Forced alignment for multilingual data

Authors: Xiaohui Zhang, Moto Hira.

This tutorial shows how to align transcript to speech for non-English languages.

The process of aligning non-English (normalized) transcript is identical to aligning English (normalized) transcript, and the process for English is covered in detail in CTC forced alignment tutorial. In this tutorial, we use TorchAudio’s high-level API, torchaudio.pipelines.Wav2Vec2FABundle, which packages the pre-trained model, tokenizer and aligner, to perform the forced alignment with less code.

import torch
import torchaudio

print(torch.__version__)
print(torchaudio.__version__)

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(device)

2.6.0.dev20241104
2.5.0.dev20241105
cuda

from typing import List

import IPython
import matplotlib.pyplot as plt

Creating the pipeline

First, we instantiate the model and pre/post-processing pipelines.

The following diagram illustrates the process of alignment.

The waveform is passed to an acoustic model, which produces the sequence of probability distribution of tokens. The transcript is passed to tokenizer, which converts the transcript to sequence of tokens. Aligner takes the results from the acoustic model and the tokenizer and generate timestamps for each token.

Note

This process expects that the input transcript is already normalized. The process of normalization, which involves romanization of non-English languages, is language-dependent, so it is not covered in this tutorial, but we will breifly look into it.

The acoustic model and the tokenizer must use the same set of tokens. To facilitate the creation of matching processors, Wav2Vec2FABundle associates a pre-trained accoustic model and a tokenizer. torchaudio.pipelines.MMS_FA is one of such instance.

The following code instantiates a pre-trained acoustic model, a tokenizer which uses the same set of tokens as the model, and an aligner.

from torchaudio.pipelines import MMS_FA as bundle

model = bundle.get_model()
model.to(device)

tokenizer = bundle.get_tokenizer()
aligner = bundle.get_aligner()

Note

The model instantiated by MMS_FA’s get_model() method by default includes the feature dimension for <star> token. You can disable this by passing with_star=False.

The acoustic model of MMS_FA was created and open-sourced as part of the research project, Scaling Speech Technology to 1,000+ Languages. It was trained with 23,000 hours of audio from 1100+ languages.

The tokenizer simply maps the normalized characters to integers. You can check the mapping as follow;

print(bundle.get_dict())

{'-': 0, 'a': 1, 'i': 2, 'e': 3, 'n': 4, 'o': 5, 'u': 6, 't': 7, 's': 8, 'r': 9, 'm': 10, 'k': 11, 'l': 12, 'd': 13, 'g': 14, 'h': 15, 'y': 16, 'b': 17, 'p': 18, 'w': 19, 'c': 20, 'v': 21, 'j': 22, 'z': 23, 'f': 24, "'": 25, 'q': 26, 'x': 27, '*': 28}

The aligner internally uses torchaudio.functional.forced_align() and torchaudio.functional.merge_tokens() to infer the time stamps of the input tokens.

The detail of the underlying mechanism is covered in CTC forced alignment API tutorial, so please refer to it.

We define a utility function that performs the forced alignment with the above model, the tokenizer and the aligner.

def compute_alignments(waveform: torch.Tensor, transcript: List[str]):
    with torch.inference_mode():
        emission, _ = model(waveform.to(device))
        token_spans = aligner(emission[0], tokenizer(transcript))
    return emission, token_spans

We also define utility functions for plotting the result and previewing the audio segments.

# Compute average score weighted by the span length
def _score(spans):
    return sum(s.score * len(s) for s in spans) / sum(len(s) for s in spans)


def plot_alignments(waveform, token_spans, emission, transcript, sample_rate=bundle.sample_rate):
    ratio = waveform.size(1) / emission.size(1) / sample_rate

    fig, axes = plt.subplots(2, 1)
    axes[0].imshow(emission[0].detach().cpu().T, aspect="auto")
    axes[0].set_title("Emission")
    axes[0].set_xticks([])

    axes[1].specgram(waveform[0], Fs=sample_rate)
    for t_spans, chars in zip(token_spans, transcript):
        t0, t1 = t_spans[0].start, t_spans[-1].end
        axes[0].axvspan(t0 - 0.5, t1 - 0.5, facecolor="None", hatch="/", edgecolor="white")
        axes[1].axvspan(ratio * t0, ratio * t1, facecolor="None", hatch="/", edgecolor="white")
        axes[1].annotate(f"{_score(t_spans):.2f}", (ratio * t0, sample_rate * 0.51), annotation_clip=False)

