v3.1.1

# VideoLingo v3.1.1 ✨

More flexible subtitle workflows, safer long-video dubbing, and official Fish Audio voice cloning.

### 🚀 What’s new

- **Official Fish Audio dubbing is here.** The new `fish_audio` provider uses Fish Audio’s official API and defaults to the `s2.1-pro` model. It can clone the voice from 15–30 seconds of the source video without storing a voice in your Fish Audio account, or use a fixed Fish Audio voice or custom voice ID.
- **OpenAI-compatible TTS can now use OpenLux.** The OpenAI TTS backend supports the OpenLux speech endpoint, with configurable base URL and model.
- **Bring your own subtitles.** Import an SRT file for translation, with or without a video, and use transcribe-only mode when you only need source subtitles.
- **Review the pipeline before it continues.** Optional checkpoints let you edit terminology before translation and review the translation spreadsheet before subtitle generation.
- **More control over subtitle output.** Set a translation style, customize font/size/color, and independently choose bilingual, translated, or source subtitles for the subtitle and dubbed videos.
- **MAI-Transcribe-2 is available as an optional ASR backend** through Azure Speech or OpenRouter, while local Qwen3-ASR remains the default.
- **Long videos are more robust.** Demucs processes long audio in chunks, Japanese NLP is batched, subtitle splitting has safer fallbacks, and translation/punctuation edge cases no longer stop the pipeline.
- **Dubbing failures are visible and recoverable.** Failed TTS requests stop after retries instead of becoming silent lines. Speaking speed is capped, lines that still cannot fit are listed in `output/log/dub_truncated.json`, and dubbed videos are produced correctly even when burned subtitles are disabled.

### ⚠️ Before upgrading

- The old `fish_tts` provider for 302.ai and `azure_tts` have been removed. Choose Fish Audio, Edge TTS, or another available provider; Edge TTS is now the default.
- Fish Audio requires `fish_audio.api_key` and an account with credit. `msgpack` is a new dependency and is installed with the normal setup flow.
- The Qwen3-ASR model size is now an advanced `config.yaml` option: `whisper.qwen_model`.

### ✅ Release validation

- Automated suite: **663 passed, 19 skipped, 12 subtests passed**.
- Five real English videos (about 7.5–11.3 minutes each) were processed end to end into Simplified Chinese, Japanese, French, Spanish, and Traditional Chinese using OpenLux GPT-5.6 Luna and Fish Audio `s2.1-pro` voice cloning.
- All **1,025** translated lines produced TTS audio, with **0 truncated lines**. All five dubbed videos passed full video/audio decode checks and frame inspection.

---

# VideoLingo v3.1.1 ✨

字幕流程更灵活，长视频配音更稳定，并正式接入 Fish Audio 声音克隆。

### 🚀 这次更新

- **新增 Fish Audio 官方配音。** 新的 `fish_audio` Provider 直接调用 Fish Audio 官方 API，默认模型为 `s2.1-pro`。可以从原视频中取 15–30 秒语音克隆音色，不会在 Fish Audio 账户中保存声音；也可以选择固定音色或自定义 Voice ID。
- **OpenAI 兼容 TTS 现在可以使用 OpenLux。** OpenAI TTS 后端支持 OpenLux 的语音端点，并可自定义 Base URL 和模型。
- **直接使用现有字幕。** 可导入 SRT 进行翻译，无论是否有视频；也可只跑转录阶段，仅生成原文字幕。
- **流程可以暂停检查。** 可选择在翻译前修改术语表，也可在生成字幕前检查并修改翻译表格。
- **字幕输出更可控。** 新增翻译风格，可调整字体、字号和颜色，并可分别为字幕视频和配音视频选择双语、译文或原文字幕。
- **新增可选的 MAI-Transcribe-2 识别后端**，支持 Azure Speech 和 OpenRouter；默认仍然是本地 Qwen3-ASR。
- **长视频处理更稳定。** Demucs 会分块处理长音频，日语 NLP 支持分批，字幕切分有更稳妥的降级方案，翻译和标点边界情况不再让整条流程中断。
- **配音失败不再被隐藏。** TTS 重试失败后会明确中止，不会生成静音冒充成功。语速有上限，仍然放不下的句子会写入 `output/log/dub_truncated.json`；关闭烧录字幕时也能正常生成配音视频。

### ⚠️ 升级前留意

- 302.ai 的旧 `fish_tts` 和 `azure_tts` 已移除。请改用 Fish Audio、Edge TTS 或其他可用 Provider；默认已改为 Edge TTS。
- Fish Audio 需要配置 `fish_audio.api_key`，账户也需要有可用额度。新依赖 `msgpack` 会随正常安装流程自动安装。
- Qwen3-ASR 模型大小现在是 `config.yaml` 中的高级选项：`whisper.qwen_model`。

### ✅ 发布验证

- 自动化测试：**663 passed，19 skipped，12 subtests passed**。
- 使用 5 个真实英文视频（每个约 7.5–11.3 分钟）完整跑通简体中文、日语、法语、西班牙语和繁体中文，翻译使用 OpenLux GPT-5.6 Luna，配音使用 Fish Audio `s2.1-pro` 声音克隆。
- 全部 **1,025** 条译文都成功生成 TTS，**0 条被截断**。5 个配音成片全部通过完整音视频解码和抽帧检查。

Full Changelog / 完整改动：[v3.1.0...v3.1.1](https://github.com/Huanshere/VideoLingo/compare/v3.1.0...v3.1.1)