tinytrashlabs/LuxTTS-mlx

LuxTTS — MLX weights (fp32)

LuxTTS converted to MLX-native safetensors, for Apple silicon. Same model, same precision — a key-remap and a conv-layout transpose, nothing retrained and nothing quantized.

Upstream ships model.pt (torch) and ONNX graphs. Neither loads in MLX, so every MLX consumer had to run the conversion itself and keep the result somewhere local. This repo is that result, published once.

Files

File	Size
lux_model.safetensors	468 MB	ZipVoiceDistill — fm_decoder + text_encoder + embeddings
lux_vocoder.safetensors	61 MB	Vocos vocoder, MLX conv layout
config.json	—	model dimensions, 24 kHz feature config
tokens.txt	—	phoneme vocabulary

fp32 on purpose. fp16 halves the download and is indistinguishable on a long, clean reference — and audibly worse on a short one. Blind A/B on a clone made from an 8-second phone recording: 4 of 8 fp16 renders had artefacts a listener rejected, against 0 of 8 at fp32. If you are cloning from short user recordings, use these.

Use

// EngineKit (Gloam Voice Studio)
let model = try await LuxSpeechModel.load(from: weightsDirectory)

The directory needs all four files side by side.

Provenance and license

Converted from YatharthS/LuxTTS, Apache-2.0, which this repo inherits. The vocoder is Vocos. Conversion and publication by Tiny Trash Labs for Gloam Voice Studio.