dots.tts-mf

Quick Start

Installation

conda create -n dots_tts python=3.10 -y
conda activate dots_tts

python -m pip install --upgrade pip
python -m pip install "git+https://github.com/studio-dots-ai/dots.tts.git" \
  -c "https://raw.githubusercontent.com/studio-dots-ai/dots.tts/main/constraints/recommended.txt"

CLI

# Few-step inference — NFE = 4 is the recommended quality / latency trade-off.
# Note: --guidance-scale is a no-op on this checkpoint (CFG is fused into the
# distilled student); leave it at the CLI default or set it to anything.
dots.tts \
  --model-name-or-path dots-studio/dots.tts-mf \
  --text "Hello, this is a zero-shot voice cloning demonstration." \
  --prompt-audio /path/to/reference.wav \
  --prompt-text "The exact transcript of the reference audio." \
  --num-steps 4 \
  --output clone.wav

Python API

from dots_tts.runtime import DotsTtsRuntime
import soundfile as sf

runtime = DotsTtsRuntime.from_pretrained(
    "dots-studio/dots.tts-mf",
    precision="bfloat16",
)

result = runtime.generate(
    text="Hello, this is a quick speech synthesis test.",
    prompt_audio_path="/path/to/reference.wav",
    prompt_text="The exact transcript of the reference audio.",
    num_steps=4,           # NFE = 4 is the recommended setting
    # guidance_scale is a no-op on dots.tts-mf — CFG is fused into the student
)

sf.write("output.wav", result["audio"].float().cpu().squeeze().numpy(), result["sample_rate"])

Recommended sampling settings

Architecture

A frozen AudioVAE encodes 48 kHz mono waveform into a continuous latent and decodes it back via a BigVGAN-style causal decoder. An autoregressive backbone predicts that latent one patch at a time:

• Semantic encoder — re-encodes each newly generated VAE patch into a compact embedding for the LLM, stripping high-variance acoustic detail.
• LLM — initialized from Qwen2.5-1.5B-Base, consumes BPE text directly (no phonemes), emits one hidden state per audio step.
• AR flow-matching head — a DiT that conditions on the LLM hidden state and the AR prefix to denoise the next VAE patch, with a frozen CAM++ speaker x-vector as side input.

CFG-aware MeanFlow distillation trains the flow-matching head as a MeanFlow student over the SCA teacher's velocity field, with classifier-free guidance directly absorbed into the student. The result is a 2–4 NFE sampler that retains the bulk of the teacher's quality with a single model evaluation per step.

Performance — dots.tts-mf

Seed-TTS-Eval (zero-shot, ~3 s reference)

At NFE = 4, dots.tts-mf essentially matches its teacher on average WER (2.94 vs. 2.95) with ~2.5× fewer model evaluations per patch and a single conditional pass per step.

CV3-Eval

See the project README for full benchmark tables including MiniMax Multilingual and EmergentTTS-Eval.

Risks and Limitations

• Misuse risk. High-fidelity zero-shot voice cloning can produce highly realistic synthetic speech. This checkpoint is intended for research and authorized deployment. Do not use it for impersonation, fraud, or disinformation. Combine downstream use with consent-aware reference-audio policies, robust synthetic-speech detection, and content watermarking. Clearly mark AI-generated audio.
• Few-step trade-off. At NFE = 2/3 there is a measurable WER and SIM regression vs. NFE = 4 (see table above). Pick the NFE that matches your latency budget.
• CFG fused. Unlike base / soar, --guidance-scale is ignored by the MeanFlow sampler — guidance is baked into the student at distillation time and cannot be adjusted at inference.
• Low-resource WER gap. A BPE backbone inherits the text LLM's language coverage at the cost of a higher data appetite. On script-divergent and under-represented languages (Arabic, Hindi, Turkish, Vietnamese) WER is higher than on high-resource languages; speaker similarity is preserved.
• Speech-heavy training. The backbone is trained on a speech-heavy mixture. Singing and unified speech + sound generation are not covered.

Citation

@article{dotstts2026,
  title   = {dots.tts Technical Report},
  author  = {dots.tts Team},
  journal = {arXiv preprint},
  year    =