使用 Wav2Vec2 进行强制对齐

作者: Moto Hira

本教程展示了如何使用 torchaudio 将转录文本与语音对齐，采用的 CTC 分割算法来自 CTC-Segmentation of Large Corpora for German End-to-end Speech Recognition。

本教程最初是为了展示 Wav2Vec2 预训练模型的一个用例而编写的。

TorchAudio 现在提供了一组专为强制对齐设计的 API。CTC 强制对齐 API 教程展示了 torchaudio.functional.forced_align() 的使用方法，这是核心 API。

如果您希望对语料库进行对齐，我们推荐使用 torchaudio.pipelines.Wav2Vec2FABundle，它结合了 forced_align() 和其他支持函数，并使用了专门为强制对齐训练的预训练模型。请参考多语言数据的强制对齐，该文档详细说明了其使用方法。

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
2.6.0
cuda

概述

对齐的过程如下所示。

从音频波形中估计逐帧的标签概率
生成表示时间步长上标签对齐概率的网格矩阵
从网格矩阵中找到最可能的路径

在本示例中，我们使用 torchaudio 的 Wav2Vec2 模型进行声学特征提取。

准备工作

首先我们导入必要的包，并获取我们要处理的数据。

from dataclasses import dataclass

import IPython
import matplotlib.pyplot as plt

torch.random.manual_seed(0)

SPEECH_FILE = torchaudio.utils.download_asset("tutorial-assets/Lab41-SRI-VOiCES-src-sp0307-ch127535-sg0042.wav")

生成逐帧标签概率

第一步是生成每个音频帧的标签类别概率。我们可以使用一个为ASR训练的Wav2Vec2模型。这里我们使用torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H()。

torchaudio 提供了对预训练模型及其相关标签的便捷访问。

在接下来的章节中，我们将在对数域中计算概率，以避免数值不稳定性。为此，我们使用 torch.log_softmax() 对 emission 进行归一化。

bundle = torchaudio.pipelines.WAV2VEC2_ASR_BASE_960H
model = bundle.get_model().to(device)
labels = bundle.get_labels()
with torch.inference_mode():
    waveform, _ = torchaudio.load(SPEECH_FILE)
    emissions, _ = model(waveform.to(device))
    emissions = torch.log_softmax(emissions, dim=-1)

emission = emissions[0].cpu().detach()

print(labels)

('-', '|', 'E', 'T', 'A', 'O', 'N', 'I', 'H', 'S', 'R', 'D', 'L', 'U', 'M', 'W', 'C', 'F', 'G', 'Y', 'P', 'B', 'V', 'K', "'", 'X', 'J', 'Q', 'Z')

可视化

def plot():
    fig, ax = plt.subplots()
    img = ax.imshow(emission.T)
    ax.set_title("Frame-wise class probability")
    ax.set_xlabel("Time")
    ax.set_ylabel("Labels")
    fig.colorbar(img, ax=ax, shrink=0.6, location="bottom")
    fig.tight_layout()


plot()

Frame-wise class probability

生成对齐概率（网格）

从发射矩阵中，我们接下来生成一个表示每个时间帧上转录标签出现概率的网格。

网格是一个二维矩阵，包含时间轴和标签轴。标签轴代表我们正在对齐的转录内容。在下文中，我们使用 \(t\) 表示时间轴的索引，使用 \(j\) 表示标签轴的索引。\(c_j\) 表示标签索引 \(j\) 处的标签。

为了生成时间步 \(t+1\) 的概率，我们查看时间步 \(t\) 的网格以及时间步 \(t+1\) 的发射值。有两种路径可以到达时间步 \(t+1\) 并带有标签 \(c_{j+1}\)。第一种情况是标签在 \(t\) 时刻已经是 \(c_{j+1}\)，并且从 \(t\) 到 \(t+1\) 没有发生标签变化。另一种情况是标签在 \(t\) 时刻是 \(c_j\)，并在 \(t+1\) 时刻过渡到下一个标签 \(c_{j+1}\)。

