title	AI Voice Transcript
sdk	streamlit
sdk_version	1.40.0
app_file	streamlit_app.py
pinned	false
license	mit

AI Voice Transcript

Local voice-to-text powered by OpenAI's Whisper (via faster-whisper). Runs fully offline on your machine. No API keys, no per-minute charges. Three ways to use it:

Mode	File	Best for
Desktop GUI	app.py	Day-to-day use on your own machine
Web app	streamlit_app.py	Sharing via a URL, deploying to the cloud
CLI	transcribe_file.py, record_and_transcribe.py	Scripting, batch jobs

The YAML block above is read by Hugging Face Spaces — it tells HF this is a Streamlit app and which file to run.

Features

• Drag-and-drop audio/video file transcription (desktop)
• Microphone capture with device selector and silent-recording detection (desktop)
• Live-streaming transcript display as Whisper decodes
• 9 pre-configured languages plus auto-detect
• 5 model sizes — pick speed vs. accuracy
• Optional timestamps
• Copy / Save As / Open Folder shortcuts
• Same backend across all three modes

Quick start (local)

git clone https://github.com/<your-username>/ai-voice-transcript.git
cd ai-voice-transcript

py -3.12 -m venv .venv
.\.venv\Scripts\Activate.ps1

pip install -r requirements.txt           # core deps (web + transcription)
pip install -r requirements-desktop.txt   # add GUI + mic recording

On Linux/macOS use bash and source .venv/bin/activate instead.

Run the desktop GUI

Windows: double-click Launch App.bat.

Any OS:

python app.py

Run the web app locally

streamlit run streamlit_app.py

Open the URL Streamlit prints (usually http://localhost:8501).

Run the CLI

python transcribe_file.py path/to/audio.m4a --model base --language en
python record_and_transcribe.py --seconds 15

Model sizes

Model	Disk	Speed	Accuracy	Memory
tiny	~75 MB	fastest	OK	~400 MB
base	~150 MB	fast	good (default)	~500 MB
small	~500 MB	medium	better	~1 GB
medium	~1.5 GB	slow	great	~3 GB
large-v3	~3 GB	slowest	best	~6 GB

Models auto-download to ~/.cache/huggingface/ on first use.

Deploying online

See DEPLOY.md for step-by-step guides:

• Hugging Face Spaces — recommended (16 GB RAM free tier, no cold starts)
• Render.com — generic web host (works but 512 MB free tier is tight for Whisper)
• Streamlit Community Cloud — native home for Streamlit apps

How it works

1. Whisper is a neural network trained on ~680,000 hours of multilingual audio.
2. faster-whisper converts Whisper to the CTranslate2 runtime with int8 quantization, giving ~4x speedup on CPU and lower memory use.
3. VAD (Voice Activity Detection) filters silence before decoding — speeds things up and prevents Whisper from "hallucinating" text on empty audio.
4. 16 kHz mono is Whisper's native input; the desktop recorder uses this directly so no resampling is needed.
5. Background threading in the GUI keeps the UI responsive during long transcriptions — workers post events to a queue, the main thread polls and renders.

Project layout

ai-voice-transcript/
├── app.py                       # Desktop GUI (Tkinter + tkinterdnd2)
├── streamlit_app.py             # Web app (Streamlit)
├── transcribe_file.py           # CLI: transcribe a file
├── record_and_transcribe.py     # CLI: record from mic + transcribe
├── mic_test.py                  # CLI: diagnose mic issues
├── Launch App.bat               # Windows: double-click to launch GUI
├── Launch App (debug).bat       # Windows: same but with visible console
├── requirements.txt             # Cloud / web deps
├── requirements-desktop.txt     # Adds GUI + mic recording deps
├── DEPLOY.md                    # Step-by-step deployment guide
├── recordings/                  # Captured audio (gitignored)
└── transcripts/                 # Generated text output (gitignored)

License

MIT