Voice Recordings and Cloning
How are audio recordings secured and which are the security safeguards in place in these collecting tools (e.g. what encryption?)
We have a range of encryption standards - all listed here.
We use AES-256 and at rest encryption. All of our endpoints are HTTPS. We also use VPC’s that are restricted from the internet.
What metadata are collected when participants submit audio samples? Where and how are those meta-data stored?
We collect only the device info we get through the browser.
This metadata is stored in an S3 bucket as metadata of S3 object which is the submitted audio file.
Metadata includes the device type (e.g. MacBook Pro) or audio interface.
Should the initiative not be offered to minors, are there any safeguards to ensure age restriction?
We do not recommend offering voice cloning to minors. Commercially available children and young peoples' voices go through rigorous processes before being included in the AudioStack library to ensure compliance.
For customers running open data collection projects for voice cloning purposes, our content moderation team will regularly check samples come through and then remove any age restricted submissions.
Data Quality & Pre-processing
What process is performed on the raw voice samples to prepare the training and test data for cloning?
AudioStack has a range of capabilities when it comes to audio processing. We may apply processes such as denoising and normalisation to get a higher quality clone when appropriate.
How can we make sure that audio being used in our project doesn't contain any offensive language or hate speech?
The AudioStack API can generate transcripts to identify what is being said in uploaded speech files if this is needed.
How many different files/formats of voice samples can be stored?
We store audio data generally in mp3, ogg and wav format (though AudioStack supports a wide range of file uploads).
Training and models
Is the AI model AudioStack uses for cloning created in-house or obtained from a third party?
We integrate with multiple providers (Eleven Labs, internal models, PlayHT, Resemble). All of these providers have a DPA signed with us, and we have contractual terms with them to ensure data deletion when it is no longer required.
Would AudioStack be able to explain how the model has created a particular voice?
We’d be able to explain the underlying principles of the models and we’d also be able to do further analysis to do some support in terms of explainability. Please contact support@audiostack.ai
Management and Deletion of Data
Who will have access to our training data, audio recordings, and AI Voice? Both internally within AudioStack and others, e.g. sub-processors?
Certain trained members of the staff have access in order to deliver our services effectively.
We have agreements with subprocessors that data will not be stored on their servers.
You can find all of our subprocessors listed here.
What is the retention policy for training data once AI model is trained. Is the information retained for maintenance, re-training, improvements etc.?
For the duration of the project, we will retain uploaded / created data
90 days after the end of the project we will delete data, as much is technically possible.
How would data subject rights be exercised?
Simply email our customer support team and we’ll delete any data we hold. Our email address is support@audiostack.ai
What is the service providers practice and approach towards such requests made against output data?
Service providers have agreements with us to delete if necessary.
EU AI Act
How does the AI Act apply to AudioStack?
We are an AI Deployer because we deploy AI systems whether provided by third parties or internal IP, under our own brand and trademark.
We run a range of evaluation metrics and auditing software. Where we provide voice cloning services.
We inform users that the results are generated by AI through a transparency guide and our documentation.
We also refer to AI throughout our documentation and have technical documentation of how we build our services.
We have an AI ethics policy, and policies for conforming with copyright law. We don’t train general purpose AI models such as foundation models.
We’ve conducted risk analysis for our models - and the biggest risk is in regards voice cloning - which could lead to impersonation.
We mitigate this by having a secure and robust enterprise level system and we don’t allow sharing of content without consent.
We also collect consent documents from all voice actors affected and they are remunerated for their voices. We are constantly evaluating this setup and investing more in our AI governance and AI safety.
IT Compliance
Are you SOC2 compliant?
Yes. You can read our SOC2 type 2 report here
Are you GDPR compliant?
Yes we follow GDPR and are registered with the ICO in the UK.
Music Licensing
Is it safe to use the music licenses that you have?
We have simple, enterprise-ready licensing deals with various music providers and are committed to working closely with them.
For our enterprise customers, all music in the AudioStack library can be used for broadcast purposes, subject to the terms of your contract. Find out more here.