"High quality" is one of those phrases that marketing pages love and ML engineers distrust. If you are evaluating vocal datasets for training a singing voice synthesis model, the word is meaningless unless you can decompose it into measurable attributes. This post does exactly that.
The quality of a vocal dataset has six orthogonal dimensions. A dataset can score well on some and badly on others. A dataset that is excellent on five out of six can still be unusable if it fails on the sixth. Below we walk through each dimension, describe what "good" looks like, and explain how to test a sample before committing to a full licensing deal.
Dimension 1: Signal quality
The first and most basic question is whether the audio itself is clean. This is not about whether the vocalist can sing. It is about whether the recording signal carries the vocalist's performance without contamination.
Sample rate and bit depth
For modern singing voice synthesis, the practical options are 44.1 kHz and 48 kHz. 44.1 kHz is the consumer music standard and captures frequency content up to 22.05 kHz (the Nyquist limit). 48 kHz is the broadcast standard and captures up to 24 kHz. For the purposes of vocal training, 44.1 kHz is usually sufficient because the human voice rarely contains useful harmonic content above 18 kHz.
Bit depth should be 24-bit for studio recordings. 16-bit indicates either consumer-grade source material or recordings that have been through a reduction step, which loses dynamic range headroom. If a vendor is offering 16-bit material for enterprise AI training, ask specifically whether the recordings were captured natively at 16-bit (legacy material) or reduced from 24-bit at some point (lossy processing).
One counterintuitive point: higher sample rates are not automatically better for training. The HiFiSinger paper (Arxiv 2009.01776) showed that moving from 24 kHz to 48 kHz created wider spectrum bands and longer waveforms that made acoustic models and vocoders struggle to converge. Dedicated architectures are needed to actually benefit from the extra bandwidth. For most production use cases, 44.1 kHz is the sweet spot.
Signal-to-noise ratio
Studio recordings should have a noise floor below -60 dBFS. Anything higher introduces background hiss that the model will learn as legitimate signal, producing outputs with baked-in noise that is impossible to remove downstream. When you receive a sample dataset, run a quick noise-floor measurement on the silent intros and outros of a few tracks. If the noise floor is inconsistent across recordings, the dataset was captured in different environments and will need normalization.
Dynamic range and clipping
Clipped recordings are common in material that has been through streaming mastering or consumer processing. Look for peaks that hit exactly 0 dBFS with flat tops in the waveform view. A clipped recording has lost information that cannot be recovered, and the model will learn the clipping as part of the signal.
Dynamic range should be wide enough to capture both quiet passages and loud crescendos. If every recording has been hit with heavy compression (DR under 6 dB), the dataset will produce outputs that sound compressed by default, which is fine for pop but limiting for classical or jazz training.
Dimension 2: Isolation
The second dimension is whether the vocal signal is actually isolated or whether it contains bleed from other sources. This is the single biggest quality differentiator between truly studio-recorded datasets and datasets built from source-separated stems.
The problem with source-separated stems
Modern source separation models like HTDemucs can extract a vocal stem from a mixed track with a signal-to-distortion ratio around 9 dB. That is impressive, but it is not equivalent to a dry studio recording. The extracted stem still contains:
- Reverb tails from the room or hall in which the mix was produced
- Harmony bleed from background vocals that share frequency content with the lead
- Phase artifacts from the separation process itself
- Transient smearing around consonants and fast vocal passages
- Spectral bleed from instrumental elements that overlap the vocal range
All of these contaminations get learned by the model. You train on separated stems, you get a model that produces outputs with inherited separation artifacts. The outputs may be indistinguishable to a casual listener but will be audibly degraded to a producer or engineer.
The research literature on singing voice synthesis is explicit about this. As the DiffSinger paper notes, research datasets use "solo vocals in controlled environments with limited effects" for a reason. The signal is cleaner and the model learns voice rather than voice-plus-room-plus-processing.
How to test for isolation
Take a few tracks from the sample dataset and do the following:
- Load them into a spectrogram viewer and look for horizontal bands of energy in the silent passages. True silence should look black. Bleed shows up as faint horizontal lines.
- Listen with headphones to the tails of vocal phrases. A dry studio recording cuts cleanly when the vocalist stops singing. Reverb tails and bleed from other instruments are audible.
- Run the tracks through a phase-invert comparison with the original mix (if available). A cleanly isolated stem should not phase-cancel with anything else in the mix.
Dimension 3: Dry vs wet, processed vs unprocessed
This is a close cousin of isolation but worth treating separately. Isolation asks "does the vocal have bleed from other sources?" Processed/unprocessed asks "has the vocal itself been altered?"
Dry stems are unprocessed vocal recordings straight from the microphone, possibly with basic gain staging but nothing else. They are the rawest form of the performance and the most flexible training material because any effect you want (reverb, compression, EQ) can be added later and the model's outputs will match your processing chain.
Wet stems are vocals with effects already applied, usually including reverb, compression, EQ, and de-essing. They are ready-to-use in a professional mix context and reflect a specific production aesthetic. A model trained exclusively on wet stems will produce outputs that sound pre-processed, which may be desirable (pop production context) or undesirable (research application that wants to apply custom processing downstream).
The best enterprise datasets include both versions of each recording: a dry version for flexible training and a wet version for production-aesthetic training. This doubles the effective dataset size without requiring additional recording sessions and gives downstream users the ability to choose their training target.
Dimension 4: Metadata
A vocal dataset is only as useful as the metadata that describes it. Without metadata, every training run requires manual labeling or automated extraction, which adds cost and introduces errors. With rich metadata, the same dataset can support multiple model architectures and conditioning strategies.
The baseline metadata for enterprise-grade vocal training data includes:
