Ethics & Transparency

We push for change.

At ExpressiveLabs, we believe it's our responsibility to be transparent about how and why we use Artificial Intelligence (AI) technology in our products, and what measures we take to ensure this happens in an ethical way. On this page, we explain what we view as "ethical AI", and how we apply these ideas to the way we work.

hrs of training data

hours trained

kWh total power use

models trained

voice provider payout

Quick overview

We use artificial intelligence technologies in our products to make them easy to use and generate high-quality audio. These technologies will never replace any human part of the creative process within our software.

Everything we use to create our AI voice models is given to us by voice actors and singers who signed a contract and give us permission to use their voice. We pay them fairly for their work and inform them of everything we do with their recordings.

Background

What is AI?

Artificial Intelligence, AI for short, is a broad term to describe software that performs its tasks through learning from data instead of following direct instructions. These technologies rely on large amounts of data to look for patterns and generate new output based on those patterns.

We only work on a tiny subset of AI technologies. Our areas of interest include linguistic processing and audio synthesis. We use relatively simple neural networks to complete these tasks faster or with better results compared to classic approaches.

Accountability

The numbers

We train all voice libraries together at once, saving on resources and using our hardware more efficiently. We monitor energy consumption closely and publish usage reports on every training session we run.

Latest model runs

Model	Duration	Energy used (Wh)	Date

Fair & Square

Data sourcing

We take pride in transparency and fair data usage. All the voice data we use in training our voicebanks is recorded especially for Mikoto Studio by professional voice actors and singers. These voice providers are paid for their work through a direct revenue-sharing agreement, letting them keep at least 15% of all product earnings. We also provide them with coaching and equipment for the job.

Besides voice libraries, we also develop a number of extra models used in the singing synthesis process. For these models, we rely on publicly available datasets that are licensed explicitly for this purpose. Scroll down to find a full, up-to-date list of all datasets that we use.

Artist-first

Respecting creativity

We put artists first in everything we do. All our tools are designed to help people express their creative vision, not replace it.

We do not believe in AI models that do all the work for you. Instead, we limit the abilities of our models to purely supportive roles. Converting natural language to linguistic information, procedurally generating high quality audio using a fine-tuned vocoder, or detecting legacy configuration types to help new users get started with ease.

We do not collect any user data to train our AI models — not now, not ever. All data we use comes from consenting, paid voice providers and publicly available datasets.

No generative modelsNo user data collectionConsented data onlyArtists stay in control

Sustainability

Environmental impact

Training AI models is an intensive process that can consume a lot of electricity. We are aware of our impact on the environment and do everything we can to minimize it. We train our models locally and get our energy from a verifiable green energy provider.

Our energy provider guarantees that all electricity they deliver is fully covered by renewable energy sources. Greenchoice's offering consists of 100% wind and solar energy produced exclusively in the Netherlands. For up-to-date information, visit their website.

On top of taking action ourselves, we are also a supporter of Trees For All. This local non-profit plants trees both within the Netherlands and around the world. We pledge to always reflect our total energy consumption in our donations. You can find the certificate of our latest donation here.

Compliance

Third-party developers

We work with an ever-growing number of partner companies who develop voice libraries for the Mikoto platform. We make sure these partners adhere to the same ethical standards we set for ourselves, both before and during development. We set guidelines for data sourcing and voice provider compensation, and we perform all model training on our own hardware.

Third-party developers are separate entities from ExpressiveLabs. See something that doesn't look right? Please get in touch with us via e-mail so we can look into it.

Memberships

Initiatives

We are proud members of a growing number of initiatives to promote fair and sustainable AI development.

AI For Music

Full transparency

Data sourcing directory

As part of our commitment to transparency, we maintain an up-to-date list of all datasets used in both our commercial products as well as our research projects. To protect innovative research, we may delay updating the list of datasets used in research projects by up to 6 months. The list was last updated on November 23, 2025.

Collection	Dataset	Projects
Commercial projects	LibriSpeech	Duration model
	CommonVoice	Duration model
	WikiPron (© 2019 Kyle Gorman, Jackson Lee, and contributors)	Grapheme-to-phoneme model, phoneme frequency model
Internal research dataset ¹	Tohoku Kiritan (© SSS)	Acoustic model research
	Kasane Teto (© TWINDRILL)	Acoustic model research (2022–2023)
	Namine Ritsu (© canon)	Acoustic model research
Voice provider data	Amarynn	Sayaka, vocoder model, duration model, pitch prediction model, Indonesian XLS support, Mandarin XLS support
	Jase	Jinshi, vocoder model, duration model, pitch prediction model, Spanish XLS support, Korean XLS support
	Yuzuyama	Sorane Mako 3510, vocoder model, duration model, pitch prediction model
	MystSaphyr	Kikyuune Aiko 3510, vocoder model, duration model, pitch prediction model
Individual data donations ²	asin_dengklang	Vocoder model, duration model, pitch prediction model
	AxelCode	Vocoder model, duration model, pitch prediction model, Italian XLS research
	GeneralNuisance	Vocoder model, duration model, pitch prediction model
	Ocean Eyes	Vocoder model, duration model, pitch prediction model
	PixPrucer	Vocoder model, duration model, pitch prediction model, Polish XLS support
	redfish	Vocoder model, duration model, pitch prediction model
	UtaUtaUtau	Vocoder model, duration model, pitch prediction model, Tagalog XLS research
	Wouter Besse	Vocoder model, duration model, pitch prediction model, Dutch XLS support

ExpressiveLabs

Legal API Support Ethics Talent Acquisition