TalkType — talk instead of typing

中文说明

Talk instead of typing. Press a hotkey anywhere on your Mac, speak, press again (or pause) — the words land where your cursor is. Slack, Notes, email, terminal… anywhere you can type, you can talk.

Built for people who mix Chinese and English in one sentence — no language switching.

press hotkey → talk → press again / pause → text at cursor

Free app (MIT). You bring your own speech API key; usage is billed by that provider. No TalkType account, no subscription, no telemetry. Not an offline / on-device engine — audio goes to the provider you pick.

---

Install (about 5 minutes)

Needs: macOS 13+ (Intel or Apple silicon), a network connection, and a Doubao (豆包) or xAI API key.

1. Download the latest macOS build:
   ⬇ latest zip
   Unzip → drag TalkType.app into Applications. Ignore 2.x zips — those were a different, on-device product.

2. Open it the first time. This build is not signed with a paid Apple Developer certificate, so macOS may block it:
   right-click TalkType → Open → Open.
   If that still fails: System Settings → Privacy & Security → Open Anyway.

3. Allow two permissions when asked:

   • Microphone
   • Accessibility (so TalkType can paste for you)

4. Paste an API key for the provider you want (dialog opens if missing):

   Provider	Menu	Where to get a key	Notes
   Doubao (豆包 / Volcengine) (default)	API Key…	Doubao Voice console → API Key 管理. Not IAM「API访问密钥」. Enable 流式语音识别 2.0 and 录音文件识别 2.0 (Streaming / Audio File Recognition 2.0).	Streams while you speak; same-provider file retry if the stream fails.
   Grok / xAI (optional)	Speech Provider → Grok, then API Key…	console.x.ai → API Keys	REST after you stop. Chinese works in practice but is not on xAI’s official formatting language list (best-effort).

Press ⌘⇧Space and start talking. Change the hotkey anytime from the menu bar.

---

What it does (and does not)

• One recognition per dictation — no second “polish” model, no local multi‑GB engine.
• One cloud provider at a time — pick Doubao (豆包) or Grok in Speech Provider. TalkType never fails over between them behind your back.
• Your own model — not a menu setting. A file-STT client is a small code change; see Building from source.
• Vocabulary — names the model mishears; sent as Doubao hot words or Grok keyterms. Nothing rewrites text after STT.
• Clipboard — every transcript is also copied, so ⌘V always works as a backup.

Privacy

What	Leaves your Mac?
Audio for this dictation + vocabulary terms	Yes — to the provider you selected (Doubao / Volcengine or xAI), under that provider’s policy
Everything else	No — no TalkType account, no telemetry, no subscription identity

No offline mode: no network → a clear error.

---

Manual

Item	Where
Hotkey	Menu bar → Change Hotkey… (default ⌘⇧Space)
Speech provider	Menu bar → Speech Provider
API key (active provider)	Menu bar → API Key…
Microphone	Menu bar → Microphone (or Automatic)
Vocabulary	Menu bar → Vocabulary
Launch at login	Menu bar → Launch at Login

Bluetooth tip: Recording through a Bluetooth headset often switches the link into headset mode; system playback may drop to 24 kHz mono for a while. That is macOS behavior, not a TalkType bug.

---

Troubleshooting

Symptom	What to try
macOS won’t open the app	Right-click → Open → Open. Still blocked: System Settings → Privacy & Security → Open Anyway
Stopped pasting after an update	Menu bar / prompt → Fix This, re-enable TalkType under Privacy → Accessibility. Builds that share the same signing cert keep the grant.
“还没填 API Key” (“API Key not set yet”)	Menu bar → API Key… for the current provider
Doubao: wrong key / “requested grant not found”	Use the Voice console project key (not IAM); enable Streaming + Audio File Recognition 2.0
Grok: wrong key	Paste a full xAI API key from console.x.ai (no spaces / masked ***)
AirPods won’t start recording	Retry once, or switch Microphone to the MacBook mic; Bluetooth handoff is flaky on macOS
CN/EN stuck together with no space	Raw model output by design — no local rewrite

---

Cost

• TalkType: free.
• Speech API: you pay the provider you pick. List prices move; check the console. As of mid-2026, a few minutes a day is pocket change:
  • Doubao Streaming Speech Recognition 2.0 (流式语音识别 2.0) ≈ ¥1 / audio hour
  • Grok REST file STT = $0.10 / audio hour (streaming, if we add it later, is $0.20)

---

How it compares

	TalkType	Apple Dictation	Wispr Flow / Superwhisper
Price	Free app; STT billed by your provider	Free	Subscription
Offline	No	Partly	Usually cloud
Voice leaves your Mac	Yes (to the provider you pick)	Sometimes	Usually
Mixed Chinese + English	Yes	Poorly	Varies
Open source	Yes (MIT)	No	No

---

Building from source

git clone https://github.com/simonsysun/talktype.git
cd talktype

./scripts/make-signing-cert.sh   # optional, once: keep Accessibility across rebuilds
./scripts/build.sh install       # Release build → /Applications/TalkType.app
# or: ./scripts/build.sh release # also writes dist/TalkType-<version>.zip

• Logic tests (no Xcode GUI): swift test
• Packaging helpers: scripts/build.sh, scripts/make-signing-cert.sh
• macOS only. There is no iOS target in this repository.
• To add a third speech API: copy GrokSTTClient, add one STTProvider case, one Keychain item, and the menu title. Streaming needs its own session; do not invent a plugin framework for one extra client.

---

Help shape TalkType

TalkType gets better through real daily use. When something feels off, a word is always misheard, or you have an idea — even a half-formed one — please put it in a GitHub Issue.

That is the main place we look: bugs, rough ideas, and tiny usability notes all count. You do not need a complete design. We can work out what to build together.

You can also browse existing discussions.

---

Under the hood (short)

• Doubao (豆包): 16 kHz PCM over WebSocket Streaming Speech Recognition 2.0 (bigmodel_nostream); on failure, same audio to Audio File Recognition 2.0 Flash (录音文件识别 2.0 极速版). Hot words in request.corpus.context as a JSON string.
• Grok: WAV upload to POST https://api.x.ai/v1/stt with optional repeated keyterm fields (no format/language=en for mixed CN/EN).
• Config / vocabulary live under ~/.talktype (override with TALKTYPE_STATE_DIR).

---

Licence

MIT — do what you like with it.