Lingo

A window for calls in English, made for Portuguese speakers:

• shows what the others say while they speak (each word appears ~1 s after it is said), with the Portuguese translation ~2 s behind the speech, and explains expressions such as circle back or touch base;
• when someone asks a question (or says your name), suggests 3 replies in English: direct, complete and diplomatic, each with its meaning in Portuguese;
• in the "Como digo…?" ("how do I say…?") field you write in Portuguese and get 3 ways to say it in English.

For a call in Portuguese, the EN→PT button in the top bar (or Ctrl+T) switches to PT mode: it only transcribes, without translation, and the suggested replies come in Portuguese.

The interface is in Brazilian Portuguese for now.

Made for Linux with PipeWire. The window is native GTK4 and opens like any other app (tiled, on Hyprland). It adapts to its width: below 460 px the top bar gets compact, and from 900 px on the suggestions, Claude's answers and the text field move to a column on the right of the conversation. On Hyprland it also has global shortcuts.

How it works

system output ─┬─pw-record──▶ gpt-live-transcribe (English, live) ──▶ window, suggestions
               └────────────▶ gpt-realtime-translate (Portuguese)  ──▶ window
microphone    ──pw-record──▶ gpt-4o-transcribe                    ──▶ context for the suggestions

Each side of the conversation is a separate stream, so there is no need to guess who spoke. The others' English arrives word by word, and the Portuguese comes from a live translator that listens to the same audio, ~2 s behind the speech, without waiting for the sentence to end. Each utterance gets its own piece of the translation below it: the translation only moves on to the next utterance when it pauses itself. The 💡 idiom notes come per sentence, from a separate call to gpt-5.4-mini.

Lingo itself closes each utterance after 700 ms without voice. Consecutive utterances from the same person are grouped in one block, and the utterance in progress has a blue bar on its left. Only the stretches with voice are sent (the translator also gets 3 s of silence after each utterance: with less, it swallows the ending).

With translate = "gpt-5.4-mini" in [models], translation goes back to sentence by sentence: the Portuguese of each sentence only appears ~2 s after it ends. It costs half, but on a long sentence the wait reaches 8–10 s. The live translator does not accept the keyword list, so jargon is sometimes translated.

Measured with TTS-generated speech, from the end of each word until it appears in the window:

	median	long sentence (p90)	worst
English	1.15 s	1.45 s	—
Portuguese, live	2.2 s	3.8 s	4.3 s
Portuguese, sentence by sentence (gpt-5.4-mini)	4.4 s	8.6 s	10.2 s

At the end of a question, both ways of translating finish ~2 s after the last word.

The automatic suggestion starts as soon as the "?" shows up, without waiting for the silence: the first option is complete ~2.3 s after the end of the question, and all three in ~3 s. If the person keeps talking after the question, it is redone when they stop; meanwhile, the previous options stay dimmed in the panel.

Accuracy

What goes wrong most in a developers' call is English jargon in the middle of Portuguese ("o pod" becomes "pode"). Three things fix it:

• keywords in [languages]: names, acronyms and project terms, passed to both models;
• in PT mode, the others' speech is tagged as Portuguese and English;
• transcribe_delay = "high": the model waits for a bit more context before writing.

On 8 developer utterances at ~230 words/min, with noise and Opus at 16 kbps, the word error rate dropped from 3.5% to 1.1% for the others and from 4.4% to 0.9% for the microphone. Every technical term came out right, including the ones that were not on the list. The cost is ~0.3 s more per word (low is faster, with three times the errors).

Asking Claude

For factual questions about your projects ("which version of the payments lib does checkout use?"), Lingo calls Claude Code (claude -p), which has its own memory and reads the code. The answer appears in the CLAUDE panel, which shows what it is looking at while it searches:

• SUPER+ALT+A, Ctrl+K or right-clicking an utterance: asks about the others' latest utterance (or about that one);
• in the bottom field, ?question + Enter asks what you typed (in PT mode, any text).

If the answer is in its memory, it arrives in ~3–4 s; with one code search, in ~6–12 s; and when it needs to look through several files, in ~30 s. It runs from ~ (Claude Code's memory is per folder; change it in [ask] cwd), only with read-only tools (Read, Grep, Glob) and without saving the session to its history. The call content goes to Claude Code, not to OpenAI.

