/LLRT_whisper

A not very efficient attempt to create a real time openai/whisper (Audio to Text Transcriber)

Primary LanguagePythonMIT LicenseMIT

LLRT_whisper (Low Level Real Time whisper)

This is a 'not very efficient' attemp to create a real time openai/whisper (Audio to Text Transcriber)

The intention is to create an open-source device to help people with hearing impairments 🧏 to see what people are talking about.

Audio to Text Transcriber

This project is a simple Python script that transcribes audio input from the microphone into text using the whisper library. The script records audio input, processes it in a queue, and transcribes the audio into text using the chosen language.

Features

  • Microphone selection
  • Background noise and speech volume calibration (For speech detection)
  • Language selection
  • Audio queue system
  • Audio storage in RAM instead of on disk (Fail)

How It Works

The script follows these steps:

  1. Microphone selection: Allows the user to choose the desired microphone to record audio input.
  2. Threshold selection: Users can choose from using the previous calibration, start a new calibration, or manually set a threshold.
  3. Language selection: Users can choose the language for transcription or use auto-detection.
  4. Record audio input: The script records audio input from the microphone and stores it in a queue.
  5. Process audio queue: Another thread processes the audio queue, transcribes the audio into text using the chosen language, and prints the result.

Notes

  • This script is compatible with Python 3.10
  • The supported audio formats for the whisper library are .wav and .mp3.

Requirements

  • Python 3.10

Setup (CPU Only version)

1a. Install Python 3.10 for your specific operating system.

Python 3.10 For Windows

Python 3.10 For MacOS

Python 3.10 For Linux/UNIX

1b. Install Miniconda if you haven't already.

2. Create a new Conda environment:

In Anaconda Prompt (miniconda3)

conda create --name LLRT_whisper python=3.10

conda activate LLRT_whisper

3. Install the required libraries:

pip install pyaudio

pip install whisper

4. To ensure the whisper library functions correctly, you also need to have ffmpeg installed on your system. Below are instructions for installing ffmpeg on different platforms.

Installing ffmpeg

# Ubuntu or Debian

sudo apt update && sudo apt install ffmpeg

# Arch Linux

sudo pacman -S ffmpeg

# MacOS (using Homebrew) If you don't have Homebrew installed, you can install it from https://brew.sh/.

brew install ffmpeg

# Windows (using Chocolatey) If you don't have Chocolatey installed, you can install it from https://chocolatey.org/.
# To install chocolatey run Anaconda Powershell Prompt (miniconda3) in Admin and run
# Set-ExecutionPolicy Bypass -Scope Process -Force; [System.Net.ServicePointManager]::SecurityProtocol = [System.Net.ServicePointManager]::SecurityProtocol -bor 3072; iex ((New-Object System.Net.WebClient).DownloadString('https://community.chocolatey.org/install.ps1'))

choco install ffmpeg

# Windows (using Scoop) If you don't have Scoop installed, you can install it from https://scoop.sh/.

scoop install ffmpeg

4. Clone the repository:

git clone https://github.com/Megumin6626/LLRT_whisper.git

5. Change to the project directory:

cd LLRT_whisper

6. Run the script to start:

python Whisper_RT_CPU_Only.py

Follow the on-screen instructions to set up the microphone, threshold, and language.

After the initial setup, the script will continuously listen to your microphone and transcribe the audio in real-time.

Enable GPU acceleration

Before continuous you should have the CPU version setup make sure it work.

This guide will help you set up and run the GPU version of the real-time transcription script using OpenAI's whisper library.

Requirements

  • CUDA Toolkit 11.7

Setup

Installing CUDA Toolkit 11.7

Follow the instructions on the NVIDIA website to install the CUDA Toolkit 11.7.

Installing Python Dependencies

In the same environment, run the following commands:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
pip install soundfile

Testing GPU Support

After installing the dependencies, run the GPU_test.py script to check if your GPU is working properly:

python GPU_test.py

If the GPU is working, you should see the following output:

Torch version: 2.0.0+cu117
CUDA available: True

Running the GPU-Enabled Transcription Script

Once your GPU is set up and working, run the Whisper_RT_GPU.py script to start the real-time transcription with GPU support:

python Whisper_RT_GPU.py

Follow the on-screen instructions to set up the microphone, threshold, and language.

After the initial setup, the script will continuously listen to your microphone and transcribe the audio in real-time using your GPU.

