Speech intelligence / Research guide

Consented voice-conversion datasets and evaluation

Consented voice-conversion datasets: source and target recordings, paired versus unpaired designs, identity permissions, benchmarks and reviewable project budgets.

FIELD NOTESBY FIUND
01Signal02Transcript03Reference

Audio · language · evaluation

Cited references
6 cited sources
Inside this guide
1 worked example
Reading time
9 minutes
Reviewed
2026-11-12

Prepared by the Fiund research desk from the cited public documentation. Technical examples describe their evidence and limits. Research method & editorial contact · Suggest a correction

Reading tools
The takeaway

Consented voice-conversion datasets organize source speech, target-voice material and the authorization needed for the proposed identity transformation. Voice conversion changes an existing speech signal, unlike ordinary text-to-speech generation from text. Useful collections distinguish linguistic content, speaker identity, expression and recording conditions so a model can be evaluated without confusing those factors. Public benchmarks can support technical comparison, but access does not automatically authorize every commercial voice use. Fiund can discuss a bounded source and review plan; availability, permissions, preparation and terms are confirmed for selected recordings and intended behavior.

Core task
Transform an existing speech signal toward a target voice representation
Source layers
Source recordings, target recordings and permitted identity relationships
Design choice
Parallel, semi-parallel or nonparallel material
Evaluation axes
Content preservation, similarity, naturalness and source artifacts
Commercial basis
Explicit source-specific use and authorization review

Your reading path

Start with the right questions.

  1. 01Voice conversion starts from speech, not only text
  2. 02Public challenge evidence
  3. 03Parallel, semi-parallel and nonparallel designs

From principle to practice

What a source/target manifest should contain

Field groupSuggested contentWhat it establishes
Open the worked example

Voice conversion starts from speech, not only text

A voice-conversion system receives an existing speech signal and changes a target property such as voice identity or style. The source performance can carry timing, pronunciation and expression that a text-only input would not contain. A dataset should describe which properties are meant to change and which should remain.

This guide concentrates on source/target identity relationships and permission-aware evaluation. Expressive TTS has a different input and control problem. A collection can potentially support both tasks, but the proposal should not imply that one training permission or one technical design covers every downstream behavior.

Public challenge evidence

The Voice Conversion Challenge 2020 site defines same-language and cross-language conditions with common source and target material. Its task design separates shared and unshared sentences, illustrating why parallelism is a deliberate dataset property rather than a generic quality label.

The challenge establishes a technical research use, not a universal commercial grant for a person’s identity. Its source files and current access conditions need direct review. The fact that several systems can be compared on common data does not establish that a particular voice may be deployed for every product or redistributed as a new dataset.

Parallel, semi-parallel and nonparallel designs

Parallel material uses matched linguistic content across source and target speakers. Nonparallel material uses different content, while semi-parallel designs share only part of the text. These arrangements affect what a model can learn and how its evaluation should be constructed.

Decide whether matching text, timing or session conditions is necessary. A larger nonparallel archive is not automatically a substitute for a smaller paired design. Conversely, a narrow matched script can conceal generalization limits. Preserve text-family and recording relationships so held-out tests can distinguish new content from repeated phrases.

Consented is a scope claim, not a decorative label

The proposed use should identify the relevant source recordings, target voice, permitted transformations and downstream access. A person can permit one use without authorizing every possible generation or transfer. Keep the reviewed scope connected to the collection rather than relying on a generic website statement.

Authorization review should establish who can grant the proposed rights and which participants or artifacts it covers. Recording access, corpus licensing and identity-related uses can differ. A public guide cannot determine the status of an unseen archive, and a technical benchmark should not be treated as a blanket answer.

A public licensed-voice example

ElevenLabs’ Hasbro announcement describes an authorized workflow involving character owners and talent. Financial consideration is undisclosed. It is a voice-use partnership, not a published price for a bulk conversion-training corpus.

ElevenLabs’ creator report describes more than USD 22 million in cumulative platform creator earnings by May 2026. That is aggregate usage remuneration, not a single dataset transaction or guaranteed supplier income. These examples illustrate different commercial structures without establishing a general price for voice-conversion data.

