Changelog - Version 3.0.0

This update primarily focuses on code refactoring, resulting in a more streamlined and readable codebase, along with improvements to the prompt and fixes for several stability issues.

🚀 New Features

• Significantly improved transcription quality: now performs ASR on the original audio, then uses Demucs-denoised audio for force alignment, greatly reducing missed sentences.
• Added support for WhisperX 302 Cloud API (recommended for users without local GPUs or who prefer not to deal with complex installation), and preliminary support for 11labs Scribe model (still in development—stability may be lower than WhisperX, use with caution).
• Enhanced segmentation stability through longer chain-of-thought reasoning.
• Improved translation prompt to optimize overly concise translations.
• Added a JSON format support button to the sidebar LLM settings.

🐛 Bug Fixes

• Increased the word deletion threshold from 20 to 30, fixing the issue of incorrectly deleting valid words.
• Fixed errors when processing longer audio/text segments, resolving WhisperX cloud audio segmentation issues.
• Implemented stricter validation of LLM response formats, fixing translation line alignment errors.

🔧 Improvements

• Refactored the project architecture, making the code more streamlined, clear, and easier to maintain.
• Removed automatic FFmpeg GPU acceleration detection, now requiring manual configuration to enhance reliability.
• Implemented pydub for more reliable audio segmentation.

更新日志 - 版本 3.0.0

本次更新主要是对代码进行了重构，现在更加精简易读，并且改进了prompt，修复了一些稳定性问题。

🚀 新功能

• 显著提升了转录质量：现在先对原声进行ASR，再用Demucs降噪后的音频做强制对齐，极大改善了漏句问题。
• 增加对 WhisperX 302 Cloud API 的支持（推荐给本地没有GPU或不想复杂安装的用户），以及初步支持 11labs Scribe 模型（仍在开发中，稳定性似乎不如WhisperX，谨慎使用）。
• 通过更长的链式推理提升分割稳定性。
• 改进翻译prompt，优化过于简洁的翻译结果。
• 侧边栏 LLM 设置中新增 JSON 格式支持按钮。

🐛 问题修复

• 将词语删除阈值从 20 提高到 30，修复误删正常词语的问题。
• 修复处理更长音频/文本片段时的错误，解决 WhisperX cloud 音频分割问题。
• 对 LLM 响应格式进行了更严格的校验，修复了翻译行数对齐错误。

🔧 改进

• 重构了项目架构，使代码更加精简清晰，结构更简单易维护。
• 移除 FFmpeg GPU 加速自动检测，现需手动在配置中设置，以提升可靠性。
• 使用 pydub 实现更可靠的音频分割。