下图展示了这种过渡。

由于我们正在寻找最可能的过渡路径，因此我们选择更可能的路径作为 \(k_{(t+1, j+1)}\) 的值，即

\(k_{(t+1, j+1)} = max( k_{(t, j)} p(t+1, c_{j+1}), k_{(t, j+1)} p(t+1, repeat) )\)

其中 \(k\) 表示网格矩阵，\(p(t, c_j)\) 表示在时间步 \(t\) 处标签 \(c_j\) 的概率。\(repeat\) 表示 CTC 公式中的空白标记。（有关 CTC 算法的详细信息，请参阅 Sequence Modeling with CTC [distill.pub]）

# We enclose the transcript with space tokens, which represent SOS and EOS.
transcript = "|I|HAD|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT|"
dictionary = {c: i for i, c in enumerate(labels)}

tokens = [dictionary[c] for c in transcript]
print(list(zip(transcript, tokens)))


def get_trellis(emission, tokens, blank_id=0):
    num_frame = emission.size(0)
    num_tokens = len(tokens)

    trellis = torch.zeros((num_frame, num_tokens))
    trellis[1:, 0] = torch.cumsum(emission[1:, blank_id], 0)
    trellis[0, 1:] = -float("inf")
    trellis[-num_tokens + 1 :, 0] = float("inf")

    for t in range(num_frame - 1):
        trellis[t + 1, 1:] = torch.maximum(
            # Score for staying at the same token
            trellis[t, 1:] + emission[t, blank_id],
            # Score for changing to the next token
            trellis[t, :-1] + emission[t, tokens[1:]],
        )
    return trellis


trellis = get_trellis(emission, tokens)

[('|', 1), ('I', 7), ('|', 1), ('H', 8), ('A', 4), ('D', 11), ('|', 1), ('T', 3), ('H', 8), ('A', 4), ('T', 3), ('|', 1), ('C', 16), ('U', 13), ('R', 10), ('I', 7), ('O', 5), ('S', 9), ('I', 7), ('T', 3), ('Y', 19), ('|', 1), ('B', 21), ('E', 2), ('S', 9), ('I', 7), ('D', 11), ('E', 2), ('|', 1), ('M', 14), ('E', 2), ('|', 1), ('A', 4), ('T', 3), ('|', 1), ('T', 3), ('H', 8), ('I', 7), ('S', 9), ('|', 1), ('M', 14), ('O', 5), ('M', 14), ('E', 2), ('N', 6), ('T', 3), ('|', 1)]

可视化

def plot():
    fig, ax = plt.subplots()
    img = ax.imshow(trellis.T, origin="lower")
    ax.annotate("- Inf", (trellis.size(1) / 5, trellis.size(1) / 1.5))
    ax.annotate("+ Inf", (trellis.size(0) - trellis.size(1) / 5, trellis.size(1) / 3))
    fig.colorbar(img, ax=ax, shrink=0.6, location="bottom")
    fig.tight_layout()


plot()

forced alignment tutorial

在上面的可视化中，我们可以看到有一条高概率的轨迹在对角线上穿过矩阵。

寻找最可能的路径（回溯）

一旦生成了网格，我们将按照高概率的元素进行遍历。

我们将从最后一个标签索引开始，选择时间步长中概率最高的点，然后回溯时间，根据转移后的概率 \(k_{t, j} p(t+1, c_{j+1})\) 或 \(k_{t, j+1} p(t+1, repeat)\) 选择停留（\(c_j \rightarrow c_j\)）或转移（\(c_j \rightarrow c_{j+1}\)）。

当标签到达起始点时，转移完成。

网格矩阵用于路径查找，但对于每个片段的最终概率，我们从发射矩阵中获取逐帧的概率。

@dataclass
class Point:
    token_index: int
    time_index: int
    score: float


def backtrack(trellis, emission, tokens, blank_id=0):
    t, j = trellis.size(0) - 1, trellis.size(1) - 1

    path = [Point(j, t, emission[t, blank_id].exp().item())]
    while j > 0:
        # Should not happen but just in case
        assert t > 0

