Skip to content
AI Tutorials
Text-to-Speech
dots.tts

Install dots.tts with Conda

Introduction

dots.tts is a 2B-parameter fully continuous, end-to-end autoregressive (AR) text-to-speech system. The backbone pairs a semantic encoder, an LLM, and an autoregressive flow-matching acoustic head over a 48 kHz AudioVAE, with no discrete tokens anywhere in the pipeline.

GitHub: https://github.com/studio-dots-ai/dots.tts
Hugging Face: https://hf.co/collections/dots-studio/dotstts
Playground: https://hf.co/spaces/dots-studio/dots.tts
Demo Page: https://studio-dots-ai.github.io/dots.tts-demo/
Report: https://arxiv.org/abs/2606.07080

dots.tts

dots.tts achieves the best average performance on Seed-TTS-Eval, with WERs of 0.94% / 1.30% / 6.60% and SIM scores of 81.0 / 77.1 / 79.5 on the zh / en / zh-hard test sets, respectively. It further attains the highest average speaker similarity (83.9) on the 24-language MiniMax multilingual benchmark. Across other benchmarks, dots.tts also consistently demonstrates open-source state-of-the-art performance, exhibiting strong generation stability, voice cloning ability, and emotional expressiveness.

dots.tts dots.tts

Prerequisites

System requirements:

  • Operating System: Windows 10/11 (64-bit), macOS, or Linux (Debian/Ubuntu).
  • Python: version >= 3.10 required
  • Disk Space: 10GB+ recommended (for dependencies and model cache). At least 400 MB for Miniconda; 3 GB+ for full Anaconda.
  • The GPU is optional but HIGHLY Recommended for Performance.
  • Internet: For downloading dependencies and models from Hugging Face Hub.
Requirements CPU RAM
Minimum Intel Core i3 or AMD Ryzen 3 8 GB
Recommended Intel Core i5 or AMD Ryzen 5 16 GB
High-performance Intel Core i7/i9 or AMD Ryzen 7/9 32 GB
Environment Run this Command
CPU only pip3 install torch torchvision
Apple Silicon pip3 install torch torchvision
AMD ROCm Install PyTorch for ROCm
CUDA 11.8 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
CUDA 12.1 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
CUDA 12.6 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu126
CUDA 12.8 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128
CUDA 13.0 pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

Video tutorial

Coming soon!

Step 1. Install Miniconda Package

Step 2. Create Conda Environment

Create an environment.yml file

name: dots_tts
channels:
- conda-forge
- defaults

dependencies:
# Python version (requires >=3.10,<3.13)
- python=3.10
- pip

# Handled via Conda to guarantee OpenFST binary mapping
- pynini

- pip:
    # PyTorch CPU only
    - torch==2.8.0
    - torchaudio==2.8.0

    # Install dots.tts directly from PyPI
    - dots.tts

Activate conda environment:

conda env create -f environment.yml
conda activate dots_tts

Step 3. Run the Inference

Command Line Interface 

# sanity check
dots.tts --help
usage: dots.tts [-h] --model-name-or-path MODEL_NAME_OR_PATH [--revision REVISION] [--cache-dir CACHE_DIR] --text TEXT [--output OUTPUT] [--precision PRECISION] [--seed SEED] [--prompt-audio PROMPT_AUDIO]
                [--prompt-text PROMPT_TEXT] [--language LANGUAGE] [--template-name {tts,instruction_tts,text_to_audio,tts_interleave}] [--ode-method ODE_METHOD] [--num-steps NUM_STEPS]
                [--guidance-scale GUIDANCE_SCALE] [--speaker-scale SPEAKER_SCALE] [--max-generate-length MAX_GENERATE_LENGTH] [--max-sequence-length MAX_SEQUENCE_LENGTH] [--normalize-text]
                [--profile-inference] [--profile-inference-calls]

dots.tts inference CLI.

