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.

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.

News

• [2026.08] 🔥 We have released dots.tts.edit for precise, instruction-controlled speech editing — download the checkpoint, try the Playground, explore the Demo Page, and read the paper.

• [2026.08] ⚡ Released dots.tts-mf-2steps, dots.tts-mf-1step, and dots.tts-mf-2steps-stts for high-quality voice cloning and double-streaming TTS. These checkpoints build on dots.tts-mf with fixed-step train–inference alignment. See Checkpoints.

• [2026.08] 🚀 SGLang Omni now supports dots.tts (mf / mf-2steps / mf-2steps-stts / soar / base) with streaming PCM and CUDA-graph backbone decode. MeanFlow, two-step sCM, and STTS use continuous batching. Current Omni STTS serving consumes complete text or an already-collected token-ID sequence when each request starts; it does not yet accept incremental tokens into the same in-flight request. According to the SGLang Omni cookbook, on Seed-TTS-Eval EN (1× H100, dots.tts-mf, num_steps=4), peak throughput reaches 4.76 req/s / 19.86 audio_s/s at concurrency 16 (WER 1.35%). See SGLang Omni Usage and the cookbook.

• [2026.07] 🚀 Shipped a high-performance inference path — under --optimize, dots.tts-soar reaches RTF p50 0.20 / 0.18 and first-chunk latency 225 ms / 69 ms (voice cloning / text-only); dots.tts-mf reaches 0.15 / 0.13 and 204 ms / 68 ms respectively. See the Efficiency section for details.

• [2026.06] 🔥 We have released dots.tts — 2B fully continuous AR TTS, with pretrained / SOAR / MeanFlow-distilled checkpoints and full inference & fine-tuning code under Apache-2.0.

---

Contents

• Quick Start
  • Installation
  • CLI
  • Checkpoints
  • Python API
  • Web Demo (Gradio)
  • Fine-tuning
  • MeanFlow Distillation
  • SGLang Omni Usage
• Usage Tips
• Architecture
• Performance
  • Seed-TTS-Eval
  • MiniMax Multilingual
  • CV3-Eval
  • EmergentTTS-Eval
• Efficiency
  • SGLang Omni Efficiency
• Community Projects
• Risks and Limitations
• Citation
• License

---

🚀 Quick Start

Installation

We recommend a fresh conda environment (Python 3.10–3.12):

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

Install from PyPI:

pip install dots.tts

Or from source (for local development / editable install):

git clone https://github.com/studio-dots-ai/dots.tts
cd dots.tts
pip install -e . -c constraints/recommended.txt

For training / linting extras:

pip install 'dots.tts[full]'
# or from source:
pip install -e .[full] -c constraints/recommended.txt

The constraints/recommended.txt file pins the reproducible versions; pyproject.toml declares compatibility ranges.

To use SGLang Omni for high-performance high-concurrency voice cloning:

git clone git@github.com:sgl-project/sglang-omni.git
cd sglang-omni

uv venv .venv -p 3.12
source .venv/bin/activate

uv pip install -v -e .

Detailed installation instructions can be found in this guidance.

CLI

The package installs a dots.tts entry point. The examples below use dots.tts-soar; replace the model path with any checkpoint in the next section and apply its checkpoint-specific settings.

# Continuation voice cloning (reference audio + transcript) — recommended, best SIM
dots.tts \
  --model-name-or-path dots-studio/dots.tts-soar \
  --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." \
  --output clone.wav

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

# 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 \
  --text "Hello, this is a quick speech synthesis test." \
  --output output.wav

Common flags:

dots.tts --help lists the full set.