        # 1. Figure out if the current position was stay or change
        # Frame-wise score of stay vs change
        p_stay = emission[t - 1, blank_id]
        p_change = emission[t - 1, tokens[j]]

        # Context-aware score for stay vs change
        stayed = trellis[t - 1, j] + p_stay
        changed = trellis[t - 1, j - 1] + p_change

        # Update position
        t -= 1
        if changed > stayed:
            j -= 1

        # Store the path with frame-wise probability.
        prob = (p_change if changed > stayed else p_stay).exp().item()
        path.append(Point(j, t, prob))

    # Now j == 0, which means, it reached the SoS.
    # Fill up the rest for the sake of visualization
    while t > 0:
        prob = emission[t - 1, blank_id].exp().item()
        path.append(Point(j, t - 1, prob))
        t -= 1

    return path[::-1]


path = backtrack(trellis, emission, tokens)
for p in path:
    print(p)

Point(token_index=0, time_index=0, score=0.9999996423721313)
Point(token_index=0, time_index=1, score=0.9999996423721313)
Point(token_index=0, time_index=2, score=0.9999996423721313)
Point(token_index=0, time_index=3, score=0.9999996423721313)
Point(token_index=0, time_index=4, score=0.9999996423721313)
Point(token_index=0, time_index=5, score=0.9999996423721313)
Point(token_index=0, time_index=6, score=0.9999996423721313)
Point(token_index=0, time_index=7, score=0.9999996423721313)
Point(token_index=0, time_index=8, score=0.9999998807907104)
Point(token_index=0, time_index=9, score=0.9999996423721313)
Point(token_index=0, time_index=10, score=0.9999996423721313)
Point(token_index=0, time_index=11, score=0.9999998807907104)
Point(token_index=0, time_index=12, score=0.9999996423721313)
Point(token_index=0, time_index=13, score=0.9999996423721313)
Point(token_index=0, time_index=14, score=0.9999996423721313)
Point(token_index=0, time_index=15, score=0.9999996423721313)
Point(token_index=0, time_index=16, score=0.9999996423721313)
Point(token_index=0, time_index=17, score=0.9999996423721313)
Point(token_index=0, time_index=18, score=0.9999998807907104)
Point(token_index=0, time_index=19, score=0.9999996423721313)
Point(token_index=0, time_index=20, score=0.9999996423721313)
Point(token_index=0, time_index=21, score=0.9999996423721313)
Point(token_index=0, time_index=22, score=0.9999996423721313)
Point(token_index=0, time_index=23, score=0.9999997615814209)
Point(token_index=0, time_index=24, score=0.9999998807907104)
Point(token_index=0, time_index=25, score=0.9999998807907104)
Point(token_index=0, time_index=26, score=0.9999998807907104)
Point(token_index=0, time_index=27, score=0.9999998807907104)
Point(token_index=0, time_index=28, score=0.9999985694885254)
Point(token_index=0, time_index=29, score=0.9999943971633911)
Point(token_index=0, time_index=30, score=0.9999842643737793)
Point(token_index=1, time_index=31, score=0.9846118092536926)
Point(token_index=1, time_index=32, score=0.9999706745147705)
Point(token_index=1, time_index=33, score=0.15352763235569)
Point(token_index=1, time_index=34, score=0.9999172687530518)
Point(token_index=2, time_index=35, score=0.6091406941413879)
Point(token_index=2, time_index=36, score=0.9997723698616028)
Point(token_index=3, time_index=37, score=0.9997134804725647)
Point(token_index=3, time_index=38, score=0.9999358654022217)
Point(token_index=4, time_index=39, score=0.986176073551178)
Point(token_index=4, time_index=40, score=0.9241712093353271)
Point(token_index=5, time_index=41, score=0.9259618520736694)
Point(token_index=5, time_index=42, score=0.01559634879231453)
Point(token_index=5, time_index=43, score=0.9998377561569214)
Point(token_index=6, time_index=44, score=0.998847484588623)
Point(token_index=7, time_index=45, score=0.10197910666465759)
Point(token_index=7, time_index=46, score=0.9999427795410156)
