Talk-type 面向 macOS 的中文、英文和中英混合语音输入。可在设置中切换中文或 English 作为主要识别语言,中英混合识别默认开启。它将实时语音识别、文本整理和跨应用输入整合在一起,适合写作、聊天、笔记和日常办公。
- 实时识别与纠错:录音时实时显示转写内容;开启 AI 整理后,文本会在录音期间同步更新,结束后可看到识别、整理和写入状态。
- 中文、英文与中英混合输入:可在设置中切换中文或英文作为主要识别语言;中英混合识别默认开启,以另一种语言辅助复核。
- 本地 Whisper 回退:Apple Speech 无结果时,可使用本机 Whisper Small 继续识别。
- 多 AI 整理:支持 DeepSeek、Anthropic(Claude)、OpenAI、xAI(Grok)、Qwen(通义千问)和 Kimi(月之暗面)。
- 上下文整理:结合光标附近文本处理标点、断句、空格和句中插入。
- 跨应用输入:通过 macOS 辅助功能写入当前输入框;不支持时保留到剪贴板。
- 全局快捷键与历史记录:支持
fn和自定义快捷键,最近输入保存在本机。
首次使用时,在「设置 → 权限」中开启麦克风、语音识别和辅助功能权限。在「设置 → 识别与整理」中选择中文或 English 作为输入语言;中英混合识别默认开启。AI 整理为可选功能,可在同一页面选择服务、填写 API Key 和模型 ID;不同服务的配置独立保存。
- 原生 macOS:使用 SwiftUI 构建界面,结合 AppKit 处理菜单栏、窗口、全局快捷键和跨应用输入。
- 实时识别链路:
AVAudioEngine同时采集麦克风、保存临时音频并输出音量数据;Apple Speech 提供录音过程中的实时结果。 - 本地回退识别:Apple Speech 没有得到有效结果时,调用本机
whisper-cli和 Whisper Small 模型继续识别,音频文件只作为本次处理的临时文件。 - 双语识别:主语言可以选择中文或 English;混合模式默认开启,录音时显示主语言结果,结束后使用另一种语言复核,中英文内容交给整理流程保守合并。
- 多服务适配:统一 AI 文本整理接口,分别适配 Chat Completions 和 Anthropic Messages API;每个服务独立保存 API Key 和模型 ID。
- 跨应用写入:通过 macOS Accessibility 读取有限的光标上下文并写入当前输入框;无法自动写入时保留到剪贴板。API Key 使用 macOS 钥匙串保存。
当前暂未提供 DMG 安装包,可从源码运行:
git clone https://github.com/cyx2333hhh/talk-type.git
cd talk-type
open Murmur.xcodeproj需要 macOS 14 或更高版本和 Xcode。使用本地 Whisper 回退时,请自行安装 whisper-cli 和 Whisper Small 模型。
Talk-type 使用 Apple Speech、AVFoundation 和 whisper.cpp。项目开发过程中使用了 OpenAI Codex 协作完成方案讨论、实现与验证。
Talk-type is a macOS voice input tool for Chinese, English, and mixed Chinese-English input. Choose Chinese or English as the primary recognition language in Settings; mixed Chinese-English recognition is enabled by default. It brings live speech recognition, text cleanup, and cross-app input together for writing, messaging, note-taking, and everyday work.
- Live recognition and cleanup: Shows the live transcript while recording. With AI cleanup enabled, the text updates during recording and the app shows recognition, cleanup, and insertion status after recording ends.
- Chinese, English, and mixed input: Choose Chinese or English as the primary recognition language. Mixed Chinese-English recognition is on by default and uses the other language as a second pass.
- Local Whisper fallback: Uses local Whisper Small when Apple Speech returns no result.
- Multiple AI providers: Supports DeepSeek, Anthropic (Claude), OpenAI, xAI (Grok), Qwen, and Kimi (Moonshot).
- Context-aware cleanup: Uses nearby cursor text to handle punctuation, sentence breaks, spacing, and inline insertion.
- Cross-app input: Writes into the focused field through macOS Accessibility, or keeps the text on the clipboard when unavailable.
- Global shortcuts and local history: Supports
fnand custom shortcuts; recent input stays on the Mac.
On first use, grant Microphone, Speech Recognition, and Accessibility permissions in Settings → Permissions. In Settings → Recognition & Cleanup, choose Chinese or English as the input language; mixed Chinese-English recognition is enabled by default. AI cleanup is optional. On the same page, choose a provider and enter its API key and model ID; settings are stored separately for each provider.
- Native macOS: The interface is built with SwiftUI, with AppKit handling the menu bar, windows, global shortcuts, and cross-application input.
- Live recognition pipeline:
AVAudioEnginecaptures the microphone, writes temporary audio, and emits level data; Apple Speech supplies partial results while recording. - Local fallback: If Apple Speech returns no usable result, the app calls local
whisper-cliwith the Whisper Small model. The audio file is temporary and used only for the current dictation. - Bilingual recognition: Choose Chinese or English as the primary language. Mixed recognition is enabled by default: the primary result is shown live, then the other language is used as a second pass before conservative cleanup.
- Provider adapters: A unified cleanup client supports both Chat Completions and the Anthropic Messages API, while keeping API keys and model IDs separate for each provider.
- Cross-app insertion: macOS Accessibility supplies limited cursor context and writes into the focused field; the clipboard is used when automatic insertion is unavailable. API keys are stored in the macOS Keychain.
There is no DMG release yet. Run from source instead:
git clone https://github.com/cyx2333hhh/talk-type.git
cd talk-type
open Murmur.xcodeprojmacOS 14 or later and Xcode are required. For local fallback transcription, install whisper-cli and the Whisper Small model.
Talk-type uses Apple Speech, AVFoundation, and whisper.cpp. OpenAI Codex was used during development for planning, implementation, and verification.



