ezzye/Text-from-Audio-for-Justice

Takes an audio recording and transcribes it into text document so that an user can annotate it. The app then can process the annotated document into audio clips and a document with hyper-links to the clips.

Text-from-Audio-for-Justice

• Clean up audio file using Audacity

• Make tools:

  • Clean up audio file using pydub and sox see video below
  • Record file locations
  • Make segments from transcription file

  • Make training files from transcription and other files:

    • text: utt_id word1 word2 word3..
    • segments: utt_id file_id start_time end_time
    • wav.scp: file_id path/file
    • utt2spk: utt_id spkr
    • spk2utt: spkr utt_id1 utt_id2 utt_id3

  • Transcribe speaker files
  • Spike - compare original transcription with speaker transcription
  • Make output rts, word etc files to be used for training and communication

Useful Links

• Automating Noise Reduction for Audio Processing, video
• Audacity
• How to clean up digital recordings using Audacity
• Quick reStructuredText
• rst-cheatsheet
• Restructured Text (reST) and Sphinx CheatSheet
• Text Markup to PDF with Python video
• Josh’s Kaldi Documentation
• Kaldi Create files for data/train
• List of English pronouciations (a very big file)
• Kaldi lab using TIDIGITS
• KALDI FOR DUMMIES
• A Basic Introduction to Speech Recognition (Hidden Markov Model & Neural Networks)
• A friendly introduction to Bayes Theorem and Hidden Markov Models
• Create and modify Word documents with Python
• FFmpeg
• How do I split an audio file into multiple?

Getting Started

• Install ffmpeg
• Install docker if not already installed
• Download Kaldi docker image (message/email for link)
• Docker load image
• List docker images and containers
• Give docker maximum resources to run
• Run taj

Tools

taj transcribe
    --audio_input_folder
    --output_folder

With:
    wav.scp: chunk file paths (files extracted from segments)
    text: (line for each chunk)
    segments: (links text to chunk file and start and end time)

taj chunk_speaker
    --audio_input_path
    --speech_segmentation_path
    --output_folder

taj clean_up
    --audio_input_folder (original recording)
    --audio_output_folder

taj convert
    --type (either rts, pdf, doc)
    --online_folder (url of online folder)
    --chunks_text_path
    --output_folder

taj create_test_data
    --input_folder
    --output_folder
    --audio_input_folder (original recording(s))

taj retrain
    --input_folder
    --audio_input_folder (original recording(s))