r4victor / afaligner

📈 A forced aligner intended for synchronization of narrated text
MIT License
85 stars 11 forks source link
forced-alignment

afaligner

Overview

afaligner is a Python library for automatic text and audio synchronization (a.k.a. forced aligner). You give it a list of text files and a list of audio files that contain the narrated text, and it produces a mapping between text fragments and the corresponding audio fragments.

afaligner is used in the syncabook command-line tool to produce EPUB3 with Media Overlays ebooks and has been developed for this specific purpose. If you want to create an ebook with synchronized text and audio, consider using syncabook instead of using afaligner directly.

afaligner works by synthesizing text and then aligning synthesized and recorded audio using a variation of the DTW (Dynamic Time Warping) algorithm. The main features of the algorithm are:

afaligner was inspired by aeneas and works in a similar way. It uses aeneas as a dependency for text synthesis and MFCC extraction.

Supported platforms

afaligner works on 64-bit Mac OS and Linux. Windows is not currently supported (you may try to use a VM).

Requirements

Installation

  1. Install Python (>= 3.6)

  2. Install FFmpeg and eSpeak

  3. Install numpy:

    pip install numpy
  4. Install afaligner:

    pip install afaligner

Or if you want to modify the afaligner's source code:

  1. Get the repository:

    git clone https://github.com/r4victor/afaligner/ && cd afaligner
  2. Install afaligner in editable mode:

    pip install -e .

Running tests

  1. Install pytest:

    pip install pytest
  2. Run tests:

    python -m pytest tests/

Installation via Docker

Installing all the afaligner's dependencies can be tedious, so the library comes with Dockerfile. You can use it to build a Debian-based Docker image that contains afaligner itself and all its dependencies. Alternatively, you can use Dockerfile as a reference to install afaligner on your machine.

Installation via Docker:

  1. Get the repository:

    git clone https://github.com/r4victor/afaligner/ && cd afaligner
  2. Build an image:

    docker build -t afaligner .
  3. Now you can run the container like so:

    docker run -ti afaligner

    It enters bash. You can run Python and import afaligner. To do something useful, you may need to mount your code that uses afaligner as a volume.

Usage

afaligner provides only one function called align() that takes a text directory, an audio directory, and a set of output parameters and returns a sync map (a mapping from text fragments to their time positions in audio files). If the output directory is specified, it also writes the result in the JSON or SMIL format to that directory. The call may look like this:

from afaligner import align

sync_map = align(
    'ebooks/demoebook/text/',
    'ebooks/demoebook/audio/',
    output_dir='ebooks/demoebook/smil/',
    output_format='smil',
    sync_map_text_path_prefix='../text/',
    sync_map_audio_path_prefix='../audio/'
)

and sync_map has the following structure:

{
    "p001.xhtml": {
        "f001": {
            "audio_file": "p001.mp3",
            "begin_time": "0:00:00.000",
            "end_time": "0:00:02.600",
        },
        "f002": {
            "audio_file": "p001.mp3",
            "begin_time": "0:00:02.600",
            "end_time": "0:00:05.880",
        },
        # ...
    },
    "p002.xhtml": {
        "f016": {
            "audio_file": "p002.mp3",
            "begin_time": "0:00:00.000",
            "end_time": "0:00:03.040",
        }
        # ...
    },
}

For more details, please refer to docstrings.

Troubleshooting

pip install afaligner may not work on macOS if it tries to compile a universal library. This seems to be because aeneas complies only on x86_64. I got an error when using Python 3.9. The following command fixes it:

ARCHFLAGS="-arch x86_64" pip install afaligner