Sesame CSM 1B: local setup and voice cloning tests

Uploaded March 14, 2025, this tutorial tests Sesame CSM 1B, the 1-billion-parameter speech model released at that time. The speaker describes it as open source and distinguishes it from Sesame's online conversational demo. The setup and results reflect that release, rather than a current installation guide.

The walkthrough covers cloning the repository, creating a Python virtual environment, installing dependencies and running a sample script that saves generated speech as a WAV file. At the time, the required models were gated on Hugging Face: users needed an account, access approval and CLI authentication with an access token. The speaker explains these prerequisites but does not demonstrate authentication.

Tests use a laptop RTX 4060 with 8GB VRAM. The speaker observes about 5.3GB total GPU memory use during a short generation, including a roughly 1.6GB baseline with OBS and other applications open. A later, longer prompt reaches about 6.5GB total. These are observations from this setup, not general hardware requirements.

Unconditioned outputs vary in voice between runs. For voice cloning, the speaker supplies four Microsoft Azure speech clips with matching transcripts, then tries recordings of their own voice. They judge the Azure clone convincing but note tone changes and an odd pause in a longer output. Their own clone has uneven timing. Converting stereo reference WAV files to mono with FFmpeg resolves a tensor shape error in their test.