Sesame has moved its CSM-1B AI voice generator into the open-source world, making the base model available under the Apache 2.0 license. The release gives developers broad room to test, adapt and use the technology commercially, while also putting a powerful voice cloning system into wider circulation.
What Sesame Released
CSM-1B is a billion-parameter base model for audio generation. Sesame has made the code available on Github, and anyone can test the audio generation capabilities directly.
The model is not only a research object. A fine-tuned version of CSM-1B also powers Maya's AI voice system, the company's conversational voice assistant demo.
The Apache 2.0 license matters because it allows broad commercial use with minimal restrictions. That makes the release more than a public showcase. It gives developers and companies a practical route to build with the model, experiment with its speech output and potentially integrate it into products.
Why The Voice Feels Different
Sesame's approach is built around what the company calls "voice presence" in AI systems. Instead of aiming for perfectly polished speech, the system intentionally includes human-like details that make conversation feel less mechanical.
Early testing highlighted subtle behaviors such as micro-pauses, emphasis variations and laughter during conversations. In one interaction, Maya responded in real-time to a user's sudden giggle, showing what was described as emotional awareness.
The system also uses mid-sentence self-corrections, apologies for interruptions and filler words. These choices make the output different from the cleaner, more corporate delivery associated with ChatGPT or Gemini, according to Techradar.
In simulated conversations about work stress or party planning, the system did not simply rely on generic replies. It produced contextually appropriate responses and questions, which is central to why the demo attracted attention.
How CSM Processes Speech
Sesame has not released a formal paper, but its blog post described the basic architecture. CSM uses a two-part transformer structure: a backbone transformer with 1-8 billion parameters for basic processing, and a smaller decoder with 100-300 million parameters for audio generation.
The system separates speech into two kinds of tokens. Semantic tokens handle linguistic properties and phonetics, while acoustic tokens represent sound characteristics such as pitch and emphasis.
Training is also split in a way that reduces the audio generation burden. The audio decoder trains on just one-sixteenth of the audio frames, while semantic processing uses the complete dataset.
The model trained on one million hours of English audio data across five epochs. It can process sequences of up to 2,048 tokens, which is about two minutes of audio, in an end-to-end architecture.
That integrated design differs from traditional text-to-speech systems because text and audio are processed together. The demo voice also reveals use of a 27-billion parameter version of Google's open-source LLM Gemma, although that detail was not directly stated in the blog post.
Performance And Limits
Testing suggests Sesame's system can come close to human speech in short bursts. In blind tests, participants could not distinguish between CSM and real humans when they heard short conversation snippets.
Longer conversations still exposed weaknesses. The source article notes occasional unnatural pauses and audio artifacts, showing that the technology is not yet indistinguishable from a person across extended dialogue.
Sesame also created custom phonetic benchmarks to evaluate performance. In listening tests, participants rated generated speech as equivalent to real recordings when heard without context. When context was provided, they still preferred the original recordings.
The Safety Tradeoff
The open-source release brings a direct safety question. Sesame's stated safety approach consists of guidelines asking developers and users to avoid unauthorized voice cloning, misleading content creation and other "harmful" activities.
The concern is that the model can clone voices with just one minute of source audio. That ability could support useful experiments in AI voice generation, but it could also enable forms of voice-based fraud.
The release also shows how difficult it can be for proprietary AI companies to preserve technical advantages in fast-moving AI fields. The source article notes that OpenAI previously chose not to release similar technology because of safety concerns, but open-source development has made that kind of protective choice less effective.
Sesame's next steps point toward more capable systems. The company has said it plans to scale up both model size and training scope, expand to over 20 languages, integrate pre-trained language models and develop fully duplex-capable systems that can learn speaker transitions, pauses and pacing directly from data.
"Building a digital companion with voice presence is not easy, but we are making steady progress on multiple fronts, including personality, memory, expressivity and appropriateness," the developers note.
Founded by former Oculus CTO Brendan Iribe and his team, Sesame AI has positioned CSM-1B as both a technical milestone and a sign of where voice assistants powered by LLMs may be heading. The result is a model that makes AI speech feel more natural, while making the governance problem harder to ignore.