Free OBS captions on CPU only: setting up Hayamimi

What Hayamimi is and why it runs on a CPU

Hayamimi is a free speech recognizer that oboroge0 publishes on GitHub, and it turns your voice into live captions while you're still talking. The name is Japanese for "quick ear." That makes it an option for anyone who wants captions without paying for a cloud service or owning a spare graphics card.

Your CPU is the general-purpose processor in every PC, while a GPU is a separate graphics chip. Hayamimi doesn't use the GPU at all. That means it runs on a laptop with integrated graphics, or on a gaming PC whose GPU is already busy with the game and the encoder.

It runs in under 2GB of RAM with no GPU or cloud API

Hayamimi does all of its recognition on your CPU, with no GPU and no cloud API involved. By default the models it keeps loaded stay under 2GB of RAM, and you can trim that to 1.35GB.

Cloud transcription APIs charge for the audio you send them, as "Gemini 3.5 Transcribe: pricing and API limits" shows. Because Hayamimi never sends audio anywhere, a six-hour stream costs the same as a five-minute test: nothing.

Japanese captions lock in about 100ms after you stop

Captions update about every 0.5 seconds while you speak, and a Japanese line locks in about 100ms after you stop. Those are the author's own measurements. With five languages mixed and every feature switched on, the average delay rose to 236ms, which still keeps the text close behind your voice.

The code is MIT-licensed and free to use

The source code is under the MIT license, so you can use Hayamimi for free on personal and monetized streams alike. The first release, v0.1.0, came out on August 25, 2026, and v0.5.0 added system-audio capture on September 16. As of October 13, the repository had 440 stars on GitHub.

How Hayamimi stays accurate without a GPU

Hayamimi detects the language of each utterance and hands it to a model built for that language, instead of running one general-purpose model for everything. Japanese goes to a Japanese model, English to an English model, and so on. That's how it gets good accuracy without a GPU.

Every model is quantized to INT8, which means its numbers are stored with fewer bits so it needs less memory ("What is LLM quantization? 4-bit vs 8-bit explained" covers the idea). The models run on sherpa-onnx, so you don't need PyTorch or CUDA.

Which model handles which language

Language	Model
Japanese	ReazonSpeech
English and 24 European languages	NVIDIA Parakeet TDT v3
Chinese	Paraformer-zh
Korean and Cantonese	SenseVoice small
About 1,600 other languages	Meta Omnilingual ASR

Each language goes to its own specialist model

Japanese is handled by ReazonSpeech, while English and 24 European languages go to NVIDIA's Parakeet. Chinese uses Paraformer, Korean and Cantonese use SenseVoice, and about 1,600 other languages fall back to Meta's Omnilingual ASR. Language detection runs during roughly the first 4 seconds of speech, and by default Hayamimi only switches languages when two detectors agree.

In Japanese tests, 3.8% errors vs Whisper's 13.8%

On real Japanese broadcast audio, Hayamimi's character error rate was 3.8%, while whisper-large-v3-turbo scored 13.8% on the same clips. Lower is better, so Hayamimi made fewer than a third as many mistakes. On English, the author's scorecard shows a 2.3% word error rate.

Speed came in at 10-50x faster than real time on a 6-core desktop CPU, which is far more headroom than live captions need. Keep in mind that the author ran all of these tests, and the README says the pipeline hasn't been independently verified yet.

Switching languages mid-sentence breaks it

Hayamimi picks one language per utterance, so a Japanese sentence with English words dropped into it comes out garbled or with words missing. If you switch languages sentence by sentence, it keeps up fine. That makes it a good fit for an interpreter alternating between two languages, but a poor one for heavy code-switching.

Installing Hayamimi and testing your microphone

Hayamimi is a Python program with no installer, so you set it up by pasting a few commands into a terminal. On Windows that's Command Prompt or PowerShell, and on a Mac it's the Terminal app. After cloning the repository and creating a virtual environment, you run three commands: install the libraries, download the models, and start Hayamimi.

What you need before you start

Item	Requirement
Python	3.10 or later
ffmpeg	Installed and on PATH
Disk space for models	About 3.1GB (about 1.1GB with --minimal)
Tested OS	Windows 11

You need Python 3.10 or later and ffmpeg

Before anything else, install Python 3.10 or later and ffmpeg, and make sure ffmpeg is on your PATH so any terminal window can find it. Then clone the repository with git and create a virtual environment inside it.

git clone https://github.com/oboroge0/hayamimi.git
cd hayamimi
python -m venv .venv

The models take about 3.1GB, or 1.1GB with --minimal

The full model download is about 3.1GB, and adding --minimal cuts it to about 1.1GB of Japanese and English models. The weights aren't stored in the repository. The download script pulls each one from its original publisher.

