MOSS-TTS v1.5：31 种语言、显式停顿控制、更稳的声音克隆，Apache 2.0 开源

复旦 NLP、上海创新研究院和模思智能联合的 OpenMOSS 团队把 MOSS-TTS v1.5 推上来了。v1.0 的底子已经在 Seed-TTS-eval 上跑出了开源 SOTA，v1.5 没有大改架构，而是做了一轮开发者真正在意的针对性打磨。

📌 GitHub：https://github.com/OpenMOSS/MOSS-TTS
HuggingFace v1.5：https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5
技术报告：https://arxiv.org/pdf/2603.18090
协议：Apache 2.0（商用友好）

v1.5 的四个实质变化

多语言扩到 31 种

保留 v1.0 的 20 种语言，新增：粤语、荷兰语、芬兰语、印地语、马其顿语、马来语、罗马尼亚语、斯瓦希里语、他加禄语、泰语、越南语。

调用时显式带上语言标签，几乎所有语言都比 1.0 有提升：

build_user_message(text="Bonjour le monde", language="French")

声音克隆更稳、方差更小

同一段参考音频反复生成，音色一致性明显提升。更重要的是，以前容易翻车的 “参考音频远长于目标文本” 场景（比如用 30 秒音频只克隆一句话）现在更可靠了——这在实际有声书和播客制作中是个高频痛点。

显式停顿控制：[pause 3.2s]

直接在文本里写停顿标记，精确到 0.1 秒：

它的名字是[pause 3.2s]静夜思！

朗诵、口播、有声书可以按节奏一字不改地排版，不需要后期剪辑对齐。

标点韵律更准

长句里逗号、顿号、句号的停顿时长更贴近真人语感，不再机械地一路平推过去。这个改进对普通话长句影响最明显。

MOSS-TTS 的技术路线

MOSS-TTS 的技术选型刻意保持克制：高质量音频 tokenizer + 自回归建模 + 大规模预训练，跟着 LLM 的范式走，没有堆外挂语义教师、没有多阶段精修流水线。

底层 MOSS-Audio-Tokenizer 把 24 kHz 音频压到 12.5 fps，32 层 RVQ 可变码率，参数量 1.6B。上层建模用纯 Causal Transformer（无 CNN），训练语料覆盖播客、有声书、影视、新闻等百万小时量级。

提供两种推理模式：

• Delay pattern：更快、长文本更稳定
• Local：参数更小，客观指标更好

另有 MOSS-TTS-Nano（0.1B 参数），设计目标是 4 核 CPU 端侧部署，不依赖 GPU。

部署支持

后端	场景
PyTorch	训练/研究/GPU 推理
GGUF（llama.cpp）	CPU 端侧部署
ONNX	跨平台推理
mlx-audio	Apple Silicon（Mac）

环境要求：Python 3.12 + Transformers 5.0.0+，可选 FlashAttention 2 加速。

conda create -n moss-tts python=3.12
pip install -r requirements.txt

权重通过 AutoModel.from_pretrained() 加载，推理接口支持批量处理。

基准表现

在 Seed-TTS-eval 零样本 TTS 评估中，MOSS-TTS v1.0 已经超越所有开源模型，接近最强闭源系统的水平（WER + 说话人相似度双指标）。v1.5 在此基础上进一步提升多语言和克隆稳定性。

适合的场景

• 有声书 / 播客：停顿控制 + 韵律改进直接上手可用
• 虚拟人 / 数字人：多语言 + 稳定克隆是基础要求
• 配音工具：Apache 2.0 协议，商用无障碍
• 端侧部署：Nano 版本 + GGUF 支持设备端运行

---

© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。

OpenMOSS — a joint team from Fudan NLP, Shanghai Innovation Institute, and MOSI.AI — has released MOSS-TTS v1.5. The v1.0 foundation already achieved open-source SOTA on Seed-TTS-eval; v1.5 keeps the architecture intact and delivers targeted improvements developers actually care about.

📌 GitHub: https://github.com/OpenMOSS/MOSS-TTS
HuggingFace v1.5: https://huggingface.co/OpenMOSS-Team/MOSS-TTS-v1.5
Technical report: https://arxiv.org/pdf/2603.18090
License: Apache 2.0 (commercial-friendly)

Four Substantive Changes in v1.5

Multilingual Support Expanded to 31 Languages

Retains all 20 languages from v1.0, adds 11 new ones: Cantonese, Dutch, Finnish, Hindi, Macedonian, Malay, Romanian, Swahili, Tagalog, Thai, Vietnamese.

With explicit language tags, v1.5 outperforms v1.0 on almost every supported language:

build_user_message(text="Bonjour le monde", language="French")

Stabler Voice Cloning, Lower Variance

Repeated generation from the same reference audio produces much more consistent voice characteristics. The previously unreliable “long reference, short target” scenario — e.g., cloning from 30 seconds of audio to synthesize a single sentence — is now significantly more reliable.

Explicit Pause Control: [pause 3.2s]

Insert pauses directly in the text, precise to 0.1 seconds:

Its name is[pause 3.2s]Jing Ye Si!

Useful for narration, podcasts, and audiobooks where timing matters — no post-processing required.

More Natural Punctuation Prosody

Pause durations for commas, enumeration pauses, and periods in long sentences now closely match natural human speech patterns. No more mechanical flat delivery through multi-clause sentences.

Technical Architecture

MOSS-TTS takes a deliberately minimal approach: high-quality audio tokenizer + autoregressive modeling + large-scale pretraining — following the LLM paradigm without stacking semantic teacher models or multi-stage refinement pipelines.

The underlying MOSS-Audio-Tokenizer compresses 24 kHz audio to 12.5 fps with 32-layer variable-bitrate RVQ (1.6B parameters). The upper modeling layer uses pure Causal Transformer blocks (no CNN). Training data covers podcasts, audiobooks, film/TV, and news at million-hour scale.

Two inference modes:

• Delay pattern: faster, more stable for long-form synthesis
• Local: smaller model, better on objective metrics

MOSS-TTS-Nano (0.1B parameters) targets CPU-only deployment on 4 cores, no GPU required.

Deployment Options

Backend	Use Case
PyTorch	Training / research / GPU inference
GGUF (llama.cpp)	CPU edge deployment
ONNX	Cross-platform inference
mlx-audio	Apple Silicon (Mac)

Requirements: Python 3.12 + Transformers 5.0.0+, optional FlashAttention 2 for speed.

Benchmark Results

On Seed-TTS-eval zero-shot TTS evaluation, MOSS-TTS v1.0 already outperformed all open-source models and rivaled the strongest closed-source systems (WER + speaker similarity). v1.5 builds further on multilingual quality and cloning stability.

Target Use Cases

• Audiobooks / podcasts: pause control and prosody improvements are immediately useful
• Virtual avatars / digital humans: multilingual + stable cloning are baseline requirements
• Dubbing tools: Apache 2.0 means no commercial restrictions
• Edge deployment: Nano + GGUF enables on-device inference

---

© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.