faraday/ema-lightning-mlx

EMA Lightning — MLX Swift voice weights

MLX-compatible weights converted from EMA Lightning, by Canberk Aslan. This is an EMA native Swift/MLX voice bundle, with the vocabulary, tensor schema, decoder architecture, sampling schedule and provenance in voice-manifest.json. It is not an mlx-lm, Transformers, or generic MLX Audio drop-in checkpoint.

Files

File	Purpose
acoustic.safetensors	FP32 text/duration, conditioning and acoustic model weights, using MLX layouts.
decoder.safetensors	Canonical FP32 decoder weights, using MLX convolution layouts.
constants.safetensors	FP32 source-generated rotary and timestep constants.
voice-manifest.json	Runtime format-v2 schema, shapes, vocabulary, architecture and asset SHA256 values.
adapter-validation.json	Original export adapter validation evidence.
conversion-manifest.json	Exporter/source provenance, asset checks and stored precision.
CHECKSUMS.sha256	SHA256 checksums for the distributed files.
LICENSE / NOTICE	Upstream Apache 2.0 license and attribution.

Precision

All tensors in the three safetensors files are stored in FP32. No training, distillation or quantization was performed for this export. Inference precision and kernel selection are controlled by the consuming application, independently of these stored weights.

Usage

Use the compatible EMA native Swift/MLX application with this folder as its model directory. For FP32 inference:

ema synthesize "Merhaba dünya." --engine mlx --variant fp32 --model-dir /path/to/ema-lightning-mlx --seed 7 --output speech.wav

The bundle contains weights and model metadata. The matching application supplies the executable, text normalizer and Metal shader library. Asset layouts and schema target mlx-swift 0.32.3.

Export validation

adapter-validation.json records the original export adapter checks. voice-manifest.json records tensor shapes and asset hashes; CHECKSUMS.sha256 covers the distributed files. These reports describe the weight export. Inference accuracy and performance depend on the consuming runtime and its configuration.

Conversion

The existing export was made by Engines/mlx/conversion/export_mlx.py in the EMA native project. PyTorch Conv1d OIK and ConvTranspose1d IOK layouts are converted to MLX OKI; scalar tensor shapes are retained. The voice manifest records the source revision/checkpoints and all transformations. A fresh export can be made from the pinned local original source and checkpoints with that script; copying this existing verified bundle does not reconvert or retrain it.