How to Evaluate a Vocal Data Vendor: 12 Questions to Ask Before You Sign

You have decided to license a vocal dataset rather than build one. Now you have to choose a vendor. The market has grown enough that there are multiple vendors to evaluate, but it has not grown enough for the vendor landscape to be well understood or standardized. Each vendor will describe their offering in their own terms, each will downplay the weaknesses you are trying to uncover, and each will pressure you to move fast.

This post is a 12-question framework for evaluating vocal data vendors. It is designed so that the answers let you compare vendors against each other on the dimensions that actually matter. Use it as a structured screening call or as a written RFQ. Either way, the goal is to force each vendor to address the same questions in comparable terms.

The 12 questions

The questions are grouped into three categories: clearance and legal, technical quality, and commercial structure. Each category has four questions.

Category A: Clearance and legal

1. Can you walk me through the chain of consent for a typical recording in your catalog?

What you are looking for: a clear narrative. "The vocalist signed an agreement on date X via mechanism Y, which granted rights Z. The recording was added to the dataset on date A with consent ID B. The consent record is retrievable on demand." If the vendor cannot produce a coherent story at this level of specificity, the chain of consent is probably informal or incomplete.

Red flag answer: "All our recordings come from contributors who signed a standard agreement."

Green flag answer: "Let me show you a sample consent record. Here is the timestamp, here is the exact language shown, here is the version of the agreement in effect at the time, and here is how the vocalist can withdraw if they want to."

2. Do your vocalist agreements explicitly grant AI training rights, and when was the current version of the agreement last updated?

What you are looking for: explicit language referencing AI training, machine learning, or generative models. Agreements drafted before 2023 often do not contemplate AI training and may require amendments. Confirm the current version date.

Red flag answer: "The agreements cover commercial use."

Green flag answer: "The current version is 3.2, dated November 2025. It includes a specific Section 4 covering AI training, model derivation, sublicensing, and withdrawal. I can send you the template."

3. What is your withdrawal process, and how does it propagate to enterprise buyers?

What you are looking for: a defined process, a defined timeline, and a contractual propagation mechanism. Withdrawal is a GDPR requirement and cannot be ignored. A vendor who has not thought through this has not thought through their compliance posture.

Red flag answer: "We don't get withdrawal requests often."

Green flag answer: "Vocalists can withdraw via their account dashboard. We are contractually required to notify enterprise buyers within 30 days. Buyers are required to remove the affected recordings from active training sets and document the model-impact decision within 60 days. We audit this annually."

4. What does the indemnification clause in your licensing agreement look like?

What you are looking for: meaningful protection against third-party IP claims arising from the use of the dataset, with a defined cap that is proportional to the license fee. The vendor should indemnify you for claims alleging their clearance process was deficient.

Red flag answer: "We do not provide indemnification."

Green flag answer: "We indemnify you against third-party IP claims arising from the use of the dataset, up to [cap], subject to standard carve-outs for your modifications to the data. The cap is proportional to the license fee and scales with larger licenses."

Category B: Technical quality

5. What is the source of the vocal recordings in your catalog?

What you are looking for: recordings captured directly from vocalists in studio environments. The goal is to confirm that the audio was never extracted from a full mix using source separation.

Red flag answer: "We use a combination of sources."

Green flag answer: "Every recording in our catalog was captured directly by a vocalist singing into a microphone in a studio. We do not use source-separated audio. I can show you the session metadata if helpful."

6. What are the technical specifications: sample rate, bit depth, isolation, and processing state?

What you are looking for: 44.1 kHz or 48 kHz, 24-bit, genuinely isolated recordings, availability of dry and wet versions. A vendor who cannot answer this quickly does not know their own catalog.

Red flag answer: "High quality studio recordings."

Green flag answer: "All recordings are 44.1 kHz, 24-bit, uncompressed WAV. Dry and wet versions are available for every recording. The noise floor is below -65 dBFS. I can send you spec files with per-recording measurements."

7. What metadata is included with each recording, and how is it generated?

What you are looking for: the 14 standard metadata fields covered in a separate post, plus clarity on whether metadata is hand-annotated or automated. Both can be fine; the important thing is knowing which.

Red flag answer: "We include full metadata."

Green flag answer: "Each recording has genre, BPM, key, vocalist ID, gender, vocal type, language, phoneme alignment, F0 contour, lyrics, sample rate, bit depth, consent ID, and license scope. Genre and BPM are hand-verified. Phoneme alignment uses Montreal Forced Aligner with a hand-correction pass on a 10% sample. F0 is extracted with RMVPE."

8. Can you send me a representative sample for technical evaluation?

What you are looking for: a willingness to share a non-trivial sample (typically 5-20 recordings covering multiple genres, vocalists, and quality tiers) under an NDA. A vendor who refuses to share samples or only shares heavily curated cherry-picked examples is hiding something.

Red flag answer: "We can share samples after we have a signed agreement."

Green flag answer: "Yes, we send sample packages under a short NDA. You can choose which genres or languages you want represented. I can send the sample within two business days of the NDA."

