Sesame: CSM 1B

Sesame: CSM 1B is a conversational speech model from Sesame. It sits in the speech category and is identified on api14 by the model id sesame/csm-1b. The provider describes it as a 1B parameter model, which makes it a relatively small option for generating spoken audio from written text.

The model takes text as its only input and produces speech as its only output, specifically English speech. It has a context length of 4096 tokens and a maximum output of 3686 tokens. Voice selection is part of what the model offers: the available voices are conversational_a, conversational_b, read_speech_a, read_speech_b, read_speech_c, read_speech_d, and none. The split between conversational and read speech styles means you can aim for a back and forth tone or a more formal narration tone depending on the voice you pick.

Those capabilities point to a fairly clear set of uses. Because the input is text and the output is audio, CSM 1B fits work where you already have a script or a block of copy and need it spoken aloud. The conversational voices suit dialogue style content, assistants that talk back, or interactive experiences where a natural speaking tone matters. The read speech voices suit narration, reading articles or documents aloud, and any case where a steady, presentational delivery is wanted. The none option is available when you do not want to pin the output to a named voice. Since the model is small at 1B parameters, it is a reasonable candidate when you want speech generation without a large model behind it.

You call it through api14's OpenAI-compatible API using the model id sesame/csm-1b. Because the API follows the OpenAI format, you send a request in the usual way and set the model field to that id, with your text as the input. The model's supported parameters list is empty, so there are no extra tuning parameters to pass beyond what the request format itself requires. Voice choice is exposed through the voices listed above, so if you need a specific delivery style, select the matching voice rather than relying on sampling settings.

There are limits worth knowing before you commit. The context window is 4096 tokens and the maximum output is 3686 tokens, so very long scripts will need to be split across requests. Output is English speech only, and input is text only, so the model does not handle audio input or non English output. If your task needs a different output modality, a longer context, or tunable generation parameters, another model on api14 may fit better.