Hear to Read is a Flask server that uses OpenAI Whisper (offline) to transscipe voice messages into readable text.

I'am a fighting voice messengers for years, they are the wild west of communication. People talking sometimes forget what they were saying halfway through.

"Where is the relevant information!?", i asked! man ey!

Nobody has time to listen to a 3min voice message, just to know that my brother thinks of visiting my parents maybe next weekend and if I want to join, if they mgith go there when the weather is good enought, because the last days it was not that good and the kids like the outside... - Nope sorry - my eyes can scrim through text much faster than my ears can listen to a voice message. So I decided to fight back and created this project.

I've also brewed up a simple Telegram bot implementation, serving as an easy-to-use front end. Simply forward those voice messages to the bot, and voila! It replies with clear text. 🚀

• Voice messages are sent to the Telegram bot.
• Telegram Bot forwards the voice message to the Flask server.
• Flask server uses OpenAI Whisper to transcribe the voice message.
• And finally, the Flask server sends the transcribed text back to the Telegram bot, which replies to the user.

cd <repo>
docker build -f docker/Dockerfile -t pyh2r .
docker run -d \
    --name hear_to_read \
    -e PYH2R_TELEGRAM_BOT_API_TOKEN=YOUR_TELEGRAM_BOT_API_TOKEN \ # required
    -e PYH2R_STORE_AUDIO_FILES=1 \ # Optional, enables storing audio files
    -v ./uploaded_files:/app/uploaded_files \ # Persistent store uploaded files
    pyh2r

Two things to note:

• OpenAI Whisper is an offline model, depending on your hardware and the length of the voice message, it might some time.
• OpenAI Whisper is not perfect
• OpenAI Whisper comes in differenct flavors, the one used here is the smallest one, so it might not be as accurate as the bigger ones.
• When you got a nvidia GPU, speed will massively increase! just add --gpus all to the docker run command, make sure you got nvidia-docker setup.

Notes:

• 30s voice message, while walking fast with background noise
• Pi-CM4: Raspberry Pi Compute Module 4
• vServer: [email protected] old Xeon VM
• Only one experiment, no warmup, no average