        for span, char in zip(t_spans, chars):
            t0 = span.start * ratio
            axes[1].annotate(char, (t0, sample_rate * 0.55), annotation_clip=False)

    axes[1].set_xlabel("time [second]")
    fig.tight_layout()

def preview_word(waveform, spans, num_frames, transcript, sample_rate=bundle.sample_rate):
    ratio = waveform.size(1) / num_frames
    x0 = int(ratio * spans[0].start)
    x1 = int(ratio * spans[-1].end)
    print(f"{transcript} ({_score(spans):.2f}): {x0 / sample_rate:.3f} - {x1 / sample_rate:.3f} sec")
    segment = waveform[:, x0:x1]
    return IPython.display.Audio(segment.numpy(), rate=sample_rate)

Normalizing the transcript

The transcripts passed to the pipeline must be normalized beforehand. The exact process of normalization depends on language.

Languages that do not have explicit word boundaries (such as Chinese, Japanese and Korean) require segmentation first. There are dedicated tools for this, but let’s say we have segmented transcript.

The first step of normalization is romanization. uroman is a tool that supports many languages.

Here is a BASH commands to romanize the input text file and write the output to another text file using uroman.

$ echo "des événements d'actualité qui se sont produits durant l'année 1882" > text.txt
$ uroman/bin/uroman.pl < text.txt > text_romanized.txt
$ cat text_romanized.txt

Cette page concerne des evenements d'actualite qui se sont produits durant l'annee 1882

The next step is to remove non-alphabets and punctuations. The following snippet normalizes the romanized transcript.

import re


def normalize_uroman(text):
    text = text.lower()
    text = text.replace("’", "'")
    text = re.sub("([^a-z' ])", " ", text)
    text = re.sub(' +', ' ', text)
    return text.strip()


with open("text_romanized.txt", "r") as f:
    for line in f:
        text_normalized = normalize_uroman(line)
        print(text_normalized)

Running the script on the above exanple produces the following.

cette page concerne des evenements d'actualite qui se sont produits durant l'annee

Note that, in this example, since “1882” was not romanized by uroman, it was removed in the normalization step. To avoid this, one needs to romanize numbers, but this is known to be a non-trivial task.

Aligning transcripts to speech

Now we perform the forced alignment for multiple languages.

German

text_raw = "aber seit ich bei ihnen das brot hole"
text_normalized = "aber seit ich bei ihnen das brot hole"

url = "https://download.pytorch.org/torchaudio/tutorial-assets/10349_8674_000087.flac"
waveform, sample_rate = torchaudio.load(
    url, frame_offset=int(0.5 * bundle.sample_rate), num_frames=int(2.5 * bundle.sample_rate)
)

assert sample_rate == bundle.sample_rate

transcript = text_normalized.split()
tokens = tokenizer(transcript)

emission, token_spans = compute_alignments(waveform, transcript)
num_frames = emission.size(1)

plot_alignments(waveform, token_spans, emission, transcript)

print("Raw Transcript: ", text_raw)
print("Normalized Transcript: ", text_normalized)
IPython.display.Audio(waveform, rate=sample_rate)

Raw Transcript:  aber seit ich bei ihnen das brot hole
Normalized Transcript:  aber seit ich bei ihnen das brot hole

preview_word(waveform, token_spans[0], num_frames, transcript[0])

aber (0.96): 0.222 - 0.464 sec

preview_word(waveform, token_spans[1], num_frames, transcript[1])

seit (0.78): 0.565 - 0.766 sec

preview_word(waveform, token_spans[2], num_frames, transcript[2])

ich (0.91): 0.847 - 0.948 sec

preview_word(waveform, token_spans[3], num_frames, transcript[3])

bei (0.96): 1.028 - 1.190 sec

preview_word(waveform, token_spans[4], num_frames, transcript[4])

ihnen (0.65): 1.331 - 1.532 sec

preview_word(waveform, token_spans[5], num_frames, transcript[5])

das (0.54): 1.573 - 1.774 sec

preview_word(waveform, token_spans[6], num_frames, transcript[6])

brot (0.86): 1.855 - 2.117 sec

preview_word(waveform, token_spans[7], num_frames, transcript[7])

hole (0.71): 2.177 - 2.480 sec

Chinese

Chinese is a character-based language, and there is not explicit word-level tokenization (separated by spaces) in its raw written form. In order to obtain word level alignments, you need to first tokenize the transcripts at the word level using a word tokenizer like “Stanford Tokenizer”. However this is not needed if you only want character-level alignments.