Speech editing uses the separate dots.tts.edit entry point. Source audio, a tagged instruction, and the output path are required. Source and target transcripts are optional: when omitted or blank, both are derived from the instruction.

dots.tts.edit \
  --model-name-or-path dots-studio/dots.tts.edit \
  --source-audio /path/to/source.wav \
  --instruction 'Hello <sub targ="small">brave</sub> world.' \
  --output edited.wav

Explicit non-empty --source-text and --target-text values override the derived transcripts. Edit speaker guidance defaults to auto: it is disabled when the instruction contains at least one operation and every operation is emo, bg, or enhance, and enabled for text, pitch, rate, pause, speaker transfer, or mixed edits. Pass bare --use-xvector (or --use-xvector on) to force it on, and --use-xvector off to force it off. TTS speaker guidance remains enabled when reference audio is provided. Supported structural tags include <del>, <ins>, <sub targ="replacement">, <emo>, <pitch>, <rate>, <enhance>, <bg>, <pause/>, and <spk_transfer/>. Malformed instructions and instructions that derive an empty transcript are rejected.

Notes:

• --prompt-audio selects the speaker voice — continuation cloning when paired with --prompt-text, x-vector-only cloning when used alone. Omitting --prompt-audio falls back to random-voice sampling, which is only meaningful on a fine-tuned single-speaker checkpoint.
• --language is useful for multilingual or code-switched text when you want to force the model-side language tag. For example, pass --language EN for English, --language ZH for Mandarin, --language Cantonese for Cantonese, or --language auto_detect to infer the tag from --text.
• Pass either a local model directory or a Hugging Face repo id.

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.

Python API

Basic TTS

from dots_tts.runtime import DotsTtsRuntime
import soundfile as sf

runtime = DotsTtsRuntime.from_pretrained(
    "dots-studio/dots.tts-soar",
    precision="bfloat16",
    optimize=True,  # torch.compile acceleration (warmup at load, faster steady-state)
)

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=10,
    guidance_scale=1.2,
)

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

The fixed-step MeanFlow artifacts read their sampling contracts directly from the model configuration, so CLI and Python calls do not need sampling options.

For low-latency playback or streaming to a client, use generate_stream instead — it yields audio chunks (torch.Tensor, shape (1, samples)) as they are produced. Arguments are identical to generate:

import torch

stream = runtime.generate_stream(
    text="Hello, this is a streaming speech synthesis test.",
    prompt_audio_path="/path/to/reference.wav",
    prompt_text="The exact transcript of the reference audio.",
    num_steps=10,
    guidance_scale=1.2,
)

chunks = []
for chunk in stream:
    chunks.append(chunk.detach().float().cpu())
    # handle_chunk(chunk)  # push to a player / websocket / etc.

audio = torch.cat(chunks, dim=-1).squeeze().numpy()
sf.write("output_stream.wav", audio, runtime.sample_rate)

Double Streaming

For duplex dialogue systems where an upstream LLM emits text tokens incrementally, use the double-streaming runtime. It accepts one text token at a time and returns either an audio chunk or None when the acoustic stream needs more text context.

The recommended released checkpoint for this path is dots-studio/dots.tts-mf-2steps-stts. Its sampling and streaming settings are stored in the artifact, so callers only provide text and optional prompt audio/text.

The same contract is available from Python:

import torch
import soundfile as sf

from dots_tts.runtime_double_streaming import DotsTtsRuntimeDoubleStreaming

runtime = DotsTtsRuntimeDoubleStreaming.from_pretrained(
    "dots-studio/dots.tts-mf-2steps-stts",
    precision="bfloat16",
    optimize=True,
    max_generate_length=500,
)

text = "你好呀，今天想聊点什么？或者有什么我能帮你的？"
text_token_ids = runtime.model.tokenizer.encode(text, add_special_tokens=False)

session = runtime.start_double_streaming(
    prompt_audio_path="/path/to/reference.wav",
    prompt_text="The exact transcript spoken in the reference audio.",
)

chunks = []
for token_id in text_token_ids:
    chunk = session.push_text_token(token_id)
    if chunk is not None:
        chunks.append(chunk.detach().cpu())

for chunk in session.finish_text():
    chunks.append(chunk.detach().cpu())

audio = torch.cat(chunks, dim=-1).float().squeeze().numpy()
sf.write("double_streaming.wav", audio, runtime.sample_rate)

--optimize is strongly recommended for double streaming. It adds a one-time torch.compile warmup at load time, but the steady-state path uses cached/compiled LLM, DiT, and vocoder steps and substantially reduces RTF and streaming gaps.

For a complete command-line example, see scripts/example_double_streaming.py.