Install

You need PipeWire (pw-record), GTK 4.14 or newer, GLib 2.80 or newer, Rust 1.92 or newer (rustup) and an OpenAI API key. To build:

sudo apt install build-essential pkg-config libgtk-4-dev libssl-dev   # Ubuntu 24.04 / Debian
sudo pacman -S --needed base-devel gtk4 openssl                        # Arch / Omarchy

Then:

tools/install.sh   # builds, installs to ~/.local/bin/lingo, creates the .desktop entry and the Hyprland rules
mkdir -p ~/.config/lingo
(umask 077; read -rsp "OpenAI key: " k && echo "$k" > ~/.config/lingo/openai_key)

The key can also come from the OPENAI_API_KEY variable. If a SUPER+ALT shortcut already does something else in your Hyprland (on Omarchy, for example), install.sh comments it out in ~/.config/hypr/lingo.conf and tells you.

Each release also has a tarball with the binary already built, on Ubuntu 24.04, so it runs there and on newer systems. Inside it, tools/install.sh installs that binary instead of building one, so Rust is not needed. To check that it was built by this repository's workflow, from the tagged commit: gh attestation verify lingo-*.tar.gz --repo heroesofcode/lingo.

Use

SUPER+ALT+L	focus Lingo, or open it
SUPER+ALT+R	suggest replies to the latest utterance
SUPER+ALT+P	pause / resume (pausing releases the microphone)
SUPER+ALT+A or Ctrl+K	ask Claude about the latest utterance
click an utterance	suggest replies to that utterance
right-click an utterance	ask Claude about that utterance
?question + Enter in the field	ask Claude what you typed
click an option or Ctrl+1..3	copy the reply
Ctrl+T or the EN→PT / PT button	switch between translating and a call in Portuguese
Ctrl+R Ctrl+P Ctrl+M Ctrl+L Esc	suggest, pause, microphone, clear, close panel

From the command line: lingo --help.

Configuration lives in ~/.config/lingo/config.toml (template in data/config.example.toml). The log is in ~/.local/state/lingo/lingo.log.

Costs

The others' speech uses gpt-live-transcribe ($0.017 per minute of voice sent; silence is not sent) and gpt-realtime-translate ($0.034 per minute). Your microphone uses gpt-4o-transcribe, $0.006 per minute (mic = "gpt-4o-mini-transcribe" costs half and makes twice the errors). In a 1-hour call where the others talk half the time, that is about $1.90; at most $3.40 if they never stop. With sentence-by-sentence translation (translate = "gpt-5.4-mini"), about $0.90. Translation and suggestions with gpt-5.4-mini are a few hundred tokens each. The counter in the top bar shows the estimated cost of transcription and translation; hover over it to see the minutes sent.

To spend ~6x less on the others' speech, set transcribe = "gpt-4o-mini-transcribe" in [models]. The text then only appears when the person stops talking.

Good to know

• Bluetooth: while the microphone is open, the headset stays in the call profile (worse sound). Pausing or closing Lingo releases the microphone.
• Call volume: only what goes above speech_level (0.003) counts as voice. If the "eles" (them) meter moves but no text appears, lower this value in [audio]; if noise turns into text, raise it.
• Screen sharing: the window shows up if you share the whole screen. Share only the call window.
• Privacy: the call audio goes to OpenAI. Check whether that is allowed by your company and your client before using it in meetings with other people.
• Fonts: Lingo uses its own fontconfig cache (~/.cache/lingo), because Edge writes a cache in a newer format to ~/.cache/fontconfig that breaks apps with an older fontconfig.
• European Portuguese: the live translator only accepts pt, with no variant, and sometimes answers "Implementámos" instead of "Implementamos". There is no way to ask it for pt-BR.

Code

Rust, with GTK4 (gtk4-rs) on the main thread and the engine on tokio, in its own threads:

src/engine.rs	the engine: receives everything on a single queue (window commands, OpenAI events) and sends events to the window
src/realtime.rs	Realtime API WebSockets (live transcription and translation), with reconnection
src/audio.rs	capture with pw-record and the voice/silence gates
src/llm.rs, src/prompts.rs, src/text.rs	streaming chat calls, prompts and parsing of the answers
src/ask.rs	questions to Claude Code
src/ui/	the window and the command line

The first version, in Python, is the repository's first commit (52a9e3e).

Tests

cargo test

mise run check runs rustfmt, clippy and the tests the way CI does; mise tasks lists the rest.

To open a second instance without touching the one that is open: LINGO_APP_ID=dev.pedro.LingoTest LINGO_CONFIG=/path/to/config.toml target/release/lingo.

License

MIT. See LICENSE.