X2Streaming-TTS-1.7B

<div align="center"> <img src="assets/x-square-logo.png" width="140" alt="X Square Robot beaver mascot holding the moon"> <h1>X2Streaming-TTS-1.7B</h1> <p><strong>The deployed CustomVoice checkpoint behind X2Streaming-TTS</strong></p> <p> <a href="https://github.com/X-Square-Robot/X2Streaming-TTS"><img src="https://img.shields.io/badge/GitHub-X2Streaming--TTS-black" alt="X2Streaming-TTS"></a> <a href="https://github.com/X-Square-Robot/Qwen3TTS-Streaming"><img src="https://img.shields.io/badge/Engine-Qwen3TTS--Streaming-6f42c1" alt="Qwen3TTS-Streaming"></a> <a href="https://arxiv.org/abs/2608.18661"><img src="https://img.shields.io/badge/arXiv-2608.18661-b31b1b" alt="paper"></a> <a href="https://huggingface.co/x-square-robot/X2-NativeCursor-Qwen3TTS-12Hz"><img src="https://img.shields.io/badge/Hugging%20Face-X2--NativeCursor-yellow" alt="X2-NativeCursor head"></a> <a href="https://arxiv.org/abs/2609.09677"><img src="https://img.shields.io/badge/arXiv-2609.09677-b31b1b" alt="X2-NativeCursor paper"></a> </p> </div> <div align="center"> <img src="assets/native_cursor_use_cases.png" width="1000" alt="X2Streaming-TTS and X2-NativeCursor: streaming speech, continuous narration, synchronized highlighting, playback-aware interruption, and dialogue-history updates"> <p><em>Speak as text arrives, track reading progress, and use the playback clock to keep interruptions and dialogue history aligned with played audio.</em></p> </div>

X2Streaming-TTS-1.7B is a Qwen3-TTS 12 Hz CustomVoice model with one built-in Mandarin voice, robot_service_v1, fine-tuned by X Square Robot from Qwen3-TTS-12Hz-1.7B-Base. It is the checkpoint served by the Qwen3TTS-Streaming engine in X Square Robot's spoken-dialogue stack and the checkpoint the X2Streaming-TTS method and the X2-NativeCursor progress observer are released against.

The files are in the standard Qwen3-TTS Hugging Face layout, so the checkpoint loads with the qwen_tts Python package for offline synthesis and exports unchanged through the Qwen3TTS-Streaming pipeline for token-level streaming on TensorRT.

Model details

Files

SHA256SUMS in the repository lists the same digests in sha256sum format.

Quick start

Download

pip install -U "huggingface_hub[cli]"
huggingface-cli download x-square-robot/X2Streaming-TTS-1.7B \
  --local-dir ./weights/X2Streaming-TTS-1.7B

Offline synthesis with qwen_tts

Install the official package from QwenLM/Qwen3-TTS, then call generate_custom_voice with the built-in speaker:

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "./weights/X2Streaming-TTS-1.7B",
    device_map="cuda:0",
    dtype=torch.bfloat16,
)
print(model.get_supported_speakers())   # ['robot_service_v1']

wavs, sr = model.generate_custom_voice(
    text="您好，我是服务助手，今天的室外温度是二十三摄氏度。",
    language="Chinese",
    speaker="robot_service_v1",
)
sf.write("output.wav", wavs[0], sr)

generation_config.json carries the sampling defaults used in production: temperature 0.9, top_k 50, top_p 1.0, repetition_penalty 1.05.

Token-level streaming with Qwen3TTS-Streaming

The engine's autorun pipeline exports the checkpoint to ONNX, builds the TensorRT engines and starts the service. The custom-1.7b variant reads its weights from workspace/models/Qwen3-TTS-12Hz-1.7B-CustomVoice, so place (or symlink) the downloaded directory there and skip the official download:

git clone --recursive https://github.com/X-Square-Robot/Qwen3TTS-Streaming.git
cd Qwen3TTS-Streaming
mkdir -p workspace/models
ln -s /path/to/weights/X2Streaming-TTS-1.7B workspace/models/Qwen3-TTS-12Hz-1.7B-CustomVoice
SKIP_MODELS=1 bash scripts/bash/autorun.sh all -m custom-1.7b --model-release-version cont7e3-f2b8ed5

Once the service is up, GET /v1/capabilities lists the exported speaker and the supported tasks; the Python and browser SDKs, the native WebSocket protocol and the OpenAI Realtime-compatible gateway are documented in the engine's user guide and deployment guide.

