kedro Whisper pipeline

Overview

This is a simple kedro pipeline that shows how custom datasets can be leveraged to process and transcribe raw audio files using torchaudio.

The pipeline includes one custom dataset class AudioDataSet and two nodes as shown in the image below:

Get Started

1. Clone repository
2. Create new conda environment

conda create -n audio_pipeline python=3.10
conda activate audio_pipeline

3. Install dependencies

pip install -r requirements.txt

4. Make sure that ffmpeg and ffmpeg-python are installed on your machine
5. Create an .env file that contains following value:

OPENAI_API_KEY= # your API key

6. Put .mp3 audio files that should be processed into directory data/01_raw
7. Modify Whisper parameters in file conf/base/parameters.yml according to your needs
8. Run pipeline

kedro run

Kedro then runs the pipeline and executes following steps:

1. Load audio files located in 01_raw
2. Reduce noise in audios by running node int_reduce_noise
3. Save audio files with reduced noise to 05_model_input
4. Transcribe audios with OpenAI Whisper by running node pri_transcription
5. Save transcripts to 07_model_output