What is Singing Voice?
Vocal recordings with melody, in both a cappella and accompanied form, used for singing voice synthesis, music information retrieval, and lyric alignment.
Singing voice data is recorded vocal performance with melody: pitch that follows a musical line rather than the contours of speech. It covers solo performance, choral work and the range between them. The technical difference from speech is that pitch and timing are constrained by a musical structure rather than by meaning.
Commercial demand comes from three places: singing voice synthesis, where a model generates a vocal performance; music information retrieval, where a model extracts pitch, lyrics or structure from a recording; and lyric alignment, where a model maps text to a performance in time.
The category carries a rights problem no other speech category has. A recording of a song involves the composition and the lyrics as separate works with separate holders, plus the performance itself. That is why we only produce original or fully cleared material, and why a supplier offering a corpus of popular songs is offering something that cannot be sold.
What buyers get wrong
Copyright is the number one problem for this category: melody and lyrics each have their own rights holder. We only produce original or fully cleared content, never existing commercial tracks.
The specification detail that decides everything
Whether the accompaniment is separated (dry vocal) or the mix is kept, and whether MIDI or pitch annotation comes with it.
This is the line item that most often separates a dataset that works from one that gets re-ordered. It belongs in the specification before production starts, not in the delivery review.
How we source it
We do not hold inventory in this category. A requirement comes in, we match it to a producer who can meet the specification, and we agree the quality bar with both sides before recording begins. A pilot batch goes first so misalignments surface early.
Scenarios this covers
Buyers rarely ask for a category by name. They describe a situation, and the situation is what tells us whether this category is the right one. If yours is in the list, name it in the brief and it goes straight into the specification. If it is not, describe it anyway, because it usually means the requirement sits in a neighbouring category.
What good Singing Voice looks like
What makes singing data usable is not the singing. It is what is recorded alongside it.
-
Dry vocal and accompaniment as separate tracks
Or a documented decision not to separate them. Separating a mixed recording after the fact is lossy, and the artifacts land in the vocal stem where a synthesis model will learn them.
-
Pitch annotation with a stated method
Extracted pitch tracks and hand-corrected ones differ in accuracy, and the difference is measurable for a synthesis model. The specification should say which one it is buying.
-
Lyric alignment at a stated granularity
Word level, syllable level or note level. Each supports a different task and each costs differently, and mixing granularities inside one delivery produces a corpus that cannot be used for any of them.
-
Vocal range and style spread
A corpus of one voice type in one genre produces a model that sings in one voice type in one genre. The range has to be designed, not left to whoever answered the recruitment advert.
-
A clean rights chain for composition, lyrics and performance
Original works written for the project, or works cleared explicitly for machine learning training. All three rights, documented, delivered with the audio.
Public datasets in this category
These are openly documented resources, not products we sell. Each one covers part of the category, and knowing where each stops is usually what decides whether a project needs a collection of its own.
-
MIR-1K
A thousand short clips of amateur singing with separate vocal and accompaniment tracks. The separation makes it useful for vocal extraction work, and the amateur, informal delivery makes it a poor match for produced singing.
-
MedleyDB
Multitrack recordings with the vocal recorded on its own track, spanning several genres. Small, licensed for research, and the multitrack structure is what most singing work actually needs.
-
MUSDB18
The standard benchmark for music source separation, 150 tracks with isolated stems. It is an evaluation set first and a training corpus second, and it is deliberately small.
-
Why the gap is wide here
Commercially released singing is heavily licensed, and the rights sit with labels rather than performers. What is publicly available is research-scale and often amateur, which is why production singing data is commissioned against a specification.
Questions to ask a supplier
This is the category where a supplier's answers about rights matter more than their answers about audio.
- Who wrote the compositions and the lyrics, and what does their agreement permit?
- Is the accompaniment separated at capture, or derived afterward?
- What is the pitch annotation method, and what is its measured error?
- How are singers recruited across vocal ranges and styles?
- Can you deliver the rights documentation per track?
