whatmodel.sh / tts / Sesame CSM-1B
Sesame CSM-1B
↗ share
+ contribute
Sesame CSM-1B is an open-weights text-to-speech model from Sesame, released March 2025. Its published size is 1B parameters. Licence: apache-2.0.
Catalogue identity notes
Sesame Conversational Speech Model 1B (sesame/csm-1b) as loaded by EmergentTTS generator.load_csm_1b ('Sesame1B'). Weight revision not pinned. Parent lane reuses this id for RW-Voice-EQ 'CSM-1B' (lucataco/csm-1b on Replicate).
Model information sources
· Weights licence ↗ · Parameter count ↗
Maker size label ("the 1B CSM variant"). Sesame sizes CSM by backbone + decoder ("Tiny: 1B backbone, 100M decoder"), so 1B is the Llama backbone class; the audio decoder and the Mimi codec are not included, and the released checkpoint total is larger.
Score within the Vapi Humanness · Bradley–Terry rating, 2026-09-25 snapshot comparison set only.
$ how it compares
The text-to-speech models ranked either side of Sesame CSM-1B, out of 148 with a score.
| # | model | score | where |
|---|---|---|---|
| 123 |  Eleven v4 Turbo | 36±4 | api $40.00/M chars |
| 124 |  VITS (VCTK multi-speaker) | 34±5 | local |
| 125 | Sesame CSM-1B | 34±5 | api $7.00/M chars +local |
| 126 |  VoiceCraft 830M (giga830M) | 32±4 | local |
| 127 |  VALL-E X (Plachtaa open implementation) | 31±4 | local |
see all text-to-speech models · find the ones that fit your machine
$ what it's good at
other tests8 not in a task · 92% of its score
| test | result | vs the others in its run | counts |
|---|---|---|---|
| EmergentTTS · 6-category preference | 13.4% ✎ | ▅▅▃▃█▅▅▃·▃·▅ beats 2/16 | 30% |
| EmergentTTS · v1 paper automatic MOS (WV-MOS) | 3.38 / 5 ✎ | ▄·█·█▄····█▄ beats 1/8 | 21% |
| EmergentTTS · 6-category WER | 32.3% ✎ | ▃▄··▂▄▄··█▄▆ beats 0/16 | 15% |
| balacoon zero-shot TTS · DAPS-celeb · ReDimNet speaker similarity | 67.3% ✎ | ▄·······█··▄ beats 1/3 | 8.8% |
| balacoon zero-shot TTS · VCTK · ReDimNet speaker similarity | 65.2% ✎ | █···█····█·█ beats 1/3 | 7.3% |
| RW-Voice-EQ TTS · Long-Form Stability (human) | 3.44 / 5 ✎ | █······█·███ beats 0/4 | 4.8% |
| balacoon zero-shot TTS · VCTK · Whisper-large-v3-turbo CER | 1.9% ✎ | █····█····██ beats 0/3 | 3.3% |
| balacoon zero-shot TTS · DAPS-celeb · Whisper-large-v3-turbo CER | 8.5% ✎ | ██········██ beats 0/3 | 3.1% |
$ where to run it
Through an API
api · DeepInfra · $7.00/M chars · deepinfra.com ↗ · price sourceUSD per 1M input characters: DeepInfra lists sesame/csm-1b at $7.00 / 1M characters (page pricing data: input_character_length billing; not deprecated in DeepInfra's catalogue). Product: DeepInfra hosted inference (text to speech). API model: sesame/csm-1b (Sesame CSM 1B checkpoint). Settings: standard serverless. Price basis: published list rate. Currency: USD. DeepInfra official model page, observed 2026-09-25.
Other ways to run it
- Run it as published · needs ~6.61 GB, Mac only · from github.com/Blaizzy
pip install mlx-audio && mlx_audio.tts.generate --model mlx-community/csm-1b --text "Hello from Sesame." --ref_audio ./reference_voice.wav --ref_text "This is what my voice sounds like." --play
Official weights and documentation: huggingface.co/sesame ↗
$ how the score was worked out · EmergentTTS-Eval six-category… 44%, balacoon speech_gen_baselines… 22%, EmergentTTS-Eval v1 paper (ar… 21% and 1 more
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 EmergentTTS-Eval six-category leaderboard | 2 | 44% |
| balacoon speech_gen_baselines · zero-shot TTS metrics (VCTK, DAPS-celeb) | 4 | 22% |
 EmergentTTS-Eval v1 paper (arXiv:2505.23009v1) · Table 1 automatic MOS | 1 | 21% |
| RW-Voice-EQ Bench (Hume AI) | 1 | 5% |
| prior · pull toward the middle | 8% | |
| weighted average | = 34.2 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json