Useful corpora and their boundaries

LibriTTS and its restored version provide read-speech research material with documented corpus conditions. They are useful comparisons for source quality and text pairing, but a corpus’s availability is not itself an unrestricted invitation to deploy every recorded identity.

The VCC2020 database repository is a more direct technical reference for conversion tasks. Review its release, source relationships and intended use. Keep original and processed variants connected; restored or re-encoded audio is not additional independent speech from new people.

What a source/target manifest should contain

The buyer should be able to reconstruct which recordings and identities participate in each permitted conversion relation. Personal contact information is usually unnecessary for a public technical manifest.

Compare the evidence
Field groupSuggested contentWhat it establishes
Recordingutterance_id, speaker group, session_id, versionIdentifies the actual source performance
Rolesource, target, reference or evaluation-onlyDefines how material is used
Text relationtranscript, text-family, parallel statusShows shared versus held-out content
Language/stylesource-supported labels and review statusSeparates identity from other factors
Acoustic conditionscapture setup, original format, processingExposes channel and preparation effects
Authorized relationscope reference and exclusionsLinks the proposed transformation with review
Partitionspeaker/session/text groupsSupports the intended independent test

Separate identity from recording conditions

A target speaker recorded in one room and a source speaker recorded through a phone can create a strong channel difference. A system may learn some of that difference instead of the intended identity property. Document the source setup and decide which conditions should be matched or deliberately varied.

Keep original capture properties apart from delivered format. Resampling, restoration and denoising can change a signal without creating a new performance. If preparation is needed, identify the operation and retain the parent relationship. An acoustic similarity result should not be interpreted without knowing what source conditions were compared.

Content preservation deserves its own reference

A conversion can sound similar to a target while changing words, dropping a short phrase or altering a number. If the application requires content preservation, specify how that is reviewed separately from perceived similarity and naturalness.

Use text and listening evidence appropriate to the task, and keep uncertain regions visible. A recognition system’s output can help flag issues but is not automatically a definitive reference. Do not let a favorable identity score conceal a content failure. The dataset should support independent analysis of these axes rather than only one combined rating.

Cross-language conversion needs a clear question

A target voice may not have been recorded in the source language. That is a different condition from same-language conversion and can confound identity, pronunciation and linguistic content. State what the model is allowed to see and what the listener or metric is asked to judge.

Keep language-specific test composition visible. A broad multilingual count does not establish balanced evidence for every conversion direction. If the source cannot support a required relation, narrow the claim or collect additional material rather than presenting a missing condition as already covered.

Evaluate similarity, naturalness and usefulness separately

Perceived speaker similarity, speech quality and task usefulness are related but not identical. A clean output may fail to resemble the intended voice; a similar output may contain artifacts or incorrect content. Define each target before choosing a review method.

If listening tests are part of acceptance, document the reference examples, prompts and reviewer procedure. A file can decode successfully without having been listened to or validated. Keep technical checks distinct from audible assessment, and report only the actual review performed. No corpus purchase can guarantee a future model’s score.

Provider roles and commercial access

A benchmark maintainer supplies a bounded research resource. A voice owner or authorized representative can discuss identity-specific uses. A recording studio can create new performances but may not control every downstream right. A generation marketplace can license a built voice without transferring its source training corpus.

Compare the actual transaction and deliverable. Ask whether the proposal covers recordings, a model, a service or a restricted usage license. A user’s ability to generate speech through an account does not automatically authorize redistribution of underlying data. Preserve source-specific limitations instead of merging these products into one price comparison.

Pricing without conflating payouts and corpus rights

No reliable public market average for consented voice-conversion training collections was established here. Price can depend on the target identity, source material, allowed use, recording design, preparation and review. Aggregate creator payouts or a partnership announcement do not supply a per-hour corpus rate.

For a hypothetical pilot involving three permitted source/target relationships, invented placeholders could allocate USD 2,400 to scoped recording/authorization, USD 1,200 to text and source preparation, USD 1,200 to evaluation references and USD 900 to independent acceptance: USD 5,700 total. These are assumptions, not observed supplier prices or a Fiund quote.

