⚠️ Important Notice: VoiceHub relies on local open-source AI models to function. Before running the app, you must download and configure the required models. Please refer to the Models Setup Guide for detailed instructions.
VoiceHub is a powerful cross-platform Flutter application providing dedicated iOS and macOS apps. It integrates advanced on-device AI models for Automatic Speech Recognition (ASR), Text-to-Speech (TTS), and Neural Machine Translation (NMT), delivering a complete voice processing pipeline right on your device.
- ⚡ Native C++ Engine: Built with a focus on maximum performance, leveraging C++ native engines for lightning-fast voice processing.
- 🔒 100% Offline & Private: All ASR, TTS, and translation processes run completely locally. No data is sent to the cloud, ensuring absolute privacy.
- 🚀 GPU-Accelerated LLMs: Integrates the
llama.cppinference framework, allowing Large Language Models (LLMs) to run directly on the device's GPU for blazing fast translations. - 📜 Unified History Management: Across all translation modes, it provides comprehensive support for automatic history saving, audio playback, precise timestamped segment comparisons, and one-click sharing.
- 🎙️ Real-time Speech Recognition (ASR): High-accuracy, offline speech-to-text processing.
- 🔊 Text-to-Speech (TTS): Natural and fast offline voice generation.
- 🌐 Neural Machine Translation (NMT): Powered by local LLMs (llama.cpp) and Opus-MT for high-quality, rapid text translation.
- 🔗 Cascade Voice Translation: A robust ASR-to-MT pipeline. Features a clean interface displaying original and translated text blocks in real-time.
- 📄 Offline Audio File Translation: Import audio and video files for fast, fully offline transcription and translation without size limits.
- 🗣️ Dual Dialogue Mode: Face-to-face real-time speech translation, designed for seamless bilingual conversations.
- 📜 History Records & Subtitle Export: Automatically saves transcription and translation history with support for exporting content as subtitle files.
- 🧠 Model Management: Built-in comprehensive manager to easily import, switch, and manage various local ONNX/GGUF AI models.
- 🔄 Simultaneous Translation (SimulST): Real-time cross-language voice translation pipeline (Note: Internal testing version, no models included).
- Cascade vs. SimulST: Unlike the cascade approach, SimulST does not rely on an ASR model using VAD or punctuation for forced sentence segmentation. Instead, it uses an end-to-end translation model that independently determines the optimal timing to output translations, bypassing the traditional ASR dependency.
- ⚡ Cross-Platform: Engineered for macOS, iOS, and other platforms.
- 🔒 Privacy First: Fully on-device data processing, requiring no internet connection.
Note: Replace the placeholders below with actual application screenshots.
| Main Screen | Cascade Translation | Dual Dialogue |
|---|---|---|
![]() |
![]() |
![]() |
| Simultaneous Translation | File Translation | History Records |
|---|---|---|
![]() |
![]() |
![]() |
VoiceHub provides extensive support for various open-source models:
- ASR (Speech Recognition): Supports
transducerarchitecture models released bysherpa-onnx. - TTS (Text-to-Speech): Supports
vitsarchitecture models released bysherpa-onnx. - NMT (Small Translation Models): Requires exporting
opus_mtmodels to ONNX format. - LLM (Large Language Models): Supports open-source LLMs in the
ggufformat. We highly recommend using the hy-mt2-1.8B model for translation tasks.
During continuous audio recording, VoiceHub primarily relies on the silero-vad model for Voice Activity Detection (VAD) and automatic sentence segmentation.
- Dynamic Adjustments: You can freely adjust the VAD parameters (Threshold, Min Silence, Min Speech) directly via the Settings (⚙️) -> VAD Settings panel in the app to optimize for your speaking pace and environment.
- Default Recommended Settings:
- General Mode (Cascade Translation / Dual Dialogue): Speech threshold is set to
0.5, with minimum speech and silence durations both set to0.3seconds (300ms) for snappy, responsive turns. - SimulST Mode (Simultaneous Translation): Minimum silence duration is increased to
0.5seconds (500ms) to allow longer continuous speech blocks before translating.
- General Mode (Cascade Translation / Dual Dialogue): Speech threshold is set to
VoiceHub's robust architecture is built upon several excellent open-source projects:
- Flutter - UI Framework
- sherpa-onnx - ASR & TTS engine
- llama.cpp - Local LLM inference engine
- Opus-MT - Neural Machine Translation framework
- Kaldi-native-fbank - Audio feature extraction
- v1.0.2: SimulST memory optimizations (prompt KV/fbank/encoder release, iOS
n_ctxtuning), Q4_K_M/Q8 quantized SpeechLLM support, and llama translation KV cache cleanup. - v1.0.1: ASR Beam Search default, hybrid VAD + Sherpa endpointing, anti-cutoff trailing silence, and chat UI cleanup.
- v1.0.0: Initial release featuring core ASR, TTS, Opus-MT, and LLaMA native integrations, File Translation, and Dual Dialogue mode for macOS and iOS.
