Spokked

Speak your language. Be heard in theirs.
Real-time voice translation for Google Meet — in your own cloned voice.

▶ Watch the 60-second demo

---

Spokked sits between your microphone and Google Meet. You speak Turkish; about a second after you finish a sentence, the other side hears it in English in your own voice. When they answer in English, you hear their original voice and read live Turkish subtitles on top of the call. When the meeting ends, a summary and a full bilingual transcript are saved for you.

It runs locally on your Mac as a small Python service plus a Chrome extension. Your own (untranslated) voice never reaches the call.

Turkish ⇄ English is the default pair. Both directions are configurable with ISO language codes.

Features

Outgoing — you speak, they hear your voice in their language

• Streaming speech recognition (ElevenLabs Scribe v2 Realtime) → translation (Claude) → speech in your cloned voice (ElevenLabs Eleven v4 Turbo).
• ~1–1.5 s from the end of your sentence to the first translated audio.
• Tone follows you: the translator tags each sentence (excited, calm, curious, apologetic…) and the voice delivers it that way.
• Pace matching: if the translation is longer than what you said, it is sped up slightly (pitch-preserving, max 1.2×).
• Language-aware early end-of-sentence detection and speculative translation that starts before the recognizer commits.
• Safety nets: the translator never answers or comments instead of translating, output is checked to be in the right language, echo of your own translated voice is filtered, silence "hallucinations" are dropped.
• Muting your mic in Meet pauses Spokked too, and drops anything still in flight.

Incoming — they speak, you read live subtitles

• Their original voice plays normally; a draggable subtitle box shows English on top and Turkish below.
• Finished sentences are locked in white while the speaker is still talking; only the unfinished tail is shown as a grey draft. Previous lines stay visible.
• Subtitle-specific translation (concise, readable, fillers removed).
• Optional automatic language detection for the other side.

Meeting helpers

• 📋 Catch me up (Alt+S): Turkish summary of the last 5 minutes, any question you were asked, and a suggested short answer.
• Automatic meeting notes: when you press Stop or leave the Meet call, a summary (topics, decisions, action items, open questions) and the full bilingual transcript are saved to a folder (default ~/Desktop/Meet Özetleri).
• Personal glossary (sozluk.txt): names and term pairs applied to recognition and both translation directions.
• Optional camera beauty filter (brightness, smoothing, warmth) applied inside Chrome.
• Local dashboard at http://127.0.0.1:8765/ with latency, counts and quota.

How it works

 OUTGOING
 Your mic ─► Scribe v2 Realtime (tr) ─► Claude Haiku (tr→en) ─► ElevenLabs v4 Turbo (your voice) ─► BlackHole 2ch ─► Meet "Microphone"

 INCOMING
 Meet "Speaker" ─► BlackHole 16ch ─► Scribe v2 Realtime (en) ─► Claude Haiku (en→tr) ─► subtitle overlay in Meet
                                   └─► original audio ─► your headphones

• BlackHole provides two virtual audio cables: Meet's microphone only ever receives the translated voice, and Meet's speaker output is captured for subtitles before being passed on to your headphones.
• The Chrome extension draws the subtitle overlay, adds Start/Stop and mute controls, syncs with Meet's mute button and detects when you leave a call.
• The local service (translator.py) runs as a macOS LaunchAgent in standby mode; it only opens the microphone and spends API credits after you press Start.

Requirements

• macOS (Apple Silicon or Intel) and Google Chrome
• Homebrew
• An ElevenLabs account with a cloned voice. A paid plan is recommended (Creator is enough); real-time translation uses roughly 800–1,000 characters per minute of your speech. Subtitles cost no ElevenLabs characters.
• An Anthropic API key (translation costs a few cents per meeting)

Installation

git clone https://github.com/Bihruze/spokked.git
cd spokked
./setup.sh

setup.sh installs the BlackHole virtual audio devices, creates a Python virtual environment, installs the dependencies and creates your .env from .env.example.

If the BlackHole devices don't show up yet, restart Core Audio (or your Mac):

sudo killall coreaudiod

1. Clone your voice

1. Go to ElevenLabs → Voices → Add a new voice → Instant Voice Clone.
2. Record 2–3 minutes of natural speech in a quiet room, in your own language. (Short samples make the clone sound generic; reading English text in a strong accent carries that accent over.)
3. Enable "Remove background noise", save, and copy the voice ID.

2. Add your keys

Edit .env (open -e .env) and fill in the first three lines:

ELEVENLABS_API_KEY=...
ANTHROPIC_API_KEY=...
ELEVENLABS_VOICE_ID=...

.env is ignored by git and created with owner-only permissions. Then verify the whole setup:

./run.sh check

Every line should show ✓ (keys, audio devices, quota, and the latency of each cloud service).

3. Load the Chrome extension

1. Open chrome://extensions and enable Developer mode.
2. Click Load unpacked and select the extension folder inside this repo.
3. Pin Spokked from the puzzle-piece menu.

4. Run it in the background (recommended)

./autostart.sh            # install as a LaunchAgent (starts in standby at login)
./autostart.sh remove     # uninstall

From now on you never need the terminal: the extension's ▶ START / ■ STOP button controls everything. Logs go to logs/service.log.

