'Whose Voice is that Synthesized Audio?' Release of Demo Version of 'Paramatch', a Speaker Identification Model that Identifies the Original Voice Actor from Unauthorized AI Voices
Parakeet Inc. has released a free demo of 'Paramatch' on Hugging Face. It is a speaker identification AI that estimates the original voice actor from AI-synthesized voices trained without permission to protect voice actors' rights.
📋 Article Processing Timeline
- 📰 Published: April 14, 2026 at 18:50
- 🔍 Collected: April 14, 2026 at 10:31
- 🤖 AI Analyzed: April 19, 2026 at 17:42 (127h 10m after Collected)
Estimating the original voice actor used for training from uploaded audio
Parakeet Inc. (Headquarters: Hirono-machi, Futaba-gun, Fukushima Prefecture; CEO: Yasutaka Nakamura; hereinafter "the Company") has released a free demo version of its speaker identification model "Paramatch" on Hugging Face. This model is a speaker identification AI that predicts and analyzes the speaker with the closest voice quality when an audio file is uploaded.
It was developed with the aim of obtaining clues to identify the owner of the original voice from audio outputted by AI voice synthesis models trained without permission.
▼ Demo URL
https://huggingface.co/spaces/Parakeet-Inc/paramatch
■ Background of Development: To Protect the Identity of "Voice"
In recent years, with the rapid development of AI voice synthesis technology, there has been a continuous stream of cases where models trained without permission on the voices of voice actors, or synthesized voices generated using them, are published and diffused on SNS and video platforms.
What is important here is to accurately grasp the essence of the problem. Under Article 30-4 of the Japanese Copyright Act, training for the purpose of information analysis by AI itself is widely permitted. In principle, there are no legal issues with an individual training audio for their own enjoyment and using it solely by themselves.
However, distributing or publishing the unauthorizedly trained AI voice synthesis model itself, providing a third party with a state where they can freely speak in a specific voice actor's voice, or publishing and diffusing the synthesized voice generated by that voice actor's voice is a completely different issue. These actions threaten the identity of the "voice" built up by voice actors and actors through years of study, and significantly infringe upon performers' rights.
A voice is the one and only proof of existence for a performer. To be made to speak words without one's knowledge, separated from one's intention—in response to this situation, we, as people involved in the research and development of audio technology, have continuously asked ourselves what we can do. One answer to that is Paramatch.
■ What is Paramatch?
Paramatch is a speaker identification AI model developed by our company. For an uploaded audio file, it predicts the top 3 speakers with the closest voice quality from the speaker database and presents the similarity of each as a score. This allows one to obtain objective clues to identify "whose voice was the basis for generating" a certain synthesized voice.
[Main Features]
- Predicts similar speakers from the speaker database
- Supports audio files of 30 seconds or less
Uploaded audio is not saved on the server, and no secondary use is made whatsoever. In addition, training data is used only to build the Paramatch model. We do not divert it to generative models like voice synthesis or other products and projects of our company.
Please note that the results of this tool are for reference only and do not serve as a basis for legal judgments. Furthermore, the recognition accuracy is not perfect, and we will continue to make improvements.
There are two main purposes for releasing this demo.
First, it is to verify, on an empirical basis, an effective countermeasure to protect the rights of the "voice," which is the identity of the performer. Rather than armchair discussions, we want to confirm with industry stakeholders how effective this technology is against actual cases of unauthorized use.
Second, it is to serve as a starting point for dialogue with voice actors, actors, their agencies, and rights-holder organizations. We are a company developing AI audio technology, but precisely because of that, we cannot turn a blind eye to the reality that this technology is being misused. As the responsibility of those creating the technology, we want to seriously work on technologies to protect voices as well.
■ To Voice Actors, Actors, and Industry Organizations: Seeking Partners to Work Together
We do not intend to let Paramatch end simply as a "public demo." To seriously implement this technology in society, the knowledge of voice actors and actors who face their "voice" every day on site, as well as the agencies and industry organizations that protect their rights, is indispensable.
