New Paraformer streaming model — words appear on screen as you speak, no more waiting until you stop recording.
新增 Paraformer 流式模型——边说边出字，不再需要录完才开始转写。
- 🎤 Real-time streaming transcription — Text appears live while you speak using sherpa-onnx OnlineRecognizer
- 📦 Paraformer Streaming model — Chinese-English bilingual, download in settings (~237MB INT8)
- 🔄 Seamless mode switching — Streaming / non-streaming toggle in settings, all existing models unaffected
- 🎤 实时流式转写 — 使用 sherpa-onnx OnlineRecognizer，录音同时文字实时显示
- 📦 Paraformer 流式模型 — 中英双语，设置窗口一键下载（~237MB INT8）
- 🔄 模式无缝切换 — 设置中开关流式/非流式，现有模型完全不受影响
Switched polishing model from gpt-4o-mini to gpt-5.4-nano — fixing critical issues where the old model would answer questions, refuse requests, or calculate results instead of just polishing text.
润色模型从 gpt-4o-mini 切换到 gpt-5.4-nano——修复了旧模型回答问题、拒绝请求、计算结果等严重问题，现在只做文字修正。
- ⚡ gpt-5.4-nano — Faster, cheaper, natively follows instructions without hacks
- 🚫 No more answering questions — Input "What's the weather today?" now correctly outputs "What's the weather today?" instead of making up an answer
- 🏷️ XML tag wrapping — User input wrapped in
<speech_transcript>tags to clearly separate data from dialogue - 🔧 Prompt rewrite — Role-locked as "text correction tool, not a conversational AI", with strict prohibition rules
- ⚡ gpt-5.4-nano — 更快更便宜，原生遵守指令无需额外 hack
- 🚫 不再回答问题 — 输入"今天天气怎么样"正确输出"今天天气怎么样？"而不是编造天气回答
- 🏷️ XML 标签包裹 — 用户输入用
<speech_transcript>标签包裹，明确区分数据和对话 - 🔧 Prompt 重写 — 角色锁定为「纯文本修正工具，非对话AI」，严格禁止回答/回应/解读
- 🌐 Translation fix — Non-English target languages (Japanese, Traditional Chinese, etc.) now translate correctly
- 🌐 Original language preserved — In "translate + show original" mode, English input stays English (was incorrectly converted to Chinese)
- 🌐 "Only translate" mode fix — Removed incorrect rule "don't translate if already target language" that caused translation failures
- ⚙️ Settings prompt persistence fix — Default prompt no longer gets "baked" into config.yaml, code updates now take effect immediately
- 🔧 Config reload order fix — Language setting now updates before transcriber creation
- 🛡️ Transcriber readiness check — Recording blocked when transcriber is not initialized
- 🧹 Temp file cleanup — Audio files properly cleaned up in concurrent early-exit paths
- ⚙️ Streaming field residual fix — Switching to Azure backend now explicitly resets streaming flag
- 🌐 翻译修复 — 非英文目标语言（日语、繁体中文等）现在能正确翻译
- 🌐 原文语言保持 — 「翻译+显示原文」模式下，英文输入保持英文原文（之前会被转成中文）
- 🌐 「仅翻译」模式修复 — 去掉了「原文已是目标语言则不翻译」的误判规则
- ⚙️ 设置提示词持久化修复 — 默认 prompt 不再被固化到 config.yaml，代码更新后立即生效
- 🔧 配置重载顺序修复 — 语言设置现在在创建转写器之前更新
- 🛡️ 转写器就绪检查 — 转写器未初始化时阻止录音
- 🧹 临时文件清理 — 并发早退路径中正确清理音频文件
- ⚙️ 流式字段残留修复 — 切换到 Azure 后端时显式重置 streaming 标志
| File | Description |
|---|---|
VoxAIInput-Setup-0.0.6.exe |
Installer (recommended) — Double-click to install |
VoxAIInput-0.0.6-win64.zip |
Portable — Extract and run |
config.example.yaml |
Config template |
- Windows 10/11 (x86_64)
- Microphone / 麦克风
- Transcription engine (choose one / 二选一):
- 🖥️ Local offline — No extra setup, download model in settings / 本地离线 — 无需额外配置，设置中下载模型即可
- ☁️ Azure cloud — Azure AI Foundry with
gpt-4o-mini-transcribe+gpt-5.4-nano