CSM-1B
Sesame
CSM (Conversational Speech Model) is a speech generation model from Sesame that produces RVQ audio codes from text and audio inputs. Its architecture pairs a Llama backbone with a smaller audio decoder that emits Mimi audio codes, and it accepts prior conversation turns as context so speech is generated in the flow of a dialogue. Sesame released the 1B variant on 13 March 2025 under Apache-2.0; it has been supported natively in Hugging Face Transformers since version 4.52.1. The released checkpoint is a base model, not tuned to any specific voice.
Strengths
- Generates speech conditioned on prior conversation turns supplied as text plus audio segments, rather than one isolated utterance at a time
- Architecture pairs a Llama backbone with a smaller audio decoder producing Mimi RVQ audio codes
- Speaker-id conditioning lets a single checkpoint voice multiple participants in a generated conversation
- Apache-2.0 licence covering both the released code and the checkpoint
- Supported natively in Hugging Face Transformers since version 4.52.1, so no bespoke runtime is needed
- Runs on consumer CUDA GPUs (tested on CUDA 12.4 and 12.6), with a hosted Hugging Face Space for testing
Caveats
- The model cannot generate text at all: Sesame states it is 'trained to be an audio generation model and not a general-purpose multimodal LLM', so a separate LLM is required for any conversational application.
- Non-English support is weak by the vendor's own account: it exists only 'due to data contamination in the training data' and 'likely won't do well'.
- This is a base generation model with no voice fine-tuning, so it produces a variety of voices rather than a consistent one; the polished voice in Sesame's public demo is a fine-tuned variant, not this checkpoint.
- Access is gated: the Hugging Face repository requires accepting conditions, and running the model also requires access to the separately gated meta-llama/Llama-3.2-1B.
- No benchmark results of any kind are published in the repository or model card, and a CUDA-compatible GPU is listed as a requirement.
- Vendor
- Sesame
- License
- Apache-2.0
- Release Date
- 2025-03-13
- Modalities
- textaudio
Capabilities
Resources
Pricing
Sesame publishes no hosted API or pricing for CSM-1B. The checkpoint is distributed on Hugging Face under Apache-2.0 for self-hosting, subject to accepting the repository's access conditions, and a free Hugging Face Space is provided for testing.