options:
  -h, --help            show this help message and exit
  --model-name-or-path MODEL_NAME_OR_PATH
                        Local pretrained directory or Hugging Face repo id
  --revision REVISION   Optional Hugging Face revision
  --cache-dir CACHE_DIR
                        Optional Hugging Face cache dir
  --text TEXT           Input text
  --output OUTPUT       Output wav file path
  --precision PRECISION
                        Inference precision
  --seed SEED           Random seed for inference.
  --prompt-audio PROMPT_AUDIO
                        Path to prompt audio
  --prompt-text PROMPT_TEXT
                        Transcript of prompt audio
  --language LANGUAGE   Language tag mode. Default: none. Supported values: none, auto_detect, or a language code/name such as EN/en/english/chinese.
  --template-name {tts,instruction_tts,text_to_audio,tts_interleave}
                        Named template preset for generation.
  --ode-method ODE_METHOD
                        ODE solver method (default: artifact setting, otherwise euler)
  --num-steps NUM_STEPS
                        Sampling steps (default: artifact setting, otherwise 10)
  --guidance-scale GUIDANCE_SCALE
                        Classifier-free guidance scale (default: artifact setting, otherwise 1.2)
  --speaker-scale SPEAKER_SCALE
                        Scale applied to the reference speaker embedding
  --max-generate-length MAX_GENERATE_LENGTH
                        Maximum total audio patch count (prompt + generated)
  --max-sequence-length MAX_SEQUENCE_LENGTH
                        Maximum LLM sequence length used for schedule validation and optimized StaticCache
  --normalize-text      Whether to normalize text before inference
  --profile-inference   Collect per-module inference timing statistics
  --profile-inference-calls
                        Log each profiled inference call timing

Checkpoints

Seven pretrained checkpoints are released on Hugging Face. They share the same backbone; choose by task and runtime entry point.

Pass the repo id directly to the entry point shown above; the snapshot is fetched on first use and cached locally. Fixed-step artifacts reject incompatible sampling overrides.

Now, run the basic example script to synthesize speech:

# Random-voice sampling (no reference) — only meaningful with a fine-tuned
# single-speaker checkpoint
dots.tts --model-name-or-path dots-studio/dots.tts-soar --precision float32 --text "Hello, this is a quick speech synthesis test." --output output.wav

The result will be the audio file output.wav

Test Voice Cloning

Download the reference audio

curl -L -O "http://cva4.com/audio/prompt_english_female1.wav"

Reference audio sample voice: prompt_english_female1.wav

Now, run the script to clone reference audio:

# X-vector-only voice cloning (reference audio only — timbre from speaker x-vector)
dots.tts --model-name-or-path dots-studio/dots.tts-soar --precision float32 --text "Hello, this is a zero-shot voice cloning demonstration." --prompt-audio prompt_english_female1.wav --output clone.wav

# Continuation voice cloning (reference audio + transcript) — recommended, best SIM
dots.tts --model-name-or-path dots-studio/dots.tts-soar --precision float32 --text "Hello, this is a zero-shot voice cloning demonstration." --prompt-audio prompt_english_female1.wav --prompt-text "In order to win, you must expect to win." --output clone.wav

The output will be an audio file named clone.wav

Python Script

Create a file named dots_test.py:

from dots_tts.runtime import DotsTtsRuntime
import soundfile as sf

runtime = DotsTtsRuntime.from_pretrained(
    "dots-studio/dots.tts-soar",
    precision="float32",
    optimize=False,  # torch.compile acceleration (warmup at load, faster steady-state)
)

result = runtime.generate(
    text="Hello, this is a quick speech synthesis test.",
    prompt_audio_path="prompt_english_female1.wav",
    prompt_text="In order to win, you must expect to win.",
    num_steps=10,
    guidance_scale=1.2,
)

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

Run the script to clone voice:

python dots_test.py

The result will be the audio file clone.wav

Web Demo (Gradio)

Clone dots.tts from GitHub (run after activating the environment)

git clone https://github.com/studio-dots-ai/dots.tts.git
cd dots.tts

Try dots.tts without coding:

python apps/gradio/app.py --model-name-or-path dots-studio/dots.tts-soar --precision float32

Open your browser and navigate to http://127.0.0.1:7860. The system will automatically load the model weights during this first run. Please wait for the model to be ready!

DiaF5-TTS