Speech Editing

The same edit contract is available from Python:

from dots_tts.edit_runtime import DotsTtsEditRuntime

edit_runtime = DotsTtsEditRuntime.from_pretrained(
    "dots-studio/dots.tts.edit",
    precision="bfloat16",
)
result = edit_runtime.generate_edit(
    source_audio_path="/path/to/source.wav",
    instruction='Hello <sub targ="small">brave</sub> world.',
    # source_text and target_text are optional overrides.
    # use_xvector defaults to "auto"; pass True or False to override it.
    num_steps=10,
    guidance_scale=1.2,
)
sf.write("edited.wav", result["audio"].float().cpu().squeeze().numpy(), result["sample_rate"])

Web Demo (Gradio)

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

Defaults to http://0.0.0.0:7860. With --optimize the first launch runs warmup (slower startup, faster steady-state).

For the local Edit Playground, build the frontend once with Node.js 20+ and then launch the application:

cd apps/edit_playground/frontend
npm ci
npm run build
cd ../../..
python apps/edit_playground/app.py \
  --model-name-or-path dots-studio/dots.tts.edit \
  --optimize

The Edit Playground ships without voice, edit-source, or noise audio presets; upload your own source/reference audio. Optional local transcription requires the ASR extra:

python -m pip install -e '.[edit_playground_asr]' -c constraints/recommended.txt
python apps/edit_playground/app.py \
  --model-name-or-path dots-studio/dots.tts.edit \
  --asr-model Qwen/Qwen3-ASR-1.7B

The generated frontend/dist directory is intentionally not committed. Pass --rebuild-frontend to install dependencies and rebuild it explicitly during launch. See apps/edit_playground/README.md for the frontend test, build, and local development workflow.

Fine-tuning

This repo exposes fine-tuning and MeanFlow distillation entry points. Fine-tune from a released checkpoint with:

accelerate launch scripts/train_dots_tts.py --config configs/dots_tts.yaml

configs/dots_tts.yaml is a smoke configuration that verifies the pipeline runs end-to-end on commodity hardware. Replace train.pretrained_model_path, train_data.sources / val_data.sources, train.output_dir, and train.max_train_steps with your own values to use it.

A helper script downloads LJSpeech-1.1-48kHz and emits a train/valid JSONL manifest for the smoke run:

python scripts/prepare_train_jsonl_manifest.py --output-dir downloaded_data

Manifest format — one JSON per line, minimum three fields:

{"fid": "sample-0001", "audio": "/abs/path/to/audio.wav", "text": "hello world"}

MeanFlow Distillation

MeanFlow distillation trains a MeanFlow DiT student against a frozen flow-matching teacher. The teacher can be the released SOAR checkpoint or any compatible flow-matching dots.tts checkpoint you have fine-tuned yourself.

To use SOAR as the teacher, download it first:

huggingface-cli download dots-studio/dots.tts-soar \
  --local-dir pretrained_models/dots.tts-soar

Then launch distillation with the MeanFlow config:

accelerate launch \
  --num_processes 2 \
  --mixed_precision bf16 \
  scripts/train_dots_tts_meanflow.py \
  --config configs/dots_tts_meanflow.yaml \
  --teacher-model-path pretrained_models/dots.tts-soar

To distill from your own fine-tuned teacher, pass that checkpoint instead:

accelerate launch \
  --num_processes 2 \
  --mixed_precision bf16 \
  scripts/train_dots_tts_meanflow.py \
  --config configs/dots_tts_meanflow.yaml \
  --teacher-model-path /path/to/your_finetuned_teacher

configs/dots_tts_meanflow.yaml is a conservative smoke configuration that uses the same LJSpeech manifests produced by scripts/prepare_train_jsonl_manifest.py. Replace train.pretrained_model_path, --teacher-model-path, train_data.sources / val_data.sources, train.output_dir, and train.max_train_steps for your own distillation run.

