Sokutsu: free interpreter app translates conversations fully offline on your Mac

A Japanese developer has released Sokutsu, an open-source app that translates conversations on your Mac in real time, as subtitles and as speech. Speech recognition, translation and text-to-speech all run locally using about 3.4 GB of models. On top of that, the app creates meeting notes you can query afterwards.

Tools Nov 16, 2026 7 min read Source: Zenn LLM (JP)
Illustrative image, AI-generated

What's going on

Japanese developer Reo Hamada has released Sokutsu (roughly "instantly understood" in Japanese), an interpreter app for the Mac. Two people sit across from each other and take turns speaking in their own language. The Mac listens, shows the translation as a large subtitle and reads it aloud in a synthetic voice. Everything happens on the machine itself – no servers, no cloud.

Hamada says the idea came to him while watching the Japanese TV series "VIVANT," which features an earpiece with real-time translation. He wanted to build something like it himself. The app is free and released under the MIT license, a very permissive open-source license. It runs on Macs with Apple chips. You can download it as a DMG file from the releases page; the source code is on GitHub. The current version is 1.1.3.

Translating before the sentence is over

The biggest annoyance with interpreter apps is the wait between the end of a sentence and the translation. Sokutsu tackles this in two ways.

Predictive translation: While someone is still speaking, the app produces a provisional translation of what's been said so far. These subtitles appear with a dotted underline. When the sentence ends and the final recognized text matches what the prediction was based on, the translation appears instantly. Otherwise the app translates again.

Detecting when a sentence is done: If someone says "um" or "and then" and pauses briefly, they're not finished. An app that translates at that point produces nonsense. Sokutsu uses the sentence ending to estimate whether a thought is complete or more is coming. That determines how long the app waits: briefly at the end of a sentence, longer when a continuation is likely. If a Japanese sentence ends with a particle like "…desu ga," (roughly "…, but"), the app treats it as incomplete.

The speech output adapts too. If a translation is long, the voice automatically speeds up. Sentences it can no longer keep up with are shown as subtitles only.

Small models, lots of fine-tuning

Large language models would give the best translations, but they don't run on a laptop. So Hamada chose models that let everything run on a laptop Mac without a server:

Task Model Size
Translation Gemma 3 4B (QAT, 4-bit) 2.53 GB
Speech recognition (default) SenseVoice-Small 0.16 GB
Speech recognition (more accurate, optional) Whisper large-v3-turbo (quantized) 0.57 GB
Speech output Kokoro, VOICEVOX (Japanese), macOS "say" –
Full package incl. speaker recognition – approx. 3.4 GB

Gemma 3 4B is a small open model from Google. Sokutsu uses the QAT version (Quantization-Aware Training), meaning the model was prepared during training to be shrunk later. Quantization means the model's numbers are stored at lower precision, in this case 4 bits. With the QAT variant, quality suffers less. The SenseVoice-Small speech recognizer takes about 0.15 seconds per speech segment. It runs via sherpa-onnx, Whisper via whisper.cpp, and the translation model via llama.cpp's llama-server.

The app makes up for the small model's weaknesses with clever engineering around it. It passes the previous utterance along as context, adds translation examples, and retries when the model invents content that isn't in the original or drops parts of long sentences. On top of that there's a built-in glossary with about 3,800 entries in Japanese, English, Chinese and Korean, plus a custom term dictionary. These let you fix typical errors – for example, English abbreviations being recognized as katakana – and translate technical terms consistently.

Benchmarks

Hamada ran the measurements on his own MacBook (M3, 16 GB). The steps to reproduce them are in the repository.

Measurement Condition Result
Time until translation is final Japanese→English, 18 sentence pairs in real time median 0.85 s
Time until speech output starts as above median 1.56 s
Translation quality (chrF) FLEURS, 12 directions JA/EN/ZH/KO, bundled 4B model approx. 39
Translation quality (chrF) same, large model on GPU server approx. 45

chrF is an automatic metric. It compares how many character sequences the machine translation shares with a reference translation. Higher is better. The developer also reports a COMET score (wmt22-comet-da) of 87.83, averaged over 1,106 translations. COMET is a neural-network-based evaluation that often comes closer to human judgment. According to Hamada, the small model lags behind large models on technical terms and long sentences.

More than an interpreter: meeting notes

Sokutsu can also take minutes. The app records everything during the conversation. When you click "End," a local language model generates a summary, decisions, action items, open questions and keywords. Each item links back to the utterance it came from, so you can check later whether it was really said. The style of the summary depends on the type of conversation: regular meeting, one-on-one, sales call, job interview, lecture, brainstorming or Q&A session.

Other features:

  • You can enter an agenda beforehand and tick items off during the meeting.
  • You can ask questions about the transcript, such as "When was the document supposed to be finished?" The answer cites line numbers as evidence, across all meetings if you like. The app also drafts follow-up emails and minutes.
  • The B key flags the current utterance (important, decision, action item, question); M lets you add your own notes.
  • Recordings and videos (m4a, mp3, wav, mp4, mov …) can be transcribed after the fact, including speaker recognition.
  • You can export as PDF, Word, Markdown or subtitles (SRT/WebVTT). The app only adds events to Calendar and tasks to Reminders when you click a button.
  • Speakers are told apart. Voices you've named are recognized again at the next meeting. The voice profiles stay on your Mac.

With the "Microphone + this Mac's audio" setting, the app also translates people you're talking to on Zoom and other video calls (macOS 14.2 or later). For online meetings, Hamada recommends subtitle-only mode, since the microphone pauses during speech output.

What you need to try it

  • A Mac with an Apple chip (M1 or newer) and macOS 13.3 or later. Intel Macs, Windows and Linux aren't supported.
  • 16 GB of RAM is recommended. The app also runs with 8 GB, but slower.
  • About 5 GB of free disk space.
  • An internet connection only once, to download the models (approx. 3.4 GB; the download can be paused and resumed).

The ready-made app doesn't need Python or LM Studio. Drag Sokutsu.app into your Applications folder. Because the app is only ad-hoc signed and not notarized by Apple, you open it the first time with a right-click and "Open." Then allow microphone access, pick the language pair and start the conversation. If the Japanese voice sounds like the default macOS voice, the optional VOICEVOX engine is missing. According to the developer, that's not a bug. If you run it from source, you'll need Python 3.10 or newer and LM Studio with a Gemma 3 4B model on port 1234.

Our take

Sokutsu shows how far local AI has come. A 4-billion-parameter model, speech recognition and speech output fit together in 3.4 GB and deliver translations in under a second on an ordinary MacBook. What really stands out is the attention to detail: predictive translation, sentence-end detection and checks for invented content contribute more here than a bigger model would. For confidential conversations, running fully offline is a real advantage over cloud services like Google Translate or the live translation features of the big platforms.

The developer is upfront about the limits. The app is designed for two-person conversations. When several people talk over each other, the word error rate on the English AMI meeting corpus was 38.5 percent. The 4B model mistranslates proper names and technical terms unless you add them to the term dictionary. All benchmarks come from the developer himself and cover Japanese, English, Chinese and Korean. The source doesn't say whether, or how well, other languages such as German work, and the built-in glossary doesn't cover them. If you have an Apple Silicon Mac and regularly talk with people in these languages, Sokutsu is a free tool well worth trying.

Sources

#translation#interpreter#mac#local#gemma#open source
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.