This Python package provides an efficient way to perform forced alignment between text and audio using Hugging Face's pretrained models. it also features an improved implementation to use much less memory than TorchAudio forced alignment API.
The model checkpoint uploaded here is a conversion from torchaudio to HF Transformers for the MMS-300M checkpoint trained on forced alignment dataset
1import torch
2from ctc_forced_aligner import (
3 load_audio,
4 load_alignment_model,
5 generate_emissions,
6 preprocess_text,
7 get_alignments,
8 get_spans,
9 postprocess_results,
10)
11
12audio_path = "your/audio/path"
13text_path = "your/text/path"
14language = "iso" # ISO-639-3 Language code
15device = "cuda" if torch.cuda.is_available() else "cpu"
16batch_size = 16
17
18
19alignment_model, alignment_tokenizer = load_alignment_model(
20 device,
21 dtype=torch.float16 if device == "cuda" else torch.float32,
22)
23
24audio_waveform = load_audio(audio_path, alignment_model.dtype, alignment_model.device)
25
26
27with open(text_path, "r") as f:
28 lines = f.readlines()
29text = "".join(line for line in lines).replace("\n", " ").strip()
30
31emissions, stride = generate_emissions(
32 alignment_model, audio_waveform, batch_size=batch_size
33)
34
35tokens_starred, text_starred = preprocess_text(
36 text,
37 romanize=True,
38 language=language,
39)
40
41segments, scores, blank_token = get_alignments(
42 emissions,
43 tokens_starred,
44 alignment_tokenizer,
45)
46
47spans = get_spans(tokens_starred, segments, blank_token)
48
49word_timestamps = postprocess_results(text_starred, spans, stride, scores)