text_raw = "关 服务 高端 产品 仍 处于 供不应求 的 局面"
text_normalized = "guan fuwu gaoduan chanpin reng chuyu gongbuyingqiu de jumian"

url = "https://download.pytorch.org/torchaudio/tutorial-assets/mvdr/clean_speech.wav"
waveform, sample_rate = torchaudio.load(url)
waveform = waveform[0:1]

assert sample_rate == bundle.sample_rate

transcript = text_normalized.split()
emission, token_spans = compute_alignments(waveform, transcript)
num_frames = emission.size(1)

plot_alignments(waveform, token_spans, emission, transcript)

print("Raw Transcript: ", text_raw)
print("Normalized Transcript: ", text_normalized)
IPython.display.Audio(waveform, rate=sample_rate)

Raw Transcript:  关 服务 高端 产品 仍 处于 供不应求 的 局面
Normalized Transcript:  guan fuwu gaoduan chanpin reng chuyu gongbuyingqiu de jumian

preview_word(waveform, token_spans[0], num_frames, transcript[0])

guan (0.33): 0.020 - 0.141 sec

preview_word(waveform, token_spans[1], num_frames, transcript[1])

fuwu (0.31): 0.221 - 0.583 sec

preview_word(waveform, token_spans[2], num_frames, transcript[2])

gaoduan (0.74): 0.724 - 1.065 sec

preview_word(waveform, token_spans[3], num_frames, transcript[3])

chanpin (0.73): 1.126 - 1.528 sec

preview_word(waveform, token_spans[4], num_frames, transcript[4])

reng (0.86): 1.608 - 1.809 sec

preview_word(waveform, token_spans[5], num_frames, transcript[5])

chuyu (0.80): 1.849 - 2.151 sec

preview_word(waveform, token_spans[6], num_frames, transcript[6])

gongbuyingqiu (0.93): 2.251 - 2.894 sec

preview_word(waveform, token_spans[7], num_frames, transcript[7])

de (0.98): 2.935 - 3.015 sec

preview_word(waveform, token_spans[8], num_frames, transcript[8])

jumian (0.95): 3.075 - 3.477 sec

Polish

text_raw = "wtedy ujrzałem na jego brzuchu okrągłą czarną ranę"
text_normalized = "wtedy ujrzalem na jego brzuchu okragla czarna rane"

url = "https://download.pytorch.org/torchaudio/tutorial-assets/5090_1447_000088.flac"
waveform, sample_rate = torchaudio.load(url, num_frames=int(4.5 * bundle.sample_rate))

assert sample_rate == bundle.sample_rate

transcript = text_normalized.split()
emission, token_spans = compute_alignments(waveform, transcript)
num_frames = emission.size(1)

plot_alignments(waveform, token_spans, emission, transcript)

print("Raw Transcript: ", text_raw)
print("Normalized Transcript: ", text_normalized)
IPython.display.Audio(waveform, rate=sample_rate)

Raw Transcript:  wtedy ujrzałem na jego brzuchu okrągłą czarną ranę
Normalized Transcript:  wtedy ujrzalem na jego brzuchu okragla czarna rane

preview_word(waveform, token_spans[0], num_frames, transcript[0])

wtedy (1.00): 0.783 - 1.145 sec

preview_word(waveform, token_spans[1], num_frames, transcript[1])

ujrzalem (0.96): 1.286 - 1.788 sec

preview_word(waveform, token_spans[2], num_frames, transcript[2])

na (1.00): 1.868 - 1.949 sec

preview_word(waveform, token_spans[3], num_frames, transcript[3])

jego (1.00): 2.009 - 2.230 sec

preview_word(waveform, token_spans[4], num_frames, transcript[4])

brzuchu (0.97): 2.330 - 2.732 sec

preview_word(waveform, token_spans[5], num_frames, transcript[5])