Point(token_index=8, time_index=47, score=0.9999943971633911)
Point(token_index=8, time_index=48, score=0.9979596138000488)
Point(token_index=9, time_index=49, score=0.035976238548755646)
Point(token_index=9, time_index=50, score=0.06177717074751854)
Point(token_index=9, time_index=51, score=4.336948768468574e-05)
Point(token_index=10, time_index=52, score=0.9999799728393555)
Point(token_index=11, time_index=53, score=0.9967018961906433)
Point(token_index=11, time_index=54, score=0.9999257326126099)
Point(token_index=11, time_index=55, score=0.9999982118606567)
Point(token_index=12, time_index=56, score=0.9990664124488831)
Point(token_index=12, time_index=57, score=0.9999996423721313)
Point(token_index=12, time_index=58, score=0.9999996423721313)
Point(token_index=12, time_index=59, score=0.8452622294425964)
Point(token_index=12, time_index=60, score=0.9999996423721313)
Point(token_index=13, time_index=61, score=0.9996007084846497)
Point(token_index=13, time_index=62, score=0.999998927116394)
Point(token_index=14, time_index=63, score=0.0035339989699423313)
Point(token_index=14, time_index=64, score=1.0)
Point(token_index=14, time_index=65, score=1.0)
Point(token_index=14, time_index=66, score=0.9999915361404419)
Point(token_index=15, time_index=67, score=0.997150719165802)
Point(token_index=15, time_index=68, score=0.9999990463256836)
Point(token_index=15, time_index=69, score=0.9999992847442627)
Point(token_index=15, time_index=70, score=0.9999997615814209)
Point(token_index=15, time_index=71, score=0.9999998807907104)
Point(token_index=15, time_index=72, score=0.9999881982803345)
Point(token_index=15, time_index=73, score=0.011422759853303432)
Point(token_index=15, time_index=74, score=0.9999977350234985)
Point(token_index=16, time_index=75, score=0.9996122717857361)
Point(token_index=16, time_index=76, score=0.999998927116394)
Point(token_index=16, time_index=77, score=0.9728758931159973)
Point(token_index=16, time_index=78, score=0.999998927116394)
Point(token_index=17, time_index=79, score=0.9949368238449097)
Point(token_index=17, time_index=80, score=0.999998927116394)
Point(token_index=17, time_index=81, score=0.9999123811721802)
Point(token_index=17, time_index=82, score=0.9999774694442749)
Point(token_index=18, time_index=83, score=0.6574353575706482)
Point(token_index=18, time_index=84, score=0.9984305500984192)
Point(token_index=18, time_index=85, score=0.9999876022338867)
Point(token_index=19, time_index=86, score=0.9993749260902405)
Point(token_index=19, time_index=87, score=0.9999988079071045)
Point(token_index=19, time_index=88, score=0.10454574227333069)
Point(token_index=19, time_index=89, score=0.9999969005584717)
Point(token_index=20, time_index=90, score=0.3973246216773987)
Point(token_index=20, time_index=91, score=0.9999932050704956)
Point(token_index=21, time_index=92, score=1.6972246612567687e-06)
Point(token_index=21, time_index=93, score=0.9860996603965759)
Point(token_index=21, time_index=94, score=0.9999960660934448)
Point(token_index=22, time_index=95, score=0.9992732405662537)
Point(token_index=22, time_index=96, score=0.9993422627449036)
Point(token_index=22, time_index=97, score=0.9999983310699463)
Point(token_index=23, time_index=98, score=0.9999971389770508)
Point(token_index=23, time_index=99, score=0.9999998807907104)
Point(token_index=23, time_index=100, score=0.9999995231628418)
Point(token_index=23, time_index=101, score=0.9999732971191406)
Point(token_index=24, time_index=102, score=0.9983194470405579)
Point(token_index=24, time_index=103, score=0.9999991655349731)
Point(token_index=24, time_index=104, score=0.9999996423721313)
Point(token_index=24, time_index=105, score=0.9999998807907104)
Point(token_index=24, time_index=106, score=1.0)
Point(token_index=24, time_index=107, score=0.999862790107727)
Point(token_index=24, time_index=108, score=0.9999980926513672)
