mlx-community/Step-Audio-EditX-bf16

Step-Audio-EditX — bf16 MLX bundle

Step-Audio-EditX (StepFun, Apache-2.0): a 3 B audio LLM that re-delivers an existing take — emotion, speaking style, inserted paralinguistics (laughter, sighs, breaths), denoise, silence trim — in the same voice, plus zero-shot cloning. This is the whole pipeline in one directory of MLX safetensors, bf16 throughout, no quantisation, converted from the stock checkpoint:

file	component	source weights
model.safetensors (7.06 GB) + config.json	step1 LM — 32 × 3072, 48 heads / 4 KV groups, sqrt-ALiBi	stepfun-ai/Step-Audio-EditX
vq02.safetensors + vq02-config.json	Paraformer-large zh-cantonese-en online encoder (the 16.7 Hz linguistic tokenizer)	FunASR dengcunqin/…-vocab8501-online (FunASR model licence)
vq06.safetensors + vq06-config.json	S3 v1 semantic tokenizer (25 Hz, 4 096 codes)	stepfun-ai/Step-Audio-Tokenizer
step-audio-tokenizer-assets.safetensors	the k-means linguistic codebook + CMVN	stepfun-ai/Step-Audio-Tokenizer
flow-model.safetensors, flow-conditioner.safetensors + configs	upsample conformer + DiT flow matching (24 kHz mel)	Step-Audio-EditX/CosyVoice-300M-25Hz/flow.pt (StepFun-trained)
hift.safetensors + hift-config.json	HiFT vocoder	…/hift.pt (StepFun-trained)
campplus.safetensors + campplus-config.json	CAM++ speaker embedding	…/campplus.onnx (Alibaba, Apache-2.0)
tokenizer.json, tokenizer_config.json	the sentencepiece Unigram text tokenizer	derived from tokenizer.model
frontend-config.json	the 24 kHz mel front end	—

Layout and conversion: appautomaton/mlx-speech (MIT) scripts/convert/step_audio_editx.py, run with a --no-quantize patch so every component stays bf16; tokenizer.json was added from tokenizer.model (LlamaTokenizerFast).

Consumers

• Swift: xocialize/mlx-step-audio-editx-swift — EditXPipeline.load(bundle: EditXBundle(root: dir)); quantises the LM to int8 at load when asked. Every stage is parity-locked against StepFun's torch code (its PORTING-SPEC.md).
• Python: mlx-speech — StepAudioEditXModel.from_dir(dir) (its resampler should be a windowed sinc; the shipped np.interp perturbs the 16 kHz tokenizers — see the Swift package's notes).

Notes

• The Paraformer front end in the upstream code dithers with an unseeded RNG at inference (only ~96 % of its own codes agree run to run). Both consumers above run it deterministically (dither off).
• Measured on an M5 Max with the Swift consumer: ≈ 9.4 GB resident, ≈ 1.9 GB per edit, real-time factor 0.72–0.75 (0.49–0.58 with the LM at int8). Content, voice and quality of the edits equal the torch reference's on the evaluation set (mlxengine-audio E19).

Licence

Apache-2.0 (StepFun) for the LM, flow, HiFT and the S3 tokenizer; the Paraformer encoder under the FunASR model licence; CAM++ Apache-2.0 (Alibaba / 3D-Speaker). Converted and re-hosted by xocialize for mlx-community.