dots.tts-mf generating 2-3 seconds of garbled audio after short text synthesis.
Raw Developer Origin & Technical Request
GitHub Issue
Jun 13, 2026
## 现象
dots.tts-mf 在**短文本**合成时,模型说完原句后**不停止**,继续生成约 2-3 秒乱码音频——听起来像中文胡言,有时混入英文单词(如 "technology"),有时 ASR 完全识别不出(如把"我吃饭吗。"识别成"诺克萨曼蒙")。
**关键对比**(附件 wav):
| 输入 | 时长 | ASR 识别 | 结果 |
|---|---|---|---|
| `"我吃饭吧?"` (5字+问号) | 1.44s | "我吃饭吧。" | ✅ 干净说完即停 |
| `"我吃饭吗。"` (5字+句号) | 3.20s | **"诺克萨曼蒙"** | ⚠️ 外语乱码 |
| `"这怎么说?"` (5字+问号含"怎么") | 3.73s | (中文胡言) | ⚠️ 问号也救不了 |
| `"嗯?"` (1字+问号) | 3.20s | (无法识别) | ⚠️ 1字 100% 必乱 |
## 复现
环境:
- 模型:`rednote-hilab/dots.tts-mf`
- 硬件:DGX Spark (GB10 / aarch64 / CUDA 13.0)
- PyTorch:2.12.0+cu130
- 配置:`precision="bfloat16", optimize=True, num_steps=6, guidance_scale=1.2, speaker_scale=1.5`
- `torch.cuda.set_per_process_memory_fraction(0.75)` 已设
- `prompt_text=None`(x-vector-only cloning)
- 参考音频 ≤6s 真人录音
```python
result = runtime.generate(
text="我吃饭吗。", # 5字+句号 → 触发乱码
prompt_audio_path="ref.wav",
prompt_text=None,
language="ZH",
num_steps=6,
guidance_scale=1.2,
speaker_scale=1.5,
)
# 期望 ~1.5s,实际 3.2s,后 1.7s 是外语乱码
```
诊断脚本(带 EOS 探针):github.com/niugtd/dots-tts-s...
## 实测规律(294 runs 系统扫描)
经过 4 个阶段共 **294 runs** 测试,发现乱码率由**字数 + 标点 + 词汇结构**三要素综合决定。
### 规律 1:句末标点是首要因素
5 字同句干"我吃饭吧" × 3 语气词 × 3 标点(27 runs):
| 标点 | 吧 | 吗 | 呢 |
|---|---|---|---|
| **问号 ?** | **0/3 ✅** | **0/3 ✅** | **0/3 ✅** |
| 句号 。 | 3/3 ⚠️ | 3/3 ⚠️ | 3/3 ⚠️ |
| 感叹号 ! | 3/3 ⚠️ | 3/3 ⚠️ | 3/3 ⚠️ |
标点变种完整对照("我吃饭吧" 5字):
- ✅ 安全:**无标点**、**逗号中间**"我吃饭,吧。"、**问号**
- ⚠️ 危险:句号、感叹号、省略号、末尾逗号、双标点"?!"、"!?"
#...
Developer Debate & Comments
No active discussions extracted for this entry yet.
Adjacent Repository Pain Points
Other highly discussed features and pain points extracted from rednote-hilab/dots.tts.
Frequently Asked Questions
Market intelligence mapped to dots.tts-mf generating 2-3 seconds of garbled audio after short text synthesis..
What problem does dots.tts-mf generating 2-3 seconds of garbled audio after short text synthesis. solve?
Are engineers actively discussing dots.tts-mf generating 2-3 seconds of garbled audio after short text synthesis.?
What are the foundational technologies related to dots.tts-mf generating 2-3 seconds of garbled audio after short text synthesis.?
Engagement Signals
Cross-Market Term Frequency
Quantifies the cross-market adoption of foundational terms like optimize and ASR by tracking occurrence frequency across active SaaS architectures and enterprise developer debates.
SaaS Metrics