Prefer running it by hand? Use ./run.sh (active immediately, Ctrl+C to quit).

Using it in a meeting

1. In Google Meet open ⋮ → Settings → Audio:
   • Microphone: BlackHole 2ch
   • Speaker: BlackHole 16ch
   • Noise cancellation: off (it can clip synthetic speech)
2. Click the Spokked extension → ▶ START.
3. Speak normally and pause briefly at the end of sentences.
4. Leave the call or press ■ STOP — your notes are saved automatically.

Use headphones for the best result: on built-in speakers Spokked mutes your microphone while a translation is playing to avoid echo; with headphones you can talk over the other side freely. Switching to AirPods mid-call is detected within ~2 seconds.

Keyboard shortcuts (in the Meet tab)

Shortcut	Action
Alt+M	Mute / unmute your translated voice
Alt+S	Catch me up (summary + question + suggested answer)
Alt+E	Show / hide the English line in subtitles
Alt + / Alt -	Subtitle size
Double-click the box header	Reset subtitle position

Commands

./run.sh                 # run in the foreground
./run.sh check           # readiness check: keys, devices, quota, latency
./run.sh devices         # list audio devices
./run.sh voices          # list your ElevenLabs voices and IDs
./run.sh translate "..." # test translation only
./run.sh say "..."       # speak a test sentence in your cloned voice into BlackHole 2ch
./run.sh summary         # Turkish notes for the most recent transcript

Configuration

All settings live in .env (see .env.example for the full, commented list). The most useful ones:

Setting	Default	What it does
SOURCE_LANG / TARGET_LANG	tr / en	Your language and the other side's language
INCOMING_STT_LANG	(target)	auto detects the other side's language; if they speak yours, subtitles show it untranslated
INCOMING_MODE	subtitles	voice reads the other side to you in a translated voice instead
TTS_MODEL	eleven_v4_turbo	Voice model; v4 supports delivery tags like [excited]
VOICE_ACCENT	American accent	Accent tag sent with every sentence
VAD_SILENCE_SECS	0.45	Silence that ends a sentence (lower = faster, more splits)
PACE_MATCH / PACE_MAX	1 / 1.2	Match the translation to your speaking pace
PREBUFFER_MS	600	Upper bound of the adaptive jitter buffer
KEYTERMS		Comma-separated names/terms to protect
MEETING_CONTEXT		One line about the meeting; helps resolve ambiguity
MEET_NOTES_DIR	~/Desktop/Meet Özetleri	Where meeting notes are saved
TRANSLATE_MODEL	claude-haiku-5-5	Fastest; any Claude model works

After changing .env or sozluk.txt, restart the service: ./autostart.sh (or restart ./run.sh).

Latency

Measured end to end on a home connection (speech end → first translated audio in Meet):

Stage	Time
End-of-sentence detection	~0.25–0.45 s
Translation, first sentence (Claude Haiku, partly hidden by speculative translation)	~0.6–1.0 s
First audio (Eleven v4 Turbo, HTTP streaming)	~0.2 s
Total	~1.0–1.5 s

Incoming subtitles: the recognizer delivers its first partial ~1–2 s into a sentence; a Turkish draft follows ~1 s later, and the final line ~1.5–2 s after the speaker stops. Turkish is verb-final, so drafts necessarily change at the end of a sentence — that tail is shown in grey.

Privacy & security

• The control server listens on 127.0.0.1 only, checks the Host header (DNS-rebinding protection), only answers the extension, the Meet page and its own dashboard, and requires a custom header plus POST for every state change.
• .env, transcripts and logs are created with owner-only permissions and are git-ignored.
• In standby the microphone is not opened and nothing is sent anywhere.
• Audio and text are processed by ElevenLabs and Anthropic according to their API terms. Meeting notes are written to MEET_NOTES_DIR; if that folder is inside iCloud Desktop & Documents, they will sync to iCloud.
• Let the people you talk to know that you use AI translation and a cloned voice.

Troubleshooting

Problem	Fix
Extension says the service isn't running	./autostart.sh or ./run.sh
The other side can't hear you	Meet microphone must be BlackHole 2ch, and Spokked must be started
No subtitles / you can't hear them	Meet speaker must be BlackHole 16ch (not 2ch)
BlackHole not listed	sudo killall coreaudiod or restart your Mac
Microphone not detected	System Settings → Privacy & Security → Microphone → allow Terminal/Python
Your clone sounds generic or deep	Re-clone with 2–3 minutes of clean speech in your own language
Anything else	./run.sh check, then logs/service.log

Project layout

translator.py        the local service: audio I/O, STT, translation, TTS, subtitles, control server, notes
extension/           Chrome extension (popup, Meet overlay, camera filter)
setup.sh             one-time installer
autostart.sh         install/remove the macOS LaunchAgent
run.sh               run any command inside the virtual environment
sozluk.txt           personal glossary
.env.example         every setting, documented

The user-facing strings in the overlay and logs are in Turkish; the code is a single, heavily commented Python file.

License

MIT