Choose the right plan to fit your organization's needs.
Our flexible pricing options cater to different business sizes and requirements, ensuring you have access to the right level of voice anonymization technology at a cost-effective price.
Frequently asked questions
Functionality
What is speech anonymisation ?
- Any information that relates to an identified or identifiable person is considered “personal data” according to global privacy regulations such as the GDPR. Speech anonymisation refers to the technique used for removing personal data in audio files. There are two types of personal data in audio files:
- Biometric information: This is the non-verbal attribute of a person’s voice such as the voice quality, timber, etc. that makes their voice distinguishable from others;
- Personally identifiable content: These are the verbal cues such as names, locations, organizations, etc. that reveal personal information.
- Nijta’s solutions handle both types of information to protect the privacy of voice data.
How many languages do you support ?
Our platform supports 90+ languages for the removal of personally identifiable content and 2 languages (English and French) for the removal of biometric information.
What file formats do you support ?
We support all the popular audio and text formats. Please refer to our documentation for a list of supported formats.
Does it work in real-time, for example, streaming audio ?
For now, our solutions do not work in real-time. We plan to release the real-time version of our solution by Q3 2025.
Can you replace the personally identifiable information ?
Yes, the PII redaction feature of our solution is very flexible and supports many entities.
Does the anonymisation work for children's voices ?
No, our solution does not accurately work for children's voices. This is an active area of research at Nijta and we are partnering up with renowned EdTech providers to build a robust solution for children’s voices.
Could the age and gender of the speaker be preserved after anonymisation ?
The gender of the output voices can be controlled, but not the age. We are working actively to provide the age preservation feature.
Could the original emotion of the speaker be preserved after anonymisation ?
We have observed that the original emotion is degraded after anonymisation, but it could be retrieved by fine-tuning the emotion detection model using anonymised voices. We are working on an anonymization technique that will preserve the original emotion with high-fidelity.
Could the non-verbal cues such as the speaking pace, pronunciation, intonation, etc. of the original speaker be preserved after anonymisation ?
Yes, such non-verbal cues are largely preserved under certain conditions. We sometimes notice small degradation in the pronunciation.
Do the speech biomarkers related to the health attributes of the original speaker be preserved after anonymisation ?
We cannot say that with certainty. We are currently working on R&D projects to preserve the speech biomarkers.
Can the anonymisation filter profane language ?
This is an active area of research. We are working with a large group to filter profane language in live calls.
Performance
What is the accuracy of your solution ?
We have observed significant improvements in privacy benchmarks after anonymisation. The probability of singling out and linkability which refer to the legal guarantees of anonymisation are improved x% and y% relative to unprotected audio. The error rate of PII redaction is less than 1%.
What is the processing time of your solution ?
- We have evaluated the speed, reliability, scalability, and efficiency of our system on this config: 15 cores, 45 GB RAM, 1 x Tesla V100S.
- We were able to anonymize 1 hour in 4 min with 1 worker processing batches of 30 files (voice only).
What is the maximum size of audio files that could be sent to the API ?
Audio files should not exceed 500 MB in size. The API will not accept files larger than this limit.
How many concurrent requests could be processed by the API without degrading the processing time ?
The system supports concurrent requests, allowing a maximum of 10 simultaneous requests per user.
Can the customer fine-tune the models hosted on their site ?
No, we do not offer fine-tuning service of our models on customer’s site.
Installation
Does it work as a SaaS or on-premise ?
We provide both SaaS and on-premise solutions.
What are the computational requirements for hosting the on-premise solution ?
- Operating system : Linux (Ubuntu 20.04 or later)
- Processor : Intel Xeon Gold 6226R or later
- RAM : 32 GB minimum
- Disk space : 100 GB minimum
- GPU : Nvidia Tesla V100S - 32 GB minimum
What are the concrete measures followed to ensure the security of the SaaS solution ?
We follow SOC2 compliance measures to ensure the security of our hosted SaaS solution. Please refer to our documentation for more information.