The vague answer here usually concerns rights rather than data. A supplier who says the material is licensed without naming the rights body, the territory and the term has described an intention. In this category the recording, the composition and the performance can each carry a separate right, and a corpus cleared on only one of them cannot be sold.
A usable answer names the source of each layer and shows one complete clearance for one track. Ask to see the chain for a single delivered item, including the performer consent and the treatment of any underlying composition. A supplier who can produce that for one item can usually produce it for the batch, and one who cannot usually cannot.
Technical attributes
The items below are the ones a specification can hold a producer to, and the ones that decide whether a delivery is usable for training rather than only present.
| Track separation | Dry vocal and accompaniment separate, or mixed by decision |
|---|---|
| Pitch annotation | None, extracted, or hand-corrected, with error stated |
| Lyric alignment | Word, syllable or note level |
| Vocal range | Documented per singer, with range and style |
| Genre coverage | Stated per track rather than as an average |
| Delivery format | Separate stems plus annotation in a documented format |
Pitch range and vocal technique coverage are the two attributes that decide whether the corpus trains a model or a caricature. A corpus drawn from one genre produces a model that sings in that genre regardless of the input. The range and technique columns have to be filled against the intended use, which for singing synthesis is usually expressive range rather than the ability to hit a note.
The numbers come from the repertoire rather than from a specification template. Decide first whether the model needs belting, falsetto, breathy delivery, vibrato control, or lyrics in several languages, then derive the coverage requirements from that list. A supplier who proposes a distribution before asking what the voice has to do is offering a sample pack.
The failure mode is specific and hard to reverse. A corpus with narrow technique coverage produces a model that sounds acceptable on the material it was trained on and collapses on anything outside it, and the collapse is audible rather than measurable, which means it survives automated evaluation. Fixing it requires new performances, since the missing techniques were never recorded.
How Singing Voice is collected
Collection starts with the composition rather than with the microphone.
- Composition and lyric commissioning. Original works are written for the project, with the writers' agreements naming machine learning training as a permitted use.
- Singer recruitment across ranges. Soprano through bass, and the styles the specification needs, screened on a sample recording before a session is booked.
- Capture in a treated room. Dry vocal and accompaniment recorded to separate tracks, with microphone placement held constant across sessions so the corpus does not vary by session.
- Annotation. Pitch extraction or hand correction, lyric alignment, and any structural markup the specification requires.
- Rights assembly and delivery. The composition agreement, the performance release and the annotation, delivered together with the audio.
After the sessions, the recordings are aligned to lyrics, segmented by phrase, annotated for technique and pitch, and packaged with the clearance documentation for each item. The lyric alignment is the step most often skipped, and it is what makes the corpus usable for anything beyond acoustic modelling.
The pilot for singing is a small set of songs recorded across the range the specification names, because the variable that decides the batch is vocal range and it cannot be judged from a single register. A few songs at the top and bottom of the range show whether the recording chain handles the levels and whether the pitch annotation is reliable at the extremes. Both are decided before the run rather than during it.
Acceptance turns on the alignment between the score or the lyrics and the audio, checked on a held-back set. The specification should state whether the alignment is at note level or phrase level before recording starts, because the two require different annotation effort and the difference is not recoverable afterwards. Where the pilot fails, the annotation level is redefined and the pilot run again on the same songs.
How a delivery goes wrong
The defects below are the ones that recur across this category. Each is decided at collection or annotation time, which is why none of them can be repaired in delivery review.
-
Vocal and accompaniment were never separated
If the delivery is mixed music, the singing cannot be isolated without a separation step that introduces its own artifacts. The separation has to exist at capture, not be reconstructed afterwards.
-
Pitch and lyric annotation promised but not verified
Note-level pitch and lyric alignment are the two hardest annotations in this category. A delivery that claims them without a stated agreement measure has usually not checked them.
-
Genre coverage is one style
A corpus of pop singing does not transfer to opera, folk, or rap. Vocal technique, register, and recording treatment differ enough that genre belongs in the specification as a dimension rather than an afterthought.
Why this matters to buyers
Singing synthesis is the obvious driver, and it needs exactly the material that is hardest to obtain: dry vocals with clean rights and accurate pitch annotation. The scarcity here is legal as much as technical.
