b2aiprep 0.14.1

A small package to generate features from acoustic

B2AI Prep

A simple Python package to prepare acoustic data for the Bridge2AI voice project.

Caution: this package is under active development and interfaces may change rapidly over the next few weeks.

Installation

Requires a Python >= 3.10, <3.12 environment

pip install b2aiprep

Usage

Four commands are available through the CLI:

b2aiprep-cli --help

1. Convert an audio file to features:

   The simplest form takes an audio file, a subject id, and a task name.

   b2aiprep-cli convert test_audio.wav s1 mpt
   

   It will save a pytorch .pt file with a dictionary of features. This can be loaded by torch.load(). The file is named following a simple convention: sub-<subject_id>_task-<task_name>_md5-<checksum>_features.pt

   To enable speech to text transcription, specify the --speech2text flag.

2. Batch process audio files

   This requires a CSV file, where there is a column header called filename and then each subsequent line is of the form: path/to/audio.wav

   This also supports a CSV file where each line is of the form: path/to/audio.wav,subject_id,task_name

   To generate this csv file from the Production directory pulled from wasabi, use command 3.

   b2aiprep-cli batchconvert filelist.csv --plugin cf n_procs=2 --outdir out --save_figures
   

   The above command uses pydra under the hood to parallel process the audio files. All outputs are currently stored in a single directory specified by the --outdir flag.

   One can also generate a hugging face dataset in the output directory by specifying the --dataset flag.

   To enable speech to text transcription, specify the --speech2text flag.

3. Generate csv file to feed to batchconvert

   b2aiprep-cli createbatchcsv input_dir outfile
   

   The input directory should point to the location of the Production directory pulled from Wasabi e.g. /Users/b2ai/production.

   This directory can have subfolders for each institution, (e.g. production/MIT), and each subdirectory is expected to have all the .wav files from each institution.

   Outfile is the path to and name of the csv file to be generated, e.g. audiofiles.csv

   The csv file will have a header named filename with all the filenames listed under.

4. Verify if two audio files are from the same speaker

   b2aiprep-cli test_audio1.wav test_audio2.wav --model 'speechbrain/spkrec-ecapa-voxceleb'
   

   This will use the speechbrain speaker recognition model to verify that the two audio files are from the same speaker.

   There is a notebook in the docs directory that can be used to interact with the library programmatically.

5. Convert the speaker in the source audio file (1st argument) into the speaker of the target audio file (2nd argument) and save the result in the output file (3rd argument)

   b2aiprep-cli convert-voice data/vc_source.wav data/vc_target.wav data/vc_output.wav
   

6. Transcribe the audio

   b2aiprep-cli transcribe data/vc_source.wav
   

   Or use a different model. Note that the large model may take some time to download.

   b2aiprep-cli transcribe data/vc_source.wav --model_id 'openai/whisper-large-v3' --return_timestamps true