okragla (1.00): 2.893 - 3.415 sec

preview_word(waveform, token_spans[6], num_frames, transcript[6])

czarna (0.90): 3.556 - 3.938 sec

preview_word(waveform, token_spans[7], num_frames, transcript[7])

rane (1.00): 4.098 - 4.399 sec

Portuguese

text_raw = "na imensa extensão onde se esconde o inconsciente imortal"
text_normalized = "na imensa extensao onde se esconde o inconsciente imortal"

url = "https://download.pytorch.org/torchaudio/tutorial-assets/6566_5323_000027.flac"
waveform, sample_rate = torchaudio.load(
    url, frame_offset=int(bundle.sample_rate), num_frames=int(4.6 * bundle.sample_rate)
)

assert sample_rate == bundle.sample_rate

transcript = text_normalized.split()
emission, token_spans = compute_alignments(waveform, transcript)
num_frames = emission.size(1)

plot_alignments(waveform, token_spans, emission, transcript)

print("Raw Transcript: ", text_raw)
print("Normalized Transcript: ", text_normalized)
IPython.display.Audio(waveform, rate=sample_rate)

Raw Transcript:  na imensa extensão onde se esconde o inconsciente imortal
Normalized Transcript:  na imensa extensao onde se esconde o inconsciente imortal

preview_word(waveform, token_spans[0], num_frames, transcript[0])

na (1.00): 0.020 - 0.080 sec

preview_word(waveform, token_spans[1], num_frames, transcript[1])

imensa (0.90): 0.120 - 0.502 sec

preview_word(waveform, token_spans[2], num_frames, transcript[2])

extensao (0.92): 0.542 - 1.205 sec

preview_word(waveform, token_spans[3], num_frames, transcript[3])

onde (1.00): 1.446 - 1.667 sec

preview_word(waveform, token_spans[4], num_frames, transcript[4])

se (0.99): 1.748 - 1.828 sec

preview_word(waveform, token_spans[5], num_frames, transcript[5])

esconde (0.99): 1.888 - 2.591 sec

preview_word(waveform, token_spans[6], num_frames, transcript[6])

o (0.98): 2.852 - 2.872 sec

preview_word(waveform, token_spans[7], num_frames, transcript[7])

inconsciente (0.80): 2.933 - 3.897 sec

preview_word(waveform, token_spans[8], num_frames, transcript[8])

imortal (0.86): 3.937 - 4.560 sec

Italian

text_raw = "elle giacean per terra tutte quante"
text_normalized = "elle giacean per terra tutte quante"

url = "https://download.pytorch.org/torchaudio/tutorial-assets/642_529_000025.flac"
waveform, sample_rate = torchaudio.load(url, num_frames=int(4 * bundle.sample_rate))

assert sample_rate == bundle.sample_rate

transcript = text_normalized.split()
emission, token_spans = compute_alignments(waveform, transcript)
num_frames = emission.size(1)

plot_alignments(waveform, token_spans, emission, transcript)

print("Raw Transcript: ", text_raw)
print("Normalized Transcript: ", text_normalized)
IPython.display.Audio(waveform, rate=sample_rate)

Raw Transcript:  elle giacean per terra tutte quante
Normalized Transcript:  elle giacean per terra tutte quante

preview_word(waveform, token_spans[0], num_frames, transcript[0])

elle (1.00): 0.563 - 0.864 sec

preview_word(waveform, token_spans[1], num_frames, transcript[1])

giacean (0.99): 0.945 - 1.467 sec

preview_word(waveform, token_spans[2], num_frames, transcript[2])

per (1.00): 1.588 - 1.789 sec

preview_word(waveform, token_spans[3], num_frames, transcript[3])

terra (1.00): 1.950 - 2.392 sec

preview_word(waveform, token_spans[4], num_frames, transcript[4])

tutte (1.00): 2.533 - 2.975 sec

preview_word(waveform, token_spans[5], num_frames, transcript[5])

quante (1.00): 3.055 - 3.678 sec

Conclusion

In this tutorial, we looked at how to use torchaudio’s forced alignment API and a Wav2Vec2 pre-trained mulilingual acoustic model to align speech data to transcripts in five languages.

Acknowledgement

Thanks to Vineel Pratap and Zhaoheng Ni for developing and open-sourcing the forced aligner API.