mlx-community/IndexTTS-2-fp16

IndexTTS-2 (MLX, fp16)

MLX conversion of IndexTeam/IndexTTS-2 — zero-shot voice-cloning TTS with decoupled emotion control and duration control (22.05 kHz output). Consolidated single-repo weight set for the Swift-MLX port xocialize/mlx-indextts2-swift (an MLXEngine tts package; also loadable by solar2ain/mlx-indextts, whose conversion lineage these weights follow — thanks to vanch007/mlx-indextts2-standard-fp16).

Contents

File	Component	Dtype
gpt.safetensors	UnifiedVoice v2 GPT AR (text→semantic) + conformer/perceiver conditioners	fp16
s2mel.safetensors	S2Mel flow-matching (DiT/WaveNet/length-regulator)	fp16
bigvgan.safetensors	BigVGAN v2 vocoder	fp16
vq2emb.safetensors	Semantic-code embedding	fp16
model.safetensors	w2v-BERT 2.0 conditioning front-end	fp32 (as upstream)
semantic_codec/model.safetensors	MaskGCT RepCodec semantic codec	fp32
campplus.safetensors	CampPlus DTDNN speaker embedding (192-d)	fp32
emotion_matrices.safetensors	8-emotion preset matrices (emo_matrix 73×1280, spk_matrix 73×192)	fp32
tokenizer.model / config.yaml / config.json	SentencePiece Unigram (12k) + resolved config	—

int8/int4 tiers are produced in-memory by the Swift package (GPT transformer Linears, group 64) — no separate quantized repos.

Parity

The Swift port is phase-gated against a captured oracle: tokenizer bit-exact; GPT latent cos 0.9999995; conditioners ≥0.999999; CFM mel 0.9999999; vocoder at the reference's own CPU↔Metal precision floor; end-to-end waveform reproduces the reference render (spectral cos 0.9998, −23.0 dBFS vs −23.2). Emotion (emoAlpha) and exact targetDuration were validated live through MLXServeEngine.

License

Weights: NonCommercial — the INDEX_MODEL_LICENSE requires written authorization from bilibili for commercial use. Port/inference code is Apache-2.0 / MIT (see the source repos).