CSM-1B

CSM-1B is built for conversational speech generation rather than plain one-shot narration. The useful part is the way it handles context. You can feed it prior turns and speaker audio cues, then generate the next line with a tone that feels more grounded in an ongoing exchange. That makes it more interesting for assistant prototypes, short dialogue tests, spoken product flows, and rough character back-and-forth than a generic TTS model that only reads isolated lines.

Start with the official Hugging Face Space and keep the first test short. Use one or two brief speaker prompts, generate a simple back-and-forth, and listen for pacing, turn-taking, and whether the tone still sounds natural once a second line answers the first. If the browser result feels unusually convincing, move to the model card and GitHub repo for the local path. Sesame documents public code and Transformers support, but local use still assumes gated model access and a capable GPU, so this is not a casual laptop workflow.

The polished Sesame voice demo on sesame.com is powered by a fine-tuned variant, not the exact open checkpoint hosted here. The open release is still strong and reproducible, but you should evaluate it as an open conversational speech workflow, not assume the browser assistant demo and the public model are identical.

• Prototype two-speaker assistant or agent conversations before building a bigger voice stack.
• Compare context-aware speech generation against plain text-to-speech on the same script.
• Draft short dialogue scenes, spoken explainers, or product voice interactions with more natural pacing.
• Pressure-test whether an open conversational speech workflow is good enough before reaching for a closed voice API.

Try the demo

View model page