Transcriber

A real-time speech to text transcription app built with Flet and OpenAI Whisper.

transcriber_demo.mp4

Setting Up An Environment

On Windows:

cd transcriber_app
py -3.7 -m venv venv
venv\Scripts\activate
pip install -r requirements.txt

On Unix:

git clone https://github.com/davabase/transcriber_app/
cd transcriber_app
python3.7 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

Running From Source

python transcriber.py

Building From Source

python cx_freeze_setup.py build

This puts the output in the build directory.

Usage

• Select an input source using the dropdown.
• Click "Start Transcribing"

Selecting a specific language will greatly improve the transcription results.

Transcriber can also be used to translate from other languages to English.

You can change the audio model to improve performance. On slow machines you might prefer the Tiny model.

You can also make the window transparent and set the text background, this is useful for overlaying on other apps. There's an invisible draggable zone just above and below the "Stop Transcribing" button, use this to move the window when it is transparent.

Hidden Features

When you stop a transcription, the lines from the transcription will be saved to transcription.txt in the same file as the app.

Starting a transcription saves the current settings to transcriber_settings.yaml. These settings will be loaded automattically the next time you use the program.

In transcriber_settings.yaml you can additionally set:

• transcribe_rate
• seconds_of_silence_between_lines
• max_record_time

which have no onscreen controls.

transcribe_rate defines how often audio data shold be transcribed, in seconds. Decreasing this number will make the app more real time but will place more burden on your computer.

seconds_of_silence_between_lines defines how much silence (audio with volume lower than the volume threshold) is required to automatically break up the transcription into separate lines.

max_record_time defines the maximum amount of time we should allow a recording to be transcribed. Longer transcriptions take longer to transcribe and may no longer be real time. If you have audio with no silence in it, consider reducing this to break up the audio in to smaller chunks.

Q&A

Why do you use cx_Freeze instead of PyInstaller like Flet recommends?

• It looks like PyInstaller and PyTorch don't get along. I tried building with PyInstaller and got errors with missing torch packages. I spent a bit of time trying to get it to work before switching to cx_Freeze which worked almost immediately.

Read more about Whisper here: https://github.com/openai/whisper

Read more about Flet here: https://flet.dev/

The code in this repository is public domain.