By default, the script initializes the student from train.pretrained_model_path, adds the MeanFlow duration embedding, freezes the non-DiT modules, and trains student.core.velocity_field_predictor. MeanFlow does not run a separate CFG branch at inference time; the default fused mode distills the guided teacher target into the student. Training checkpoints save the MeanFlow student only; the frozen teacher is not written into the checkpoint model directory. Pass --train-all-parameters only if you want to update the full dots.tts model.

Common MeanFlow flags:

---

SGLang Omni Usage

SGLang Omni serves dots.tts behind an OpenAI-compatible /v1/audio/speech API with continuous batching for MeanFlow, two-step sCM, and prebuilt STTS interleave schedules, plus streaming PCM and CUDA-graph backbone decode. SGLang Omni does not yet expose the model's end-to-end, same-request token/audio double-streaming runtime. Full details live in the SGLang Omni dots.tts cookbook.

Install Omni as in SGLang Omni Installation, then from the sglang-omni checkout:

hf download dots-studio/dots.tts-mf

sgl-omni serve \
  --model-path dots-studio/dots.tts-mf \
  --config examples/configs/dots_tts.yaml \
  --port 8000

Use the model-specific config for two-step sCM or STTS:

# Fixed two-step sCM
sgl-omni serve \
  --model-path dots-studio/dots.tts-mf-2steps \
  --config examples/configs/dots_tts_scm.yaml \
  --port 8000

# Streaming TTS with artifact-defined text/audio interleave
sgl-omni serve \
  --model-path dots-studio/dots.tts-mf-2steps-stts \
  --config examples/configs/dots_tts_stts.yaml \
  --port 8000

Use the config file — it enables the compiled acoustic tail / vocoder and backbone decode CUDA graph. MeanFlow, sCM, and STTS use continuous batching; SOAR/base remain single-request because their CFG conditional/unconditional branches are not yet implemented by the batched acoustic tail.

Voice cloning (reference audio + transcript required):

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dots-studio/dots.tts-mf",
    "input": "Have a nice day and enjoy south california sunshine.",
    "references": [{
      "audio_path": "docs/_static/audio/male-voice.wav",
      "text": "Hey, Adam here. Let'\''s create something that feels real, sounds human, and connects every time."
    }],
    "seed": 42
  }' \
  --output output.wav

import requests

resp = requests.post(
    "http://localhost:8000/v1/audio/speech",
    json={
        "model": "dots-studio/dots.tts-mf",
        "input": "Have a nice day and enjoy south california sunshine.",
        "references": [{
            "audio_path": "docs/_static/audio/male-voice.wav",
            "text": "Hey, Adam here. Let's create something that feels real, sounds human, and connects every time.",
        }],
        "seed": 42,
    },
)
resp.raise_for_status()
with open("output.wav", "wb") as f:
    f.write(resp.content)

ref_audio / ref_text are accepted as a shorthand for references[0].audio_path / references[0].text.

The same request format works with dots.tts-mf-2steps and dots.tts-mf-2steps-stts. STTS requires reference audio plus its transcript and reads its interleave cadence from the checkpoint. Each Omni request must start with complete text or an already-collected text_token_ids (or input_ids) array. The /v1/audio/speech/stream WebSocket can receive text chunks, but it buffers them and starts separate TTS requests at sentence/clause boundaries; it does not inject new tokens into an in-flight STTS request. Therefore Omni does not yet provide same-request, token-by-token double-streaming inference.

Streaming (raw 48 kHz PCM; set "stream": true and "response_format": "pcm"):

curl -N -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{
    "model": "dots-studio/dots.tts-mf",
    "input": "Get the trust fund to the bank early.",
    "references": [{
      "audio_path": "docs/_static/audio/female-voice.wav",
      "text": "By repeating what students say, teachers can demonstrate that they are listening. By extending what students say."
    }],
    "stream": true,
    "response_format": "pcm",
    "seed": 42
  }' \
  --output output.pcm

ffmpeg -f s16le -ar 48000 -ac 1 -i output.pcm output.wav

Solver knobs (speaker_scale, guidance_scale, eos_threshold, num_steps, …) go under stage_params.latent_engine — not as top-level fields. temperature / top_p / top_k do not apply (continuous latent; no token sampler). MeanFlow fixes num_steps=4 engine-wide for continuous batching. The two-step sCM and STTS checkpoints read Euler, NFE 2, CFG 0, tau_mid, and (for STTS) cadence from the artifact and reject incompatible overrides.

