LoCapt

LoCapt is a local live-captioning and translation tool for Windows. It can listen to system audio, microphone audio, or both at the same time, then transcribe or translate speech using Faster Whisper.

The app is designed for local/offline use after the model files are downloaded.

Features

• Local speech transcription using Faster Whisper
• Local translation to English
• System audio capture using WASAPI loopback
• Microphone capture
• Split panes for system audio and microphone when both are enabled
• Device sweep tools for finding the active speaker or microphone
• Cache for last-used microphone and system audio devices
• Adjustable chunk size, overlap, silence cutoff, CPU thread count, and max lag
• Optional source-language text shown beside English translation
• Collapsible settings area
• Always-on-top mode

Requirements

• Windows 10 or Windows 11
• Python 3.10 or newer
• A working microphone or system audio output device
• Enough disk space for Whisper model files

Recommended:

• Python 3.11
• 8 GB RAM minimum
• 16 GB RAM recommended
• Modern Intel or AMD CPU

NVIDIA GPU mode untested, edit this line in LoCapt.py for now to try for yourself:

self.model = WhisperModel(model_key, device="cpu", compute_type="int8", cpu_threads=threads)

and change it to:

self.model = WhisperModel(model_key, device="cuda", compute_type="float16")

change compute_type to int8_float16 for lower vram gpus.

Project Structure

A simple setup can look like this:

LoCapt/
├─ LoCapt.py
├─ models/
│  └─ fast-whisper-small/
│     ├─ config.json
│     ├─ model.bin
│     └─ other model files...
├─ .venv/
└─ locapt_settings_cache.json# LoCapt

LoCapt is a local live captioning and translation tool for Windows. It can listen to your microphone, system audio, or both, then transcribe and optionally translate speech using local Faster-Whisper models.

The main goal is to make meetings, videos, calls, and language learning easier without needing a cloud transcription service.

---

Features

• Local speech-to-text using Faster-Whisper
• Optional translation output
• Microphone capture
• System audio capture
• Separate mic and system caption panes when both sources are enabled
• Optional source-language text beside translated text
• Collapsible settings panel
• Cached device settings between sessions
• Configurable model size, compute type, and worker/thread settings
• Simple Windows Python setup

---

---

Project Files

Typical project layout:

LoCapt/
├─ LoCapt.py
├─ requirements.txt
├─ README.md
├─ .venv/
└─ locapt_device_cache.json

locapt_device_cache.json is created automatically. It stores the last selected microphone/system devices and related settings so you do not need to re-select them every time.

---

Setup

Open PowerShell in the project folder/use the terminal in your IDE.

1. Create a virtual environment

python -m venv .venv

2. Activate the virtual environment

.\.venv\Scripts\Activate.ps1

If PowerShell blocks activation scripts, run:

Set-ExecutionPolicy -Scope CurrentUser RemoteSigned

Then activate again:

.\.venv\Scripts\Activate.ps1

3. Install dependencies

python -m pip install --upgrade pip
pip install -r requirements.txt

---

Faster-Whisper Models

LoCapt uses Faster-Whisper models. Models are not downloaded automatically, I have tested these:

whisper large-v2

whisper small

The large model does not run well on my thinkpad, but runs decent with 8 threads on the small version.

Recommended starting points:

Hardware	Recommended Model	Notes
Slow CPU / low RAM	tiny or base	Fastest, least accurate
Decent CPU	small	Good balance
Strong CPU	medium	Better accuracy, slower
NVIDIA GPU	medium, large-v2, or large-v3	Best accuracy if VRAM allows

The first run may take a while because the model has to download. After that, it loads from the local cache.

---

Manual Model Pre-download

You can force a model to download before running LoCapt.

Activate your virtual environment, then run:

python -c "from faster_whisper import WhisperModel; WhisperModel('small', device='cpu', compute_type='int8')"

Replace small with the model you want:

python -c "from faster_whisper import WhisperModel; WhisperModel('medium', device='cpu', compute_type='int8')"

For GPU use, try:

python -c "from faster_whisper import WhisperModel; WhisperModel('medium', device='cuda', compute_type='float16')"

---

Model Storage Location

Faster-Whisper downloads models through Hugging Face and stores them in the local Hugging Face cache.

Usually this is somewhere under:

C:\Users\YOUR_USERNAME\.cache\huggingface\hub

You can change the cache location by setting HF_HOME before running the app:

$env:HF_HOME="C:\AIModels\huggingface"
python LoCapt.py

To make that permanent for your Windows user:

setx HF_HOME "C:\AIModels\huggingface"

Restart PowerShell after using setx.

---

Running LoCapt

From the project folder:

.\.venv\Scripts\Activate.ps1
python LoCapt.py

Or run directly with the venv Python:

.\.venv\Scripts\python.exe LoCapt.py

---

Audio Sources

LoCapt can listen to:

• Microphone audio
• System audio
• Both microphone and system audio at the same time

When both are enabled, LoCapt shows them separately so mic speech and system audio do not get mixed into one caption stream.