The second driver is analysis rather than generation. Lyric alignment and pitch extraction models are used in music production tooling, in catalogue processing and in rights administration, where the question is what a recording contains rather than how to make a new one. Those models need annotated audio, and annotated singing is far rarer than annotated speech.
A singing model trained on narrow material is discovered to be narrow in public. The failure is audible to anyone, which means it surfaces in demonstrations rather than in metrics, and by then the model has been shown to customers. The engineering cost of widening it later is a full collection, because the missing techniques are performance data.
The second cost is legal. Rights problems in this category are not usually found by the buyer, they are found by whoever acquires the buyer. A corpus that cannot show a complete chain for every item becomes a liability in diligence, and the usual remedy is to withdraw the model built on it, which costs far more than clearing the rights at collection time.
A singing corpus that passed the pilot gives a model pitch and timing labels it can trust across the range the product needs. The synthesis or transcription work starts from a corpus whose alignment level is stated, so the team knows whether it is building on note-level or phrase-level ground truth rather than discovering it later.
The second gain is rights clarity per performance. Singing carries the composition, the lyrics and the performance as separate rights, and a delivery that documents all three per item can be extended with a second batch from another repertoire without renegotiating the first. The documentation is what makes the corpus a library rather than a one-off.
Specification at a glance
| Category | Singing Voice |
|---|---|
| Group | Speech data |
| Critical spec | Whether the accompaniment is separated (dry vocal) or the mix is kept, and whether MIDI or pitch annotation comes with it. |
| Included by default | Pilot batch, metadata schema, source and consent documentation |
| Pricing | Quoted from specification. No published rates |
Singing Voice across languages
This category is available in every language we cover. The difficulty changes with the language, which is why each language page documents its own constraints.
Related categories
-
Call Center Speech
Customer-service calls, simulated rather than intercepted, capturing both the agent and the caller side, used for call center QC, intent recognition, and dialogue systems.
-
Conversational Speech
Speech from natural conversation between two or more people, on open or semi-structured topics, used for conversational AI, voice assistants, and small talk models.
-
Speech Commands
Targeted recordings of short command words or phrases, usually with many speakers reading each entry several times over, used for wake words and on-device command recognition.
-
Read Speech
Recordings of speakers reading specified text, with clear pronunciation and known text, the base material for TTS and ASR.
-
Podcast Speech
Long-form podcast and interview audio, either solo monologue or two-person conversation, used for long-form speech recognition and speaker modeling.
-
Multilingual Speech
Speech data covering multiple languages within one project, used for multilingual ASR, cross-lingual transfer, and language identification.
Questions about Singing Voice
Can you use existing commercial songs?
No. Melody and lyrics are separate works with separate rights holders, and a corpus of existing tracks cannot be licensed for training at the scale a model needs. Everything in this category is original or cleared explicitly for the purpose.
Do you separate the accompaniment?
Per the specification, and the decision belongs at capture rather than afterward. Separating a mixed recording after the fact leaves artifacts in the vocal stem, and a synthesis model learns whatever is in that stem.
What pitch annotation do you provide?
Extracted pitch tracks by default, with hand correction available. The specification should state which, because the accuracy difference is measurable and the cost difference is real.
Can you cover a specific genre?
Yes, as a recruitment and composition constraint. Tell us the genre and the vocal range, and we will state what the contributor pool looks like before quoting.
Do you have Singing Voice available now?
No. We do not hold inventory. Everything is produced against a specification. That means the timeline starts when the spec is agreed rather than immediately, and it also means the data matches your requirement instead of approximately matching something already built.
Can this be combined with other categories in one delivery?
Yes, and it is common. A single specification can cover several categories across several languages, with one pilot batch and one quality bar. It usually reduces total cost because contributor recruitment and project setup are shared.
What does the consent documentation cover?
Signed contributor consent stating the intended use, the collection methodology, and the annotation guideline the data was produced under. For projects involving EU data subjects the chain is built to satisfy GDPR, including the transfer mechanism. The terms for moving the data across a border.
Request Singing Voice
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.