💡 Usage Tips

• Keep the reference audio around 10s. Longer audio won't yield better results.
• --prompt-text should match what's actually spoken in the reference audio. Mismatches degrade stability and may cause word-level errors.
• Higher-quality references give better clones — prefer a high sample rate, low background noise, no trailing noise, and natural-sounding speech.
• Try different --seed values for prosody variation. Each seed produces a different rhythm and intonation — resample a few times if the default doesn't feel right.
• For flow-matching checkpoints, increase --num-steps if quality isn't good enough. Fixed-step MeanFlow artifacts reject incompatible sampling overrides.
• Force a pronunciation with Pinyin for polyphones. Replace the character in the input text with its tone-marked pinyin — e.g. write 我生平不hào此道 to force 好 to be read as hào. Use tone-marked pinyin only (hǎo, hào, bā); numbered forms like hao4 or ha4o are not recognized. Useful when reseeding doesn't fix a polyphone misread.

---

🏛 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, in three components:

• 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), and 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.

Two sequence layouts: plain mode places the full text as a prefix before the audio span (standard TTS); double-streaming interleaved mode lets a caller push BPE text tokens incrementally while audio patches are decoded online using the checkpoint's declared streaming cadence. See the technical report for full architectural and training details.

---

📊 Performance

Baselines are taken from original publications or default-configuration open-source releases.

Seed-TTS-Eval

Zero-shot, ~3 s reference prompt, scored by the benchmark's reference ASR and WavLM-SV similarity.

MiniMax Multilingual (24 languages)

Per-language WER / SIM on the MiniMax-Speech multilingual test set (100 utterances × 2 reference speakers per language). Highest average SIM (83.9, SOAR), with a dots.tts variant taking the per-language SIM lead outright on 19 of 24 languages and tying on 2 more. Content fidelity is on par with the strongest systems on high-resource / Western European splits, and trails on low-resource long-tail languages where SIM is still preserved.

Per-language WER / SIM (click to expand)