Point(token_index=25, time_index=109, score=0.9988560676574707)
Point(token_index=25, time_index=110, score=0.9999798536300659)
Point(token_index=26, time_index=111, score=0.8575499653816223)
Point(token_index=26, time_index=112, score=0.9999847412109375)
Point(token_index=27, time_index=113, score=0.987017810344696)
Point(token_index=27, time_index=114, score=1.898651862575207e-05)
Point(token_index=27, time_index=115, score=0.9999796152114868)
Point(token_index=28, time_index=116, score=0.9998251795768738)
Point(token_index=28, time_index=117, score=0.9999990463256836)
Point(token_index=29, time_index=118, score=0.9999732971191406)
Point(token_index=29, time_index=119, score=0.0008991437498480082)
Point(token_index=29, time_index=120, score=0.9993476271629333)
Point(token_index=30, time_index=121, score=0.9975395202636719)
Point(token_index=30, time_index=122, score=0.0003041217278223485)
Point(token_index=30, time_index=123, score=0.9999344348907471)
Point(token_index=31, time_index=124, score=6.082251275074668e-06)
Point(token_index=31, time_index=125, score=0.9833292961120605)
Point(token_index=32, time_index=126, score=0.9974585175514221)
Point(token_index=33, time_index=127, score=0.0008251372491940856)
Point(token_index=33, time_index=128, score=0.9965135455131531)
Point(token_index=34, time_index=129, score=0.017435934394598007)
Point(token_index=34, time_index=130, score=0.9989168643951416)
Point(token_index=35, time_index=131, score=0.9999697208404541)
Point(token_index=36, time_index=132, score=0.9999842643737793)
Point(token_index=36, time_index=133, score=0.9997639060020447)
Point(token_index=37, time_index=134, score=0.5117325186729431)
Point(token_index=37, time_index=135, score=0.9998301267623901)
Point(token_index=38, time_index=136, score=0.08520185202360153)
Point(token_index=38, time_index=137, score=0.004068952519446611)
Point(token_index=38, time_index=138, score=0.9999815225601196)
Point(token_index=39, time_index=139, score=0.012018151581287384)
Point(token_index=39, time_index=140, score=0.9999980926513672)
Point(token_index=39, time_index=141, score=0.000581191445235163)
Point(token_index=39, time_index=142, score=0.9999070167541504)
Point(token_index=40, time_index=143, score=0.9999960660934448)
Point(token_index=40, time_index=144, score=0.9999980926513672)
Point(token_index=40, time_index=145, score=0.9999916553497314)
Point(token_index=41, time_index=146, score=0.9971164464950562)
Point(token_index=41, time_index=147, score=0.9981791973114014)
Point(token_index=41, time_index=148, score=0.9999310970306396)
Point(token_index=42, time_index=149, score=0.9879276156425476)
Point(token_index=42, time_index=150, score=0.999763548374176)
Point(token_index=42, time_index=151, score=0.9999536275863647)
Point(token_index=43, time_index=152, score=0.9999715089797974)
Point(token_index=44, time_index=153, score=0.3192700445652008)
Point(token_index=44, time_index=154, score=0.9997826218605042)
Point(token_index=45, time_index=155, score=0.016051672399044037)
Point(token_index=45, time_index=156, score=0.999901294708252)
Point(token_index=46, time_index=157, score=0.46622487902641296)
Point(token_index=46, time_index=158, score=0.9999994039535522)
Point(token_index=46, time_index=159, score=0.9999996423721313)
Point(token_index=46, time_index=160, score=0.9999995231628418)
Point(token_index=46, time_index=161, score=0.9999996423721313)
Point(token_index=46, time_index=162, score=0.9999996423721313)
Point(token_index=46, time_index=163, score=0.9999996423721313)
Point(token_index=46, time_index=164, score=0.9999995231628418)
Point(token_index=46, time_index=165, score=0.9999995231628418)
Point(token_index=46, time_index=166, score=0.9999996423721313)
Point(token_index=46, time_index=167, score=0.9999996423721313)
Point(token_index=46, time_index=168, score=0.9999995231628418)