Category C: Commercial structure

9. How is pricing structured?

What you are looking for: pricing that reflects the scope of use, the exclusivity, and the size of the catalog subset being licensed. A vendor with flat pricing is not thinking about your use case; a vendor with custom pricing is.

Red flag answer: "$X per recording."

Green flag answer: "Pricing is based on the scope of your use (fine-tuning vs full training), the size of the catalog subset (all vs filtered), exclusivity (non-exclusive, semi-exclusive, exclusive), and the term. A typical fine-tuning license on a subset is in the range of X, a full training license on the complete catalog is in the range of Y."

10. What is the term of the license and what happens when it expires?

What you are looking for: clear language about license duration, renewal, and post-expiration use. Specifically: are you permitted to continue using the trained model after the license expires, even if you cannot train new models on the data?

Red flag answer: "Standard terms apply."

Green flag answer: "The license is typically 2 years, renewable. During the term, you can use the dataset for training. After expiration, you can continue to deploy and serve models that were trained during the term, but you cannot train new models on the dataset without renewal."

11. Is the license exclusive, semi-exclusive, or non-exclusive, and what is the pricing difference?

What you are looking for: clarity on which other companies can also license the same dataset. Non-exclusive licenses are cheapest and most common. Exclusive licenses are expensive and often unrealistic for a dataset that is still being actively built. Semi-exclusive (vendor agrees not to license to direct competitors) is a useful middle ground.

Red flag answer: "We do not discuss other licensees."

Green flag answer: "Standard licenses are non-exclusive. Semi-exclusive options are available, where we agree not to license to a defined list of your direct competitors for the term, typically at a 50-100% price premium. True exclusivity is available in limited cases but is typically 3-5x the non-exclusive price."

12. What happens if a vocalist in the catalog sues you? Who bears the legal cost?

What you are looking for: a clear statement about defense costs, indemnification scope, and whether the licensee is protected if the upstream vendor relationship breaks down.

Red flag answer: "That has never happened."

Green flag answer: "The indemnification clause specifically covers defense costs and damages arising from claims by contributors alleging defective consent. The obligation survives the licensing term for claims arising from your use during the term. The cap is X. We have insurance coverage of Y."

How to use the framework

The framework works best as a structured vendor comparison. Run each shortlisted vendor through all 12 questions, ideally in writing, and score their answers on a simple 1-to-5 scale. After two or three vendors, the differences become visible.

What this framework catches

The framework is specifically designed to catch vendors who market aggressively but have not invested in the underlying operation. The most common failure modes:

• No per-recording audit trail. Vendor can describe the dataset but not produce a specific consent record.
• Pre-2023 agreements. Vendor built the catalog years ago, never amended the contracts, and cannot actually grant AI training rights.
• Separated stems marketed as "isolated." Vendor uses source separation to extract vocals from mixed audio and calls the results isolated stems.
• No indemnification. Vendor transfers all legal risk to the buyer.
• Unclear withdrawal propagation. Vendor cannot describe what happens when a vocalist withdraws.

Any of these is disqualifying for an enterprise deal. All of them are common enough that you will catch at least one during a vendor evaluation process.

How The Vocal Market answers the 12 questions

If you are running this evaluation and want a calibration point, here is how we answer each question.

1. Chain of consent: Every vocalist signs a timestamped digital agreement via their marketplace account. The consent record includes user ID, timestamp, IP address, agreement version, and exact consent language. Records are immutable and retrievable on demand.
2. Explicit AI rights: Current agreement version is 3.2, updated November 2025. Section 4 explicitly covers AI training, model derivation, sublicensing to enterprise buyers, and the specific scope of permitted uses.
3. Withdrawal: Vocalists can withdraw via their account. We notify enterprise buyers within 30 days. Buyers remove the affected recordings from active training within 60 days and document model-impact decisions.
4. Indemnification: We indemnify enterprise buyers against third-party IP claims arising from the use of the dataset, with a cap proportional to the license fee and standard carve-outs for buyer modifications.
5. Source: Every recording is captured by a vocalist singing into a microphone in a studio. No source separation. Session metadata is available.
6. Specs: 44.1 kHz, 24-bit, uncompressed WAV. Dry and wet versions of every recording. Noise floor below -65 dBFS.
7. Metadata: 14 fields per recording. Genre, BPM, and key are hand-verified. Alignment is automated with a quality-check pass.
8. Samples: Yes, we send samples within 2 business days of a signed NDA, with choice of genres and languages.
9. Pricing: Structured by scope, catalog subset, exclusivity, and term. No flat pricing.
10. Term: Typically 2 years, renewable. Post-expiration, you can continue to serve models trained during the term but cannot train new models.
11. Exclusivity: Non-exclusive standard. Semi-exclusive available at a premium. True exclusivity case-by-case.
12. Legal defense: Defense costs and damages from contributor claims covered by the indemnification clause, with insurance backing.

If you want to put our answers directly alongside other vendors' answers in a scorecard format, request a sample dataset and ask for the vendor scorecard kit. We will send it in a format that maps cleanly onto the 12-question framework above so you can compare apples to apples.