Sesame AI has open-sourced its speech generation model, CSM-1B, which contains 1 billion parameters. This model can generate human-like voice through text and audio input and was released under the Apache 2.0 license on March 13, 2025. Developers and researchers can now utilize this technology freely with minimal restrictions.
CSM-1B forms the core technology of the viral voice assistant Maya, using Residual Vector Quantization (RVQ) to encode and reproduce various voices. Built upon Meta’s Llama model and advanced audio decoder technology, it allows for efficient voice compression even at a low bitrate of 1.1 kbps. Additionally, it offers the ability to clone a specific person’s voice from just a one-minute audio sample, providing a high degree of realism.
This release is expected to accelerate the development of AI voice generation technology, but it simultaneously raises concerns over the potential misuse of voice cloning technology. Unlike other AI voice cloning companies, Sesame AI has not imposed strict technical limitations, instead offering only ethical guidelines discouraging unauthorized voice mimicry and misinformation creation. Industry experts warn of the growing threat that AI voice technology might erode public trust and call for additional safeguards to prevent misuse.
Sesame AI asserts that AI voice technology should contribute to more natural and immersive human-machine interactions. The company is focused on developing realistic voice assistants using AI and emphasizes responsible use of the technology. However, given the potential ethical conflicts arising from the openness of the technology, discussions on the need for regulations around AI voice cloning technology are expected to intensify in the future.