Skip to content

Open-source State-of-the-art TTS model which runs on a CPU 😻

Python

Update api.md

Kitten55eb6789e5cc
diff --git a/docs/api.md b/docs/api.md
index 63404f9..606d2b1 100644
--- a/docs/api.md
+++ b/docs/api.md
@@ -10,13 +10,14 @@ Loads whichever model family the repository's `config.json` declares.
|---|---|---|---|
| `model_name` | `str` | -- | Hugging Face repo ID, or a path to a local repository directory |
| `cache_dir` | `str` | `None` | Local directory for caching downloaded files |
-| `device` | `str` | `None` | `"cuda"` or `"cpu"`; auto-detected when omitted. CPU works, at roughly 2-3x slower than real time -- and on a many-core host `torch.set_num_threads(8)` is markedly faster than letting it use every core |
+| `device` | `str` | `None` | `"cuda"` or `"cpu"`; auto-detected when omitted. The Python package runs slower than real time on most CPUs. For real-time performance, please refer to our [optimized implementation](https://github.com/KittenML/KittenTTS#running-on-cpu) |
| `hf_token` | `str` | `None` | Token, only for downloading a private model repository |
| `decoder` | `str` | repo default | Which decoder to load; see [decoders.md](decoders.md) |
| `weights` | `str` | `"packed"` | Which weight packing to load: `"packed"` (lossless), `"emb4"` (smaller, quantised embedding), `"full"` (bf16) |
| `backend` | `str` | `None` | Default PyTorch inference, or `"vllm"` for optional NVIDIA GPU inference on Linux |
| `vllm_options` | `dict` | `None` | Engine memory and context settings; see [vLLM](vllm.md) |
+
# `model.generate(text, voice=None, reference=None, ...)`
Returns a float32 NumPy array at `model.sample_rate` (24 kHz). The speaker comes from exactly one