Dots Tts Mlx

Languages: same as upstream dots.tts — all 24 (Arabic, Cantonese, Chinese, Czech, Dutch, English, Finnish, French, German, Greek, Hindi, Indonesian, Italian, Japanese, Korean, Polish, Portuguese, Romanian, Russian, Spanish, Thai, Turkish, Ukrainian, Vietnamese). The 5-language check in Quality is only the quantization spot-check, not the supported set.

These are converted + LLM-quantized MLX safetensors, not PyTorch. They load only with the dots-tts-mlx runtime on Apple Silicon (Metal). For the original PyTorch model, see rednote-hilab/dots.tts-soar.

⚡ Two decoders, one voice. int4/int8 are the standard soar build (reference quality). mf-int4/mf-int8 are MeanFlow — a distilled few-step decoder (rednote-hilab/dots.tts-mf) that runs the acoustic DiT at NFE=4 with no classifier-free guidance, ~2× faster than the 10-step path with no measurable quality loss. Same model, same 24-language cloning — pick the folder that fits your speed budget. For cloning, reference mode (audio + transcript) is the recommended quality path.

Variants

Only the Qwen2.5-1.5B LLM trunk (≈70% of the weights) is quantized (group-wise affine, group size 64); the precision-sensitive flow-matching DiT, the BigVGAN vocoder, and the CAM++ speaker encoder stay bf16.

MeanFlow (mf-*) is the distilled rednote-hilab/dots.tts-mf checkpoint — the same architecture plus a small duration embedder — auto-detected from config.json (no flag): point the runtime at an mf-* folder and it uses the few-step solver. Measured 1.9–2.2× faster than the 10-step path on reference cloning (EN/HI/ZH). It drops classifier-free guidance, so --guidance-scale is ignored. Use reference cloning (audio + transcript) for best quality.

Quality

Quantization is validated to be lossless relative to the full-precision MLX build: on a small multilingual acceptance check (EN/DE/ES/FR + Hindi), int8 and int4 showed no transcription-accuracy or voice-similarity regression vs bf16. This is a sanity check, not a dataset-scale benchmark — evaluate on your own content.

Correctness of the port itself is gated per-stage against the original PyTorch model (AudioVAE PSNR ≈ 56 dB; attention / DiT / LLM / semantic-encoder cosine ≥ 0.9999) — see the runtime repo.

Usage

# 1. install the quant-aware runtime (>= v0.2.0)
pip install "git+https://github.com/sb1992/[email protected]"

# 2. download the variant you want  (use "mf-int4/*" for the faster MeanFlow decoder)
hf download shraey/dots-tts-mlx --include "int4/*" --local-dir ./dots-tts-mlx-weights

# 3. run (files land in ./dots-tts-mlx-weights/int4/)
dots-tts --model ./dots-tts-mlx-weights/int4 \
    --text "Hello from MLX." --ref-audio reference.wav --language EN \
    --out-path out --out-prefix clone

The runtime auto-detects the quantization block in config.json, so nothing changes at the CLI/API level vs an unquantized directory. Python API and the full flag set: see the runtime repo.

• Memory: peak RAM scales with generation + reference length — roughly ~6 GB for a short clip, up to ~13 GB for a ~30 s clip (int4); resident weights are ~2.4 GB. The render peak is activation-bound (the bf16 DiT + vocoder working set), so it's the same for soar and MeanFlow and isn't reduced by quantization. MLX's allocator may cache up to its memory limit, but that cache is releasable.
• Requires: Apple Silicon (MLX is Metal-only), Python ≥ 3.10.

Attribution & licenses

Derivative quantized weights of rednote-hilab/dots.tts-soar (Apache-2.0) — you must comply with the upstream license. Components:

• dots.tts — model · code — Apache-2.0, © the dots.tts team at rednote-hilab.
• Qwen2.5-1.5B-Base (LLM backbone) — Apache-2.0.
• CAM++ / 3D-Speaker (speaker x-vector encoder) — Apache-2.0.
• BigVGAN (vocoder/decoder architecture style) — MIT, © NVIDIA.

MLX port + quantization code: github.com/sb1992/dots-tts-mlx (Apache-2.0).

Responsible use

This performs zero-shot voice cloning — it can reproduce a person's voice from a few seconds of audio. Only clone voices you own or for which you have explicit, informed consent; do not use it for impersonation, fraud, or deception; and disclose AI-generated audio wherever it's shared. See the upstream risks guidance.