.venv\Scripts\pip install -r requirements.txt
.venv\Scripts\python scripts\download_models.py --minimal

Run it once to test your microphone

Start realtime_transcribe.py with no options, say a few words, and your speech should appear as text right away. It's worth doing this before you touch OBS, so you know the microphone and models work. On macOS or Linux, replace .venv\Scripts with .venv/bin in each command.

.venv\Scripts\python scripts\realtime_transcribe.py

How to add OBS captions with a Browser Source

Hayamimi serves its captions as a local web page, and OBS shows that page on stream through a Browser Source. A Browser Source is the OBS source type that renders a web page inside your scene. Since it's built into OBS, there's no plugin to install.

Start the local caption server with --serve

Adding --serve starts a small web server on your PC that serves three pages. The dashboard at http://localhost:8833/dashboard shows in-progress text, finalized lines, translations and a cleaned-up transcript. The page you want for OBS is http://localhost:8833/, which shows only the captions.

.venv\Scripts\python scripts\realtime_transcribe.py --serve

Add the overlay URL as a Browser Source

In OBS, add a new Browser Source to your scene and paste http://localhost:8833/ into its URL field. Set the width and height to fit your canvas, then drag the source to where you want the captions. OBS and Hayamimi need to run on the same PC, because the page only exists locally.

Put final and in-progress lines in separate sources

Adding ?show=final to the URL shows only finished lines, and ?show=partial shows only the line you're still speaking. Add each one as its own Browser Source and you can place them separately, for example finished captions along the bottom and the live line just above. The README doesn't describe settings for fonts or colors.

Translation, speaker labels, and system audio

Hayamimi adds features through command-line options, and they combine freely. Anything you turn on alongside --serve shows up on the dashboard and in the OBS overlay too.

Options that add to your captions

Option	What it does
--translate en	Translates Japanese captions into English
--translate zh,ko	Translates into Chinese, Korean and other languages
--speakers	Labels each line S1, S2 by speaker
--input speaker	Captions audio playing on the PC (Windows only)
--input mix	Captions the mic and PC audio together (Windows only)
--transcript PATH	Appends the cleaned-up transcript to a file

--translate adds a translated line under each caption

With --translate en, every Japanese caption gets an English translation underneath it in real time. English uses FuguMT, a dedicated Japanese-to-English model, and other targets such as Chinese, Korean and Spanish use the multilingual M2M-100.

You can list several targets at once, as in --translate en,zh,ko. The author has measured translation quality for Chinese, Korean and Spanish.

--speakers labels each line by speaker (S1, S2)

Turning on --speakers tags each caption line with a speaker label based on the voice. That helps viewers follow who said what during a collab or a podcast-style stream. Expect the count to run high, though: in the author's tests on four-person meetings, it estimated 4-8 speakers.

Capturing system audio works on Windows only

--input speaker captions whatever your PC is playing, such as the other side of a call or a video, and --input mix combines that with your mic. Mix mode is handy for meeting notes that need both sides of the conversation.

Both rely on Windows audio loopback, so they don't work on macOS or Linux yet. The README recommends headphones so the call audio isn't picked up twice by your mic.

Limitations to know before you go live

Hayamimi was first released in August 2026, and its README is upfront about what doesn't work yet. Do a test stream with your own voice and setup before relying on it for a live audience.

Main limitations listed in the README

Area	Limitation
Tested OS	Windows 11 only
Overlapping speech	Two voices at once aren't separated
Translation	Numbers can break in Japanese-to-Chinese and Japanese-to-Korean
Japanese proper nouns	--hotwords has no effect (use --replace)

It has only been tested on Windows 11

The author developed and tested Hayamimi on Windows 11, and macOS and Linux haven't been tested end to end. Every runtime it uses is cross-platform, so the author expects them to work. If you try it on a Mac, the README asks you to report your results in an issue.

Overlapping speech is not separated

When two people talk at the same time, Hayamimi can't split their voices into separate captions. The speaker label goes to whoever was processed first, so lines can end up attributed to the wrong person during lively crosstalk. Very short phrases right after a jingle or background music can also trip up language detection.

Japanese-to-Chinese or Korean translation can garble numbers

The translation models are small, and numbers aren't reliably preserved when translating Japanese into Chinese or Korean. Hayamimi falls back to showing the original text when it detects a broken number, but don't trust translated prices or figures without checking. Separately, --hotwords doesn't affect Japanese recognition, so fix names it keeps getting wrong with a --replace dictionary written as wrong=right pairs.

Summary: free OBS captions without a GPU

Hayamimi gives you free OBS captions that run entirely on your CPU, in under 2GB of RAM, with Japanese lines finalizing about 100ms after you stop talking. Start it with --serve and add http://localhost:8833/ as a Browser Source to put the captions on stream.

Translation and speaker labels are one option away, but it has only been tested on Windows 11 and it struggles when two languages share a sentence. A good first step is the --minimal install and a short test with your own voice.