluctuation. Instead, we aimed to provide realistic music scores that are useful for SVS and singing-related tasks. In our pipeline, we first extracted F0 with RMVPE, then used ROSVOT to derive the initial MIDI / score representation. After that, music experts manually listened to the recordings and references, checked the generated scores, and adjusted tempo, clef, key, note pitch, note duration, and note type where needed.

At the time, ROSVOT was already a strong SOTA system for singing voice transcription. In the ROSVOT paper, it achieved COnPOff F1 of 77.4 / 77.0 on clean / noisy inputs, compared with 65.8 / 62.1 from the previous method, and also reached Pitch AOR around 97.0 and Melody RPA around 87.6. We used it because pure-vocal MIDI annotation at this scale is genuinely difficult, and this was the best practical pipeline available to us.

That said, I am sorry that the released annotations did not meet your expectation in some cases. We did put a lot of effort into the annotation and manual checking, but for a dataset of this scale and across multiple languages/singers/techniques, it is still possible that some files have imperfect note timing or score alignment.

If you can provide a few concrete examples, such as file paths, timestamps, and what kind of mismatch you observe, we would be happy to check them carefully and consider fixing or documenting those cases.

I’ve been auditing the GTSinger MusicXML files for a study that pairs written notes with recorded vocal pitches to analyze pitch classes and scale degrees. I found many clips where the key declared in the MusicXML differs from the key I hear in the recording. Here are some examples:
'Italian/IT-Soprano-1/Mixed_Voice_and_Falsetto/caro mio ben/Mixed_Voice_Group/0000.wav', clip should be A major, musicxml reads C major.
'German/DE-Soprano-1/Mixed_Voice_and_Falsetto/Stille Nacht Heilige Nacht/Mixed_Voice_Group/0001.wav', clip should be Gb/F# major, musicxml reads F minor.
'English/EN-Alto-2/Mixed_Voice_and_Falsetto/Call Me Maybe/Mixed_Voice_Group/0003.wav', clip should be Gb/F# major, musicxml reads C major.
'Spanish/ES-Bass-1/Mixed_Voice_and_Falsetto/Devuélveme El Corazón/Mixed_Voice_Group/0004.wav, clip should be Eb/D# minor, musicxml reads C major.

I’ve also encountered passages where the written rhythm is difficult to reconcile with the recording; the transcription switches between adjacent semitones during what sounds like a single sustained note, or audible notes appear to be missing. These issues affect the analysis in different ways. An incorrect tonic shifts every calculated scale degree in a clip, while missing or extra notes and mismatched boundaries can cause the analysis to pair a written note with the wrong part of the recording. Taken together, these issues make reliable score-audio paired analysis impractical for my study without substantial manual review.

The paper says music experts listened to the recordings, consulted the original accompaniments, and checked key, tempo, pitch, and duration. Given that review process, I expected more reliable key metadata and note annotations. The quality of the released files I’ve checked falls short of what I expected from the method described in the paper.

Could you verify the examples I provided and clarify whether the key fields in the released MusicXML were checked against the actual recorded takes? If you have a corrected key list, a score revision, or documentation identifying files with known annotation issues, that would help researchers assess which clips are usable for this kind of analysis.

@nonocut Thanks for the examples, and sorry for the trouble this has caused. We extract F0 from the actual singing, derive MIDI from it, and estimate the key from that MIDI automatically. Errors can remain in both note timing and key estimates, so the released annotations aren't guaranteed to be precise for every clip.

We plan to revise the MusicXML annotations and will test SheetSage2, the audio-to-score model released alongside YuE2, as a candidate for re-transcription. It outputs vocal MIDI and key estimates; we'll check its results on your examples before deciding whether to use it.

Sign up or log in to comment