VoxGo 0.6.0
本页内容由 GitHub Release 生成。
更新内容
- 支持中文(普通话)、英语、日语、韩语、俄语五种语言之间20个有向在线语音翻译方向,按实际源/目标语言直接翻译。
- 日语、韩语、俄语明确标为 Beta 实验性语言;源/目标选择统一显示 Beta 标签,并按识别/译文方向显示非阻断说明。标签不改变语言代码或内部 ASR 策略。
- 软件界面新增日语、韩语、俄语,覆盖设置、浮窗、首次配置、模型下载、更新、错误提示及帮助;界面语言独立于翻译方向保存。
- 中文译其他语言时,识别原文在浮窗和手机镜像中显示简体,保留原始 ASR、送翻译文本和实际译文。
- 日语、韩语内部使用兼容识别策略;俄语在极速、非极低影响模式下使用既有语音完整性策略。正式界面不提供策略开关,标准、精准及自定义调校保留原逻辑。
- 日/韩/俄相关方向需要用户自行配置在线 Provider;本地 OPUS 继续仅支持英中双向,不自动回退到其他语言或收费服务。英语专用 `.en` 模型不会用于非英语识别。
- 语言切换清理旧方向任务和音频,过期结果不会覆盖当前字幕;翻译输出达到长度上限时提示不完整,不展示或提供复制截断正文。
- 切换识别模型后,后台自动检查缓存、下载或加载;无须等待下一段语音,暂停时仍可准备,显示持续的模型准备状态。
- 保留现有英中识别、离线 OPUS、正式 ASR 预设和翻译 Prompt;不集成日/韩/俄离线 PoC,不增加降噪、音源分离、OCR 或官方付费服务。
- Full 匿名诊断按 Fast/Balanced/Accurate/Custom 分类统计 ASR;在线翻译失败按固定原因和服务商分类,对应 Dashboard 展示。原63指标口径保持不变,不上报语音、字幕、Key或请求地址。
- 新增语言使用计数及 Dashboard 源/目标排行、20方向矩阵和每日趋势。按实际任务的语言方向统计成功失败,缓存单列,历史保持未分类。统一跟随原有隐私开关,无需单独启用;新增协议仅在兼容服务端就绪后协商启用。新安装默认勾选,保留已有关闭选择。
What's new
- Supports 20 directed online speech-translation pairs across Mandarin Chinese, English, Japanese, Korean and Russian, translating directly between the selected languages.
- Japanese, Korean and Russian are explicitly experimental Beta languages. Both language selectors show Beta labels and a non-blocking recognition/translation notice, without changing language codes or internal ASR strategies.
- Added Japanese, Korean and Russian interface languages across settings, the overlay, setup, model downloads, updates, errors and help. Interface language is saved independently of translation direction.
- Chinese source transcripts display in Simplified Chinese in the overlay and mobile mirror. Raw ASR, translation input and real translated output are preserved.
- Japanese and Korean internally use compatibility recognition. Russian uses the existing speech-completeness policy in Fast mode except Ultra Low impact. Settings has no strategy switch; Balanced, Accurate and Custom behavior is retained.
- JA/KO/RU directions require a user-configured online provider. Offline OPUS remains English–Chinese only, without silent provider substitution. English-only `.en` models are restricted to English recognition.
- Language changes cancel and isolate old tasks/audio so stale results cannot replace current subtitles. Length-limited translations show an incomplete-output notice without displaying or copying partial output.
- Changing the recognition model starts a background cache check, download or load without waiting for speech, including while paused, with persistent preparation status.
- Preserves existing English–Chinese recognition, offline OPUS, formal ASR presets and translation prompts. No experimental offline languages, denoising, source separation, OCR or official paid API service is added.
- Full anonymous diagnostics classify ASR by Fast/Balanced/Accurate/Custom and API failures by fixed reason/provider, with matching Dashboard views. The original 63 metric meanings are preserved. No audio, captions, keys or request addresses are uploaded.
- Added language usage counts with source/target rankings, a 20-direction matrix and daily trends. Counts use each task's direction, distinguish cache, and leave history unclassified. The original Full privacy switch controls these counts, with no extra opt-in; new setup stays checked by default and existing disabled choices are preserved. Its additive protocol requires a ready, compatible server capability.
下载 / Downloads
- VoxGo-v0.6.0-cuda-runtime.zip 563.2 MiB
- VoxGo-v0.6.0-full-cuda.zip 1248.6 MiB
- VoxGo-v0.6.0-full.zip 685.2 MiB
- VoxGo-v0.6.0-lite.zip 132.3 MiB
附件与 SHA256 校验信息请以 GitHub Release 为准。