Please refer to the following documentation in the @docs directory for build instructions, model setup, and platform-specific guides:
- 🍎 macOS Build & Distribution Guide / macOS 构建与打包指南
- 📱 iOS Build Guide / iOS 编译指南
- 🧠 Models Setup Guide / 模型下载配置指南
- 🌍 Opus-MT Setup Guide / Opus-MT 部署指南
⚠️ 重要提示:VoiceHub 的所有核心功能均依赖于本地开源 AI 模型。在运行应用之前,您必须先下载并配置相关模型。请务必仔细参考 模型配置指南 进行操作。
VoiceHub 是一款强大的跨平台 Flutter 应用,目前已提供原生的 iOS 和 macOS 应用版本。它集成了先进的端侧 AI 模型,提供自动语音识别 (ASR)、文字转语音 (TTS) 以及神经机器翻译 (NMT) 等完整的语音处理能力。
- ⚡ 原生 C++ 引擎驱动:底层基于 C++ 原生引擎构建,注重极致的性能与运行效率。
- 🔒 100% 离线与隐私保护:所有的语音识别、合成与翻译处理均在设备本地完成,无需网络连接,绝对保障您的数据隐私。
- 🚀 GPU 加速的端侧大模型:通过深度集成
llama.cpp推理框架,VoiceHub 能够将大语言模型 (LLM) 直接运行在设备的 GPU 上,从而大幅加速翻译推理速度。 - 📜 统一的历史记录管理:在所有功能模式下,应用均提供统一的基础架构支持:自动保存历史记录、录音回放、查看带时间戳的精确分段对照以及一键分享。
- 🎙️ 实时语音识别 (ASR):高精度的离线流式与非流式语音转文本能力。
- 🔊 语音合成 (TTS):自然、极速的离线语音生成。
- 🌐 机器翻译 (NMT):基于本地大语言模型 (llama.cpp) 和 Opus-MT,提供高质量的文本翻译。
- 🔗 级联语音翻译 (Cascade Translation):结合 ASR 与 MT 的标准翻译流水线。界面直观展示实时双语对照翻译结果。
- 📄 离线文件翻译:支持导入本地音视频文件,快速进行离线语音提取、转写与翻译。
- 🗣️ 双人对话模式:专为跨语言面对面交流设计的双语实时翻译界面。
- 📜 历史记录与字幕导出:自动保存所有的转写与翻译历史记录,并支持将结果一键导出为字幕文件。
- 🧠 模型管理:内置完善的模型管理器,支持方便地导入、管理和切换本地各类 AI 模型(支持压缩包直通导入)。
- 🔄 同声传译 (SimulST):实时的跨语言语音翻译流水线 (注:内部测试版本,无开源模型提供)。
- 级联 vs 同传的区别:传统的级联翻译强依赖 ASR 模型通过 VAD 或标点符号进行强制断句。而同传 (SimulST) 则采用端到端的翻译模型,能够自主判断最优时机进行增量输出,完全不再依赖 ASR 的前置断句机制。
- ⚡ 跨平台支持:专为 macOS、iOS 等多平台设计与深度优化。
- 🔒 隐私优先:完全在端侧本地处理数据,无需连接网络。
注:请将下方的占位图替换为实际的应用截图。
| 主界面 | 级联翻译 | 双人对话 |
|---|---|---|
![]() |
![]() |
![]() |
| 同声传译 | 文件翻译 | 历史记录 |
|---|---|---|
![]() |
![]() |
![]() |
VoiceHub 为多种开源模型提供深度的本地推理支持:
- ASR (语音识别):支持由
sherpa-onnx发布的transducer架构模型。 - TTS (语音合成):支持由
sherpa-onnx发布的vits架构模型。 - 小模型翻译 (NMT):需要将
opus_mt模型导出为 ONNX 格式后使用。 - LLM (大语言模型):支持使用
gguf格式的开源大模型。针对翻译场景,强烈推荐使用腾讯的 hy-mt2-1.8B 翻译大模型。
在持续收音过程中,应用主要依赖于自带的 silero-vad 模型来进行实时的语音端点检测与自动断句。
- 界面动态调节:您可以直接在主界面的 “设置 (⚙️)” -> “VAD 参数设置” 面板中,通过滑动条动态调节各种 VAD 参数(语音阈值、最小静音判定时长、最少发言时长),以适应您的语速与背景噪音。
- 默认推荐参数:
- 通用模式 (General Mode):用于级联翻译和双人对话,语音阈值设为
0.5,最小静音 (minSilence) 和最小语音 (minSpeech) 时长均为0.3秒 (300ms),以实现极速响应。 - 同传模式 (SimulST Mode):为了收集更长、更完整的上下文,最小静音触发断句的时长被放宽至
0.5秒 (500ms)。
- 通用模式 (General Mode):用于级联翻译和双人对话,语音阈值设为
VoiceHub 的强大功能离不开以下优秀的开源项目:
- Flutter - 跨平台 UI 框架
- sherpa-onnx - ASR 与 TTS 引擎
- llama.cpp - 用于翻译的本地 LLM 推理引擎
- Opus-MT - 神经机器翻译模型
- Kaldi-native-fbank - 音频特征提取库
- v1.0.2:同传内存优化(段落结束立刻释放 fbank/encoder/KV,iOS
n_ctx下调防 OOM)、支持 Q4_K_M/Q8 量化 SpeechLLM,以及翻译后立刻释放 llama KV cache。 - v1.0.1:ASR 默认 Beam Search、VAD + Sherpa 双轨断句、尾端防吞字,以及聊天页 UI 精简。
- v1.0.0:首发版本,包含核心 ASR、TTS、Opus-MT 以及 LLaMA 的原生集成,加入了离线文件翻译、双人对话等核心业务场景,完美支持 macOS 及 iOS 平台。
关于环境搭建、编译打包和模型配置,请参考项目 @docs 目录下的详细教程:
- 🍎 macOS Build & Distribution Guide / macOS 构建与打包指南
- 📱 iOS Build Guide / iOS 编译指南
- 🧠 Models Setup Guide / 模型下载配置指南
- 🌍 Opus-MT Setup Guide / Opus-MT 部署指南





