weisrc / strong-align

Forced alignment using Wav2Vec2

🦾 Strong Align 🎯

Forced alignment using Wav2Vec2

Installation

pip install git+https://github.com/weisrc/strong-align.git

⚠️ Warning: This package is still in development. The API may change in the future.

Usage

Basic

import torchaudio
from strong_align import align

text = "Hello world! This is a test."
audio, sr = torchaudio.load("test.wav")
audio = audio[0]
audio = torchaudio.transforms.Resample(sr, 16000)(audio)
audio = audio.to("cuda") # or keep it on the CPU
out = align(text, audio, "en", on_progress=print)
print(out)

Custom normalization

You can use your own normalization function by passing it to the align function.

Special use case:

• Romanizing the text to use the English model for languages such as Chinese, Japanese, Korean, etc.

from strong_align.preprocess import NORMALIZE_FUNCS

def my_normalize_normalize(text, mappings, language, labels):
    # do something with text and mappings
    return text, mappings

out = align(text, audio, "en",
      normalize_func=NORMALIZE_FUNCS+[my_normalize_normalize])

Please refer to the normalize.py file for examples of normalization functions.

License

MIT. Wei (weisrc)