Sesame CSM-1B
Free.ai (self-hosted)
·
tts
·
~500 Ii-token clip
Sesame CSM-1B — Apache 2.0. Conversational Speech Model designed for low-latency, real-time voice. 24 kHz output, sounds best with a short reference-audio context turn. Self-hosted on Free.ai for the /voice/realtime/ tool.
Sebenzisa nge-API
curl -X POST https://api.free.ai/v1/tts/ \
-H "Authorization: Bearer sk-free-..." \
-H "Content-Type: application/json" \
-d '{"model":"csm-1b","text":"hello world"}'Imibuzo ebuzwa rhoqo
Sesame CSM-1B supports a wide range of languages. The exact list depends on the engine; the form on this page accepts any text and the engine will render in its supported languages. See /voice/ for the full multi-engine picker if you need a specific language.
Most engines render neutral-American English by default and a region-appropriate accent for non-English languages. Premium engines may expose accent variants — paste a sample to compare.
SSML support varies by engine. Pause, prosody, and emphasis tags are honored on most premium engines and on a few self-hosted ones. Plain text always works — no markup required.
Streaming TTS is available on premium engines via the /v1/tts/ API endpoint with stream=true. The web UI on this page returns the full clip once rendering finishes.
Sesame CSM-1B runs on our own GPUs. Generation draws from your daily free pool first. Once depleted, paid tokens start at $5 → 200,000 tokens. Roughly ~5 tokens per character, minimum 100 per clip.
Up to 5,000 characters per request on the web UI. For longer pieces (audiobooks, full chapters), use /voice/audiobook/ which chunks and stitches automatically, or call the API in a loop.
Yes — POST a list of strings to /v1/tts/batch/, or use the workspace UI at /workspace/ to chain TTS into a longer pipeline (e.g., translate → speak → stitch).
Yes — POST text to /v1/tts/ with model="Sesame CSM-1B" (or the slug on this page). Returns WAV or MP3. See /api/ for full reference + SDK snippets.
This page is text-to-speech, not voice cloning — the voice is the engine's default. For voice cloning (uploading a reference audio), see /voice/clone/, which requires you to either own the voice rights or have explicit written consent.
Self-hosted engines run on Free.ai-owned GPUs; nothing leaves our servers. Premium engines pass text to upstream model providers under our DPA. We do not train on your inputs and do not sell data.
Yes — Free.ai grants commercial use of generated audio. The engine's underlying license (Apache 2.0, MIT, or vendor terms) is shown above and on the model reference page; in practice this means voiceovers, ads, podcasts, and apps are all in-scope.
Yes — failed jobs auto-refund to the source (daily pool or paid tokens). If a refund does not show up the same day, email contact@free.ai.