WhisperUI

A desktop GUI and command-line tool for OpenAI Whisper — record or open audio/video, transcribe or translate in 99 languages, choose models, and export to Text, SRT, WebVTT, TSV, or JSON.

Features

• Desktop GUI (PySide6) with a dark theme
• CLI with full option parity — no display required
• Importable Python API for use in other projects
• Model selector with descriptions, sizes, and in-app download with live progress bar
• Microphone recording with real-time VU meter and elapsed timer
• Open any audio/video format supported by ffmpeg
• Transcription (keeps source language) and translation (always to English)
• Language auto-detection or manual selection across 99 languages
• CUDA GPU acceleration (auto-detected; falls back to CPU)
• Word-level timestamps
• Output formats: plain text, SRT, WebVTT, TSV, JSON

Requirements

• Python 3.8+
• ffmpeg on your system PATH (Whisper uses it for audio decoding)
• A CUDA-capable GPU is optional but strongly recommended for medium/large models

Installation

git clone https://github.com/applebiter/whisper-ui.git
cd whisper-ui
python -m venv .venv
source .venv/bin/activate       # Windows: .venv\Scripts\activate
pip install -r requirements.txt

Install ffmpeg if you don't already have it:

sudo apt install ffmpeg          # Debian / Ubuntu / Mint
brew install ffmpeg              # macOS
winget install ffmpeg            # Windows

Usage

Desktop GUI

python main.py

Select a model (download it if needed), choose an audio source (file or microphone), configure options, and click Transcribe.

CLI

python main.py [--help] {transcribe,models} ...

Run without arguments to open the GUI. Pass a subcommand for headless operation.

transcribe — process an audio/video file

# Basic — output goes to stdout
python main.py transcribe interview.mp3

# Save as an SRT subtitle file
python main.py transcribe lecture.mp4 --format srt --output lecture.srt

# Translate from French to English, auto-download the model if missing
python main.py transcribe discours.mp3 \
    --model large-v3-turbo \
    --task translate \
    --auto-download \
    --output english.txt

# Quiet mode — only the transcript, nothing else
python main.py transcribe audio.wav --model tiny --quiet > transcript.txt

All options for transcribe:

Flag	Default	Description
--model, -m	large-v3-turbo	Model name (see models command)
--task, -t	transcribe	transcribe or translate (→ English)
--language, -l	auto	ISO-639-1 code, e.g. en, fr, de
--device, -d	auto	cuda or cpu
--format, -f	txt	txt, srt, vtt, tsv, json
--output, -o	stdout	Write to this file instead of stdout
--word-timestamps	off	Add word-level timing to each segment
--no-condition	off	Disable conditioning on previous segment (reduces looping)
--initial-prompt	—	Seed text to guide transcription style or vocabulary
--auto-download	off	Download the model automatically if not present
--verbose	off	Print each decoded segment to stderr
--quiet, -q	off	Suppress all status messages

models — list and download models

# Show all models and whether they're downloaded
python main.py models

# Download a specific model
python main.py models --download large-v3-turbo

Python API

whisper_ui.core exposes a clean, Qt-free API for use in other applications:

from whisper_ui.core import (
    available_models,       # () → list[str]
    model_info,             # (name) → {"size": str, "description": str}
    is_model_downloaded,    # (name) → bool
    download_model,         # (name, progress_callback=None) → path str
    transcribe,             # (audio_path, model_name, ...) → dict
    format_result,          # (result, fmt) → str
    FORMAT_KEYS,            # ["txt", "srt", "vtt", "tsv", "json"]
)

# Check and download a model
if not is_model_downloaded("base"):
    download_model(
        "base",
        progress_callback=lambda pct, speed: print(f"{pct}%  {speed}")
    )

# Transcribe
result = transcribe(
    "audio.mp3",
    model_name="base",
    device=None,            # None → auto-detect CUDA or CPU
    task="transcribe",
    language=None,          # None → auto-detect
    word_timestamps=False,
    status_callback=print,
)

print(result["text"])       # full transcript
print(result["language"])   # detected language code

# Format the result
from whisper_ui.core import format_result
print(format_result(result, "srt"))

Format converters are also importable individually:

from whisper_ui.formats import to_text, to_srt, to_vtt, to_tsv, to_json

Models

Model	Size	Notes
tiny / tiny.en	~75 MB	Fastest, lowest accuracy
base / base.en	~142 MB	Fast with reasonable accuracy
small / small.en	~466 MB	Good speed/quality balance
medium / medium.en	~1.5 GB	High accuracy, moderate speed
large-v1 / v2 / v3	~2.9 GB	Highest accuracy
large-v3-turbo	~1.5 GB	★ Recommended — near large-v3 quality at ~8× the speed

.en variants are English-only and slightly more accurate for English audio. Models are cached in ~/.cache/whisper/ and shared with the Whisper CLI.

License

MIT