Microphone Mode

Use this for:

• Your own speech
• Language practice
• Calls where you only need your side transcribed

System Mode

Use this for:

• YouTube
• Teams
• Zoom
• Browser audio
• Any audio playing through Windows

Both Mode

Use this for:

• Meetings
• Calls
• Language learning
• Comparing your spoken response against the system audio

---

Device Sweep / Device Selection

If LoCapt does not hear audio, use the sweep/test feature to detect working devices.

Recommended process:

1. Select microphone mode and sweep/test mic devices.
2. Select system mode and sweep/test system devices.
3. Confirm the correct input devices are selected.
4. Start listening.

Once working devices are selected, LoCapt saves them to the local cache file so you should not need to sweep every time.

---

Device Cache

LoCapt stores the last used audio devices in a small local cache file.

Example:

locapt_device_cache.json

This file may save:

• Last microphone device
• Last system audio device
• Last source mode
• Last language/translation options
• Last display preferences

If your audio devices change or LoCapt opens the wrong device, delete the cache file and restart the app.

Remove-Item .\locapt_device_cache.json

Then run LoCapt again and re-select/sweep devices.

---

Translation Display Options

LoCapt can show translated text by itself, or show the source-language text beside the translation.

This can be useful for learning because you can compare what was said with the translated version.

There are separate settings for:

• Mic source text beside translation
• System source text beside translation

That means you can enable it for one source and disable it for the other.

---

Settings Panel

The settings panel can be collapsed once your devices and model options are configured.

The app does not auto-collapse settings when listening starts. Use the small collapse button manually when you want more caption space.

---

Performance Settings

Model size

Larger models are more accurate but slower.

Suggested order for testing:

base → small → medium → large-v3

If captions lag badly, lower the model size.

Threads / workers

Increasing thread count does not always improve performance. More threads can help until the CPU is saturated, but too many threads can make performance worse because of overhead and contention. You must reload the model to change the number of threds used.

Good starting points:

CPU Type	Suggested Threads
4-core CPU	4
6-core CPU	4 to 6
8-core CPU	6 to 8
12-core+ CPU	8 to 12

If changing the thread value appears to do nothing, the current pipeline may be bottlenecked by audio chunking, model inference, queue handling, or the selected backend rather than raw CPU thread count.

---

Troubleshooting

The model download is slow

The first run downloads the selected model. Use a smaller model like base or small to test the app first.

The model download fails

Try:

pip install --upgrade huggingface-hub faster-whisper

Then manually pre-download:

python -c "from faster_whisper import WhisperModel; WhisperModel('small', device='cpu', compute_type='int8')"

If you are on a corporate network, the download may be blocked by proxy or SSL rules. Download the model on another network or configure the required proxy/certificate settings. I downloaded manually from hugging face to solve this same issue and dragged the files into a directory of my chosen name. The program searches for folders in the models folder that contain valid config.json and model.bin files, then displays the name of the corresponding folder to choose from.

No microphone audio is detected

Check:

• Windows microphone permissions
• Correct input device selected
• Mic is not muted
• Another app is not taking exclusive control
• Run the mic sweep/test again

No system audio is detected

Check:

• Audio is playing
• Correct output device is selected in Windows
• System capture device was swept/tested
• Bluetooth headphones are not switching modes
• Try another output device

Stop button crashes the app

This usually means one of the audio threads, queues, or device streams is still being accessed while shutdown is happening.

Try:

1. Start the app again.
2. Select only mic or only system mode.
3. Test stop.
4. Then test both mode.

If it only crashes in both mode, the shutdown issue is likely in the dual-stream cleanup logic.

Captions are delayed

Try:

• Smaller model
• Shorter audio chunks
• CPU int8
• GPU float16, if available
• Fewer simultaneous sources
• Lower thread count if CPU usage is maxed out

Translation quality is poor

Try:

• Larger model
• Confirm the source language is correct
• Avoid noisy audio
• Use headphones to prevent mic echo

Wrong device loads every time

Delete the cache:

Remove-Item .\locapt_device_cache.json

Then restart LoCapt and select devices again.

---

Updating Dependencies

Activate the virtual environment:

.\.venv\Scripts\Activate.ps1

Then update packages:

pip install --upgrade faster-whisper sounddevice numpy scipy

---

Notes

• Larger models are not always better for live captions if they cause too much delay.
• Device caching improves convenience, but deleting the cache is the easiest fix after changing headsets, docks, monitors, or audio drivers.
• System audio capture is usually more sensitive to Windows device changes than microphone capture.
• For learning, enabling source text beside translation can make the output more useful than translation alone.

---

TODO

Possible future improvements:

• Installer or one-click launcher
• Automatic model download UI
• GPU impl
• Better crash-safe stop/start handling
• Saved profiles for different audio setups
• Export transcript to text file
• Hotkeys for start/stop and collapse settings