Pilot acceptance and held-out design

Inspect matched and unmatched text where relevant, a difficult name or number, different expressive delivery and a source condition unlike the target capture. Confirm which source/target relations are within scope and which material is evaluation-only. Keep ambiguous or excluded cases explicit.

Group related recordings and derivatives according to the experiment. A speaker-held-out test is not the same as new-text evaluation on familiar voices. If shared sentences exist, record them rather than assuming content independence. The acceptance report should identify inspected examples and review layers, not imply an exhaustive archive audit.

Reference roles and scope matrices

An evaluation reference can be permitted for comparison without being part of the training material. Mark those roles in the collection rather than allowing every delivered recording to flow into the same pipeline. A source/target relationship table should identify which transformations are within the reviewed project scope and which remain excluded.

This matrix also helps prevent experimental leakage. A target reference used during model selection should not quietly become a final held-out sample. Keep preparation versions and the purpose of each recording visible. If the project changes from research evaluation to a deployed identity-specific product, revisit the actual source scope instead of assuming the original inquiry settled it. Those are project-design and source-review questions, not permission that can be inferred from a speaker name or a public download.

Discuss a voice-conversion brief with Fiund

Email hello@fiund.com with the intended transformation, languages, source/target structure, recording design and permitted product use. State whether the collection needs parallel content, existing archives or newly recorded performances, and what reference and authorization evidence is required.

Fiund can discuss candidate source fit, handling and a bounded pilot. Availability, permission, preparation, pricing and delivery are confirmed for selected material. This research did not establish a largest publicly priced voice-conversion corpus sale; software usage, creator remuneration and broader IP partnerships remain different economic measures.

Related sourcing decisions

Expressive TTS datasets: prosody, style and controllable speech — Expressive TTS data guide covering prosody, style labels, benchmark limits, public commercial-license pricing, voice permissions and source-specific sample acceptance.

Accented-speech evaluation datasets — Accented-speech evaluation data: supported coverage labels, benchmark limits, channel confounders, independent references and sourcing budgets.

Questions buyers ask

Is voice conversion the same as TTS?

No. Conversion starts from an existing speech signal; ordinary TTS starts from text. Their source pairings and evaluation requirements can differ.

What does parallel mean?

It means source and target material share linguistic content under a defined design. It does not automatically mean sample-level alignment or unrestricted permission.

Does benchmark access authorize every commercial voice use?

Do not assume that. Review the actual recordings, participants, scope and intended transformation rather than treating access as a blanket grant.

Are creator payouts a dataset price?

No. Usage-based aggregate earnings are different from the price of a particular training-corpus license.

What should a Fiund proposal resolve?

The selected source/target material, permitted relations, preparation, reference design, acceptance and commercial terms. A category page does not guarantee a named voice or model result.

Sources & research notes

Public documentation reviewed 2026-11-12. Provider descriptions and published terms are attributed below. Published facts are attributed to their sources. Worked files are labeled as illustrative or as an inspected editorial source; neither implies commercial dataset clearance.

  1. OpenSLR LibriTTSwww.openslr.org · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

  2. OpenSLR LibriTTS-Rwww.openslr.org · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

  3. ElevenLabs Hasbro voice partnershipelevenlabs.io · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

  4. ElevenLabs creator payout reportelevenlabs.io · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

  5. Voice Conversion Challenge2020 original sitevc-challenge.org · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

  6. VCC2020 original databasegithub.com · Reviewed 2026-11-12

    Primary source reviewed November 12, 2026; check the current release and project-specific terms.

Sourcing with Fiund

Turn the requirements into a sourcing brief.

Send the model task, required source characteristics, target volume, annotations, intended use and review criteria. Fiund can assess sourcing options and the preparation needed. Availability, source documentation, permitted uses and pricing are confirmed for the proposed project.

Discuss consented voice-conversion datasets and evaluation ↗

Build with AIPodcast

Start with the data
your model needs.

Share your use case. Get a focused next step.

Request a sample