Flow Chart to illustrate how the code works:

1. Start
|
2. List available microphones and prompt user to choose one
|
3. Choose threshold (use previous, measure new, or enter custom value)
|
4. Choose language for transcription
|
5. Initialize audio_queue and start two threads:
   |
   5.1. Thread 1: record_audio()
   |     |
   |     5.1.1. Continuously record audio from microphone
   |     |
   |     5.1.2. If audio level exceeds threshold, start recording
   |     |
   |     5.1.3. If audio level falls below threshold for 0.6 seconds, stop recording
   |     |
   |     5.1.4. Save recorded audio as a wave file and add it to audio_queue
   |
   5.2. Thread 2: process_audio_queue()
         |
         5.2.1. Load Whisper model
         |
         5.2.2. Continuously process audio files in audio_queue
               |
               5.2.2.1. Transcribe audio using chosen language
               |
               5.2.2.2. Print transcribed text to console
               |
               5.2.2.3. Remove processed audio file
|
6. Wait for both threads to finish (This should run indefinitely until interrupted)
|
7. End

Code explan

STOP HERE TO KEEP YOUR SANITY

get_unique_devices(p)

This function takes a PyAudio object and returns a dictionary of unique audio devices (microphones) found on the system.

save_microphone_index(mic_index) and load_microphone_index()

These functions save and load the previously chosen microphone index to a file ("microphones_setting.txt").

record_audio(threshold, audio_queue)

This function records audio from the selected microphone and stores the recorded audio as a wave file. It uses a threshold value to determine when to start and stop recording. The recorded audio files are added to a queue for further processing. How fast the recording will stop when it is silence can be chage here

if silent_frames > 0.6 * RATE / CHUNK: # stop recording after 0.6 seconds of silence

Sidenote: When this is set too low and the file is created too fast, somehow the audio file might be deleted before being processed. This method also poses another problem. If the environment noise suddenly becomes very loud and swamp the threshold, the recording might not be able to stop itself.

process_audio_queue(audio_queue, language)

This function takes the audio queue and language as input, loads the Whisper model, and transcribes the audio files in the queue. The transcribed text is then printed on the screen.

measure_threshold()

This function measures the sound intensity of background noise and speech levels to determine a suitable threshold for starting and stopping audio recording. The threshold calculated using threshold = (background + speech) / 2

save_threshold(threshold) and load_threshold()

These functions save and load the previously measured threshold value to a file ("threshold.txt").

choose_threshold()

This function allows the user to choose how to set the threshold value: use the previously saved threshold, measure a new threshold, or enter a custom threshold value.

save_language(language) and load_language()

These functions save and load the previously chosen language for transcription to a file ("language_settings.txt").

choose_language()

This function allows the user to choose the input language for transcription.

Further Development

I have identified several areas where the project can be improved. If you are interested in contributing or have ideas of your own, feel free to open an issue or submit a pull request. Some of the areas for future development include:

1. Save audio files in RAM

Storing audio files in RAM instead of writing them to the SSD can help extend the lifespan of the storage device, as frequent writing and erasing of files can wear out an SSD.

Note : attempted but fail due to me suck at codeing. :(

2. Improve voice activity detection

Explore more efficient methods for determining when the user is speaking, such as implementing a Voice Activity Detector (VAD).

Note : in my attempt the effectiveness of VAD may vary depending on the hardware being used so.

3. Modify record_audio function

Update the record_audio(threshold, audio_queue) function to stop and create a new audio file after a certain amount of time(eg. 30s). This will prevent the threshold from being overwhelmed by noise and generating infinitely long audio clips.

Note : Problem is addressed with another approach.Now the Slience time will turn into 0s after a certain of time. For more detail check Update Log - 2023-03-23 V1.1

4. Speaker identification

Implement speaker identification by first applying a Fast Fourier Transform (FFT) to the audio data and determining the overall speaking pitch. This information can then be used to identify who is speaking in the conversation.

Note : Just saying not attempted yet.

5. Add support to Faster Whisper by guillaumekln

which is up to 4 times faster than openai/whisper for the same accuracy while using less memory. https://github.com/guillaumekln/faster-whisper

Feel free to contribute to these ideas or propose new ones to enhance the functionality and performance of the project.

License

Both Whisper's code and model weights, as well as this project's code, are released under the MIT License. This permissive license allows for reuse, modification, and distribution, both for commercial and non-commercial purposes, provided that the copyright notice and the license's permission notice are included in all copies or substantial portions of the software. See the LICENSE file for further details.