Language	MiniMax	ElevenLabs	Fish-Audio S2	VoxCPM 2	dots.tts (Pre.)	dots.tts (SOAR)	dots.tts (MF$_4$)
Arabic	1.67 / 73.6	1.67 / 70.6	3.50 / 75.0	13.05 / 79.1	37.91 / 77.5	36.19 / 79.1	39.65 / 77.6
Cantonese*	34.11 / 77.8	51.51 / 67.0	30.67 / 80.5	38.58 / 83.5	37.91 / 84.7	42.32 / 85.0	37.82 / 84.0
Chinese	2.25 / 78.0	16.03 / 67.7	0.73 / 81.6	1.14 / 82.5	1.08 / 82.3	0.77 / 82.5	1.01 / 81.8
Czech	3.88 / 79.6	2.11 / 68.5	2.84 / 79.8	24.13 / 78.3	5.05 / 83.8	4.25 / 84.2	5.67 / 83.9
Dutch	1.14 / 73.8	0.80 / 68.0	0.99 / 73.0	0.91 / 80.8	1.20 / 81.4	1.39 / 82.2	1.30 / 82.1
English	2.16 / 75.6	2.34 / 61.3	1.62 / 79.7	2.29 / 85.4	1.06 / 86.9	1.03 / 87.5	1.09 / 86.9
Finnish	4.67 / 83.5	2.96 / 75.9	3.33 / 81.9	2.63 / 89.0	3.44 / 88.0	4.08 / 88.3	3.61 / 88.3
French	4.10 / 62.8	5.22 / 53.5	3.05 / 69.8	4.53 / 73.5	3.82 / 78.2	3.56 / 78.6	3.26 / 78.5
German	1.91 / 73.3	0.57 / 61.4	0.55 / 76.7	0.68 / 80.3	1.03 / 79.5	1.70 / 80.6	0.91 / 79.5
Greek	2.02 / 82.6	0.99 / 73.3	5.74 / 79.5	2.84 / 86.0	2.97 / 87.6	3.00 / 87.6	3.19 / 87.3
Hindi	6.96 / 81.8	5.83 / 73.0	14.64 / 82.1	19.70 / 85.6	14.32 / 84.5	14.24 / 84.7	14.75 / 84.8
Indonesian	1.24 / 72.9	1.06 / 66.0	1.46 / 76.3	1.08 / 80.0	2.71 / 80.8	2.96 / 80.8	3.91 / 81.2
Italian	1.54 / 69.9	1.74 / 57.9	1.27 / 74.7	1.56 / 78.0	3.16 / 84.5	3.12 / 84.7	2.16 / 84.3
Japanese	3.52 / 77.6	10.65 / 73.8	2.76 / 79.6	4.63 / 82.8	7.16 / 83.1	5.28 / 83.7	5.17 / 83.1
Korean	1.75 / 77.6	1.87 / 70.0	1.18 / 81.7	1.96 / 83.3	5.30 / 84.3	5.66 / 83.6	3.93 / 84.9
Polish	1.42 / 80.2	0.77 / 72.9	1.26 / 81.9	1.14 / 88.4	2.72 / 87.3	3.59 / 87.8	3.42 / 87.5
Portuguese	1.88 / 80.5	1.33 / 71.1	1.14 / 78.1	1.94 / 83.7	1.64 / 83.1	2.00 / 84.3	2.40 / 83.1
Romanian	2.88 / 80.9	1.35 / 69.9	10.74 / 73.3	21.58 / 79.7	3.36 / 86.2	3.87 / 87.1	3.38 / 86.1
Russian	4.28 / 76.1	3.88 / 67.6	2.40 / 79.0	3.63 / 81.1	3.64 / 83.0	4.28 / 83.2	4.42 / 83.2
Spanish	1.03 / 76.2	1.08 / 61.5	0.91 / 77.6	1.44 / 83.1	0.96 / 83.9	1.27 / 84.0	0.80 / 84.0
Thai	2.70 / 80.0	73.94 / 58.8	4.23 / 78.6	2.96 / 84.0	7.45 / 83.8	7.86 / 83.9	8.03 / 84.2
Turkish	1.52 / 77.9	0.70 / 59.6	0.87 / 83.5	0.82 / 87.1	5.45 / 87.4	4.96 / 87.3	6.20 / 86.8
Ukrainian	1.08 / 73.0	1.00 / 64.7	2.30 / 74.7	6.32 / 79.8	1.61 / 80.5	1.27 / 81.2	1.66 / 80.0
Vietnamese	0.88 / 74.3	73.42 / 36.9	7.41 / 74.0	3.31 / 80.6	3.85 / 80.7	3.89 / 81.6	5.43 / 80.5
Average	2.8 / 76.6	7.5 / 65.5	3.7 / 78.0	5.7 / 82.3	6.6 / 83.5	6.8 / 83.9	6.8 / 83.5

_{*Cantonese WER reflects an ASR-faithfulness floor common to all systems; SIM remains comparable.}

CV3-Eval

Hard-subset Chinese/English plus a cross-lingual voice-cloning split. Takes the table top on hard-en (MF$_4$ at 4.37) and leads both cross-lingual SIM subsets (SOAR at 75.0 / 72.8), with the post-trained variants bracketing the prior leader on the hardest English subset.

EmergentTTS-Eval

Win-rate judged head-to-head against gpt-4o-mini-tts by Gemini-2.5-Pro-0506 across six expressiveness-oriented scenarios. SOAR takes the top Syntactic Complexity score in the table (65.7%) — above every closed-source system — and Pretrain posts the best Emotions score among open-source systems (72.7%).

_{* Closed-source / commercial. Table shows a selected subset for brevity — for the full leaderboard, see EmergentTTS-Eval-public.}

---

⚡ Efficiency

Streaming-inference benchmarks under --optimize on a Seed-TTS-Eval mix (100 utterances across zh / en / zh-hard, first post-warmup request excluded, N=99 per group). voice_cloning uses reference audio + transcript; text_only uses text with no reference. Common config: precision=bfloat16, guidance_scale=1.2, seed=42; SOAR uses num_steps=10, MF uses num_steps=4. Hardware / stack: single H800, torch 2.8 + CUDA 12.8. The --optimize path also accelerates non-streaming generate() calls; numbers below are the streaming path.

