Skip to content

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

Python

Update mlx.md

Kitten3d7b2699a853
diff --git a/docs/mlx.md b/docs/mlx.md
index 871515d..6dd7394 100644
--- a/docs/mlx.md
+++ b/docs/mlx.md
@@ -77,32 +77,3 @@ reference is unchanged.
KittenTTS's sentence chunking, and `generate_stream` yields completed chunks.
- Voice cloning without a supplied `reference_text` still loads Whisper through
transformers, which runs on the CPU.
-
-## Benchmarks
-
-Measured on an Apple M4 Max with 12 GB of memory, macOS 15.6, Python 3.13, PyTorch
-2.14.1 and MLX 0.32.4, using `weights="packed"`, the default decoder and the stable
-preset. These are median end-to-end synthesis times from five runs after one warmup
-per voice (three runs for PyTorch, which is slow enough that the medians were already
-stable). Short, medium and long inputs contain 61, 205 and 522 characters; the long
-input is split into two chunks. Model initialization and file writes are excluded.
-
-The two backends sample differently, so each produces a slightly different length of
-audio from the same text β€” hence the realtime factors rather than a bare time ratio.
-
-| Input / voice | PyTorch CPU | MLX | PyTorch | MLX | Speedup |
-|---|---:|---:|---:|---:|---:|
-| Short / Bruno | 9.23 s | 0.86 s | 0.41x realtime | 4.76x realtime | 11.6x |
-| Short / Bella | 13.06 s | 1.12 s | 0.39x realtime | 5.07x realtime | 13.0x |
-| Medium / Bruno | 23.45 s | 2.55 s | 0.53x realtime | 4.77x realtime | 9.0x |
-| Medium / Bella | 30.25 s | 3.31 s | 0.57x realtime | 5.21x realtime | 9.1x |
-| Long / Bruno | 54.98 s | 6.14 s | 0.54x realtime | 4.68x realtime | 8.7x |
-| Long / Bella | 71.21 s | 8.42 s | 0.55x realtime | 4.91x realtime | 8.9x |
-
-Of the 0.86 s short run, the language model takes about 0.68 s β€” roughly 150 tokens
-per second, against the codec's 25 tokens per second of speech β€” and audio decoding
-about 0.16 s. Decoding on the CPU instead (`device="cpu"`) costs about 1.0 s, which
-is why MPS is the default for it.
-
-This machine is a virtual machine, which costs some memory bandwidth; a Mac running
-the same model natively should be at least as fast.