With the X2Streaming-TTS method

X2Streaming-TTS adds causal commitment and causal speech-state inheritance on top of the engine through two hook patches. Its README walks through applying the patches and running scripts/run_checkpoint_e2e.py against this checkpoint.

With the X2-NativeCursor progress observer

X2-NativeCursor-Qwen3TTS-12Hz is an 8.2 MB observer head that reads this model's codebook-0 tokens inside the engine and publishes a reading cursor into the source text. Its model card explains how to enable it (text_progress.estimator: native). The X2-NativeCursor paper describes the observer and its evaluation.

Training summary

The voice was obtained by fine-tuning Qwen3-TTS-12Hz-1.7B-Base in two stages on an internal synthetic Mandarin corpus for a single service-assistant speaker:

The speech tokenizer (speech_tokenizer/) is the official Qwen3-TTS-Tokenizer-12Hz, included unchanged so that the directory is self-contained.

Intended use and limitations

• Intended use. Mandarin spoken-dialogue and service-assistant applications that need low first-audio latency, in particular token-level streaming from an upstream language model through Qwen3TTS-Streaming.
• Single voice. The checkpoint contains one speaker. Instruction-based style control (instruct) is supported by the architecture but was not a training target for this voice.
• English. English text is synthesized, with lower naturalness than Mandarin.
• Streaming caveats. Token-level streaming can still produce hallucinated, repeated or dropped words on some inputs; the rate depends on the text and the segmentation policy. The engine's known limitations and the X2Streaming-TTS paper describe the failure modes and the mitigations (causal commitment, health-gated state inheritance).
• Numbers, dates and units. Read the text through a text normalizer before synthesis; the X2Streaming-TTS release includes the Chinese normalizer used in the paper. Some date, time and unit formats are misread when passed raw.
• Evaluation. The streaming evaluation protocol and results for the method are in the paper.

Citation

@article{wen2026x2streamingtts,
  title   = {X2Streaming-TTS: Causal Token-Level Text-to-Speech from Streaming Text with Speech-State Inheritance},
  author  = {Wen, Rime and Liu, Zehan and Qin, Shawn and Shi, Lights and Gan, Roy and Wang, Hao and Wang, Qian},
  journal = {arXiv preprint arXiv:2608.18661},
  year    = {2026},
}

If you use the X2-NativeCursor progress observer with this checkpoint, also cite:

@article{liu2026x2nativecursor,
  title   = {X2-NativeCursor: Native-Token Text Progress Tracking for Incremental-Text Streaming Codec TTS},
  author  = {Liu, Zehan and Chen, Carl and Wen, Rime and Fu, Kaiqi and Lin, Altman and Qin, Shawn and Shi, Lights and Gan, Roy and Wang, Hao and Wang, Qian},
  journal = {arXiv preprint arXiv:2609.09677},
  year    = {2026},
  url     = {https://arxiv.org/abs/2609.09677},
}

License

See LICENSE for the full Apache-2.0 terms and NOTICE for component provenance and attribution.

The weights are released under the Apache License 2.0, Copyright (c) 2026 XSquareRobot. They derive from Qwen3-TTS-12Hz-1.7B-Base and include the Qwen3-TTS-Tokenizer-12Hz, both released by the Qwen team under Apache-2.0; their notices apply to the corresponding components.

中文简介

X2Streaming-TTS-1.7B 是 X Square Robot 基于 Qwen3-TTS-12Hz-1.7B-Base 微调的 CustomVoice 模型，内置一个中文服务助手音色 robot_service_v1。它是 Qwen3TTS-Streaming 引擎在线服务所用的 checkpoint，也是 X2Streaming-TTS 方法与 X2-NativeCursor 进度观察器 发布时所对应的权重。文件采用 Qwen3-TTS 的标准 Hugging Face 布局：用 qwen_tts 包可以 离线合成，交给 Qwen3TTS-Streaming 的 autorun 流水线即可导出 TensorRT 做 token 级流式 合成。合成前请先做文本正则化（数字、日期、单位）。许可为 Apache-2.0。

论文：X2Streaming-TTS；如使用朗读进度观察器， 请同时引用 X2-NativeCursor。