Note: --optimize triggers a one-shot torch.compile warmup that walks every DiT compile bucket + KvPrefill + vocoder chunk sizes. Cold start takes ~3 minutes on H800; every subsequent request runs at the steady-state RTF above. Pass warmup_on_optimize=False to DotsTtsRuntime if you want to skip warmup and accept the first request paying the compile cost.

Steady-State Latency

Memory Footprint by Length Bucket

Bucket = total prompt + generated audio in latent patches (one patch ≈ 160 ms).

* From explicit long-audio probes (actual spans within 256–512 patches).
** mf / text_only did not reach the 256–512 bucket under either synthetic (x4) or real long-text probes; longest observed 237 patches at 5.73 GB.

SGLang Omni Efficiency

Serving throughput on Seed-TTS-Eval EN against a single Omni server started from examples/configs/dots_tts.yaml (max_running_requests=16, bf16, num_steps=4, backbone decode CUDA graph + graph-captured acoustic tail). Each row is the mean of two runs, seed 42. Hardware: 1× H100. Full write-up: SGLang Omni cookbook — Performance.

Zero failed requests in every run, and no sample above 50% WER. WER is measured with Qwen/Qwen3-ASR-1.7B on the first run of each row.

To reproduce (server already running as above):

python -m benchmarks.eval.benchmark_tts_seedtts \
  --meta zhaochenyang20/seed-tts-eval-arrow \
  --model dots-studio/dots.tts-mf \
  --ref-format references \
  --base-url http://127.0.0.1:8000 --port 8000 \
  --lang en --max-concurrency 16 --warmup 8 --seed 42 \
  --generate-only --use-existing-server \
  --output-dir results/dots-seedtts-en-c16

python -m benchmarks.eval.benchmark_tts_seedtts \
  --meta zhaochenyang20/seed-tts-eval-arrow \
  --model dots-studio/dots.tts-mf \
  --ref-format references --lang en --seed 42 \
  --transcribe-only --port 8000 \
  --output-dir results/dots-seedtts-en-c16

🤝 Community Projects

Third-party ports and integrations of dots.tts, maintained by the community.

---

⚠️ Risks and Limitations

• Misuse risk. High-fidelity zero-shot voice cloning can produce highly realistic synthetic speech. The released checkpoints are intended for research and authorized deployment. Do not use dots.tts 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.
• 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) the WER gap visible on the MiniMax benchmark reflects this, and the same long tail surfaces on the Foreign Words and Complex Pronunciation scenarios of EmergentTTS-Eval. Speaker similarity is preserved across these languages.
• Speech-heavy training. Although the AudioVAE is trained at 48 kHz and is modality-agnostic in principle, the backbone is trained on a speech-heavy mixture. Singing and unified speech + sound generation are not covered in this release.

---

📖 Citation

If you find dots.tts or dots.tts.edit useful, please consider citing the corresponding technical report and starring the repository.

@article{dotstts2026,
  title         = {dots.tts Technical Report},
  author        = {dots.tts Team},
  year          = {2026},
  eprint        = {2606.07080},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
}

@article{wang2026dotsttsedit,
  title         = {dots.tts.edit: Precisely Controlled Speech Editing with a Continuous Autoregressive Model},
  author        = {Wang, Hankun and Li, Bohan and Lian, Shi and Gu, Xiaoyu and Peng, Jing and Zheng, Da and Zhang, Colin and Yu, Kai},
  year          = {2026},
  eprint        = {2608.02673},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
}

📄 License

dots.tts code and released checkpoints are licensed under Apache-2.0.

🙏 Acknowledgments

• Qwen2.5 — LLM backbone initialization.
• DiTAR and ARDiT — for the continuous-AR + per-patch diffusion design.
• HoliTok — for the AudioVAE design.
• BigVGAN — for the vocoder design.
• CAM++ — for speaker x-vector encoder.