IndexTTS2-MLX-fp16

This is an fp16 artifact export of IndexTeam/IndexTTS-2 for Apple Silicon runtime evaluation.

Runtime support is not implied by this model card. The bundle preserves upstream key names and is intended for a speech-swift Swift/MLX port.

Runtime and Documentation

• speech-swift runtime repo: https://github.com/soniqo/speech-swift
• speech-models exporter repo: https://github.com/soniqo/speech-models
• Voice cloning docs: https://soniqo.audio/voice-cloning
• CLI docs: https://soniqo.audio/cli
• Benchmark article: https://soniqo.audio/blog/voice-cloning-benchmarks

Model

Field	Value
Display name	IndexTTS2
Source	IndexTeam/IndexTTS-2
Source revision	740dcaff396282ffb241903d150ac011cd4b1ede
Publish repo	aufklarer/IndexTTS2-MLX-fp16
Parameter count	1.5B-class
Sample rate	24000 Hz
Voice conditioning	reference audio plus optional emotion reference/control path
Streaming	needs runtime validation
License posture	check upstream/commercial terms before product use
Runtime status	artifact-export; Swift native inference not implemented

Files

• soniqo_manifest.json - export manifest with source, runtime status, and file metadata
• config.json - upstream or generated root config for runtime metadata
• *.safetensors - fp16-converted model weights with upstream key names preserved
• aux/ - auxiliary model configs and fp16-converted weights required by this runtime, when applicable

Auxiliary Models

This expanded export includes auxiliary assets that upstream IndexTTS2 downloads at runtime:

Component	Source	Purpose
w2v-BERT 2.0	facebook/w2v-bert-2.0	SeamlessM4T features and semantic hidden states for reference audio
MaskGCT semantic codec	amphion/MaskGCT	Quantizes reference semantic features and maps generated codes to embeddings
CAMPPlus	funasr/campplus	Extracts the 192-d global style vector from the speaker reference
BigVGAN	nvidia/bigvgan_v2_22khz_80band_256x	Decodes generated 80-band mel spectrograms to waveform

Notes

• Public upstream weights include GPT/S2Mel checkpoints plus a Qwen emotion merge directory.
• The expanded export also bundles the upstream auxiliary models required by IndexTTS2 v2 synthesis.
• This exporter converts PyTorch pickle checkpoints to safetensors and preserves subdirectory layout.
• Native Swift inference still needs the graph ports for w2v-BERT, MaskGCT semantic codec, GPT/S2Mel, CAMPPlus, and BigVGAN.

Usage

These artifacts are for runtime implementation and evaluation. They are not a drop-in Transformers checkpoint.

git clone https://huggingface.co/aufklarer/IndexTTS2-MLX-fp16

Source

• Upstream model: https://huggingface.co/IndexTeam/IndexTTS-2
• Converted bundle: https://huggingface.co/aufklarer/IndexTTS2-MLX-fp16