Specifically, we would like to request cooperation in forms such as the following:
- Verification of Paramatch's effectiveness against actual cases of unauthorized use
- Consultations to improve the accuracy of the speaker database
- Discussions on industry guidelines and rights
Parakeet Inc. (Headquarters: Hirono-machi, Futaba-gun, Fukushima Prefecture; CEO: Yasutaka Nakamura; hereinafter "the Company") has released a free demo version of its speaker identification model "Paramatch" on Hugging Face. This model is a speaker identification AI that predicts and analyzes the speaker with the closest voice quality when an audio file is uploaded.
It was developed with the aim of obtaining clues to identify the owner of the original voice from audio outputted by AI voice synthesis models trained without permission.
▼ Demo URL
https://huggingface.co/spaces/Parakeet-Inc/paramatch
■ Background of Development: To Protect the Identity of "Voice"
In recent years, with the rapid development of AI voice synthesis technology, there has been a continuous stream of cases where models trained without permission on the voices of voice actors, or synthesized voices generated using them, are published and diffused on SNS and video platforms.
What is important here is to accurately grasp the essence of the problem. Under Article 30-4 of the Japanese Copyright Act, training for the purpose of information analysis by AI itself is widely permitted. In principle, there are no legal issues with an individual training audio for their own enjoyment and using it solely by themselves.
However, distributing or publishing the unauthorizedly trained AI voice synthesis model itself, providing a third party with a state where they can freely speak in a specific voice actor's voice, or publishing and diffusing the synthesized voice generated by that voice actor's voice is a completely different issue. These actions threaten the identity of the "voice" built up by voice actors and actors through years of study, and significantly infringe upon performers' rights.
A voice is the one and only proof of existence for a performer. To be made to speak words without one's knowledge, separated from one's intention—in response to this situation, we, as people involved in the research and development of audio technology, have continuously asked ourselves what we can do. One answer to that is Paramatch.
■ What is Paramatch?
Paramatch is a speaker identification AI model developed by our company. For an uploaded audio file, it predicts the top 3 speakers with the closest voice quality from the speaker database and presents the similarity of each as a score. This allows one to obtain objective clues to identify "whose voice was the basis for generating" a certain synthesized voice.
[Main Features]
- Predicts similar speakers from the speaker database
- Supports audio files of 30 seconds or less
Uploaded audio is not saved on the server, and no secondary use is made whatsoever. In addition, training data is used only to build the Paramatch model. We do not divert it to generative models like voice synthesis or other products and projects of our company.
Please note that the results of this tool are for reference only and do not serve as a basis for legal judgments. Furthermore, the recognition accuracy is not perfect, and we will continue to make improvements.
There are two main purposes for releasing this demo.
First, it is to verify, on an empirical basis, an effective countermeasure to protect the rights of the "voice," which is the identity of the performer. Rather than armchair discussions, we want to confirm with industry stakeholders how effective this technology is against actual cases of unauthorized use.
Second, it is to serve as a starting point for dialogue with voice actors, actors, their agencies, and rights-holder organizations. We are a company developing AI audio technology, but precisely because of that, we cannot turn a blind eye to the reality that this technology is being misused. As the responsibility of those creating the technology, we want to seriously work on technologies to protect voices as well.
■ To Voice Actors, Actors, and Industry Organizations: Seeking Partners to Work Together
We do not intend to let Paramatch end simply as a "public demo." To seriously implement this technology in society, the knowledge of voice actors and actors who face their "voice" every day on site, as well as the agencies and industry organizations that protect their rights, is indispensable.
Specifically, we would like to request cooperation in forms such as the following:
- Verification of Paramatch's effectiveness against actual cases of unauthorized use
- Consultations to improve the accuracy of the speaker database
- Discussions on industry guidelines and rights