可视化

def plot_trellis_with_path(trellis, path):
    # To plot trellis with path, we take advantage of 'nan' value
    trellis_with_path = trellis.clone()
    for _, p in enumerate(path):
        trellis_with_path[p.time_index, p.token_index] = float("nan")
    plt.imshow(trellis_with_path.T, origin="lower")
    plt.title("The path found by backtracking")
    plt.tight_layout()


plot_trellis_with_path(trellis, path)

The path found by backtracking

看起来不错。

分割路径

现在，这条路径包含了相同标签的重复部分，因此让我们将它们合并，使其更接近原始转录文本。

在合并多个路径点时，我们只需对合并的片段取平均概率。

# Merge the labels
@dataclass
class Segment:
    label: str
    start: int
    end: int
    score: float

    def __repr__(self):
        return f"{self.label}\t({self.score:4.2f}): [{self.start:5d}, {self.end:5d})"

    @property
    def length(self):
        return self.end - self.start


def merge_repeats(path):
    i1, i2 = 0, 0
    segments = []
    while i1 < len(path):
        while i2 < len(path) and path[i1].token_index == path[i2].token_index:
            i2 += 1
        score = sum(path[k].score for k in range(i1, i2)) / (i2 - i1)
        segments.append(
            Segment(
                transcript[path[i1].token_index],
                path[i1].time_index,
                path[i2 - 1].time_index + 1,
                score,
            )
        )
        i1 = i2
    return segments


segments = merge_repeats(path)
for seg in segments:
    print(seg)

|       (1.00): [    0,    31)
I       (0.78): [   31,    35)
|       (0.80): [   35,    37)
H       (1.00): [   37,    39)
A       (0.96): [   39,    41)
D       (0.65): [   41,    44)
|       (1.00): [   44,    45)
T       (0.55): [   45,    47)
H       (1.00): [   47,    49)
A       (0.03): [   49,    52)
T       (1.00): [   52,    53)
|       (1.00): [   53,    56)
C       (0.97): [   56,    61)
U       (1.00): [   61,    63)
R       (0.75): [   63,    67)
I       (0.88): [   67,    75)
O       (0.99): [   75,    79)
S       (1.00): [   79,    83)
I       (0.89): [   83,    86)
T       (0.78): [   86,    90)
Y       (0.70): [   90,    92)
|       (0.66): [   92,    95)
B       (1.00): [   95,    98)
E       (1.00): [   98,   102)
S       (1.00): [  102,   109)
I       (1.00): [  109,   111)
D       (0.93): [  111,   113)
E       (0.66): [  113,   116)
|       (1.00): [  116,   118)
M       (0.67): [  118,   121)
E       (0.67): [  121,   124)
|       (0.49): [  124,   126)
A       (1.00): [  126,   127)
T       (0.50): [  127,   129)
|       (0.51): [  129,   131)
T       (1.00): [  131,   132)
H       (1.00): [  132,   134)
I       (0.76): [  134,   136)
S       (0.36): [  136,   139)
|       (0.50): [  139,   143)
M       (1.00): [  143,   146)
O       (1.00): [  146,   149)
M       (1.00): [  149,   152)
E       (1.00): [  152,   153)
N       (0.66): [  153,   155)
T       (0.51): [  155,   157)
|       (0.96): [  157,   169)

可视化

def plot_trellis_with_segments(trellis, segments, transcript):
    # To plot trellis with path, we take advantage of 'nan' value
    trellis_with_path = trellis.clone()
    for i, seg in enumerate(segments):
        if seg.label != "|":
            trellis_with_path[seg.start : seg.end, i] = float("nan")

    fig, [ax1, ax2] = plt.subplots(2, 1, sharex=True)
    ax1.set_title("Path, label and probability for each label")
    ax1.imshow(trellis_with_path.T, origin="lower", aspect="auto")

    for i, seg in enumerate(segments):
        if seg.label != "|":
            ax1.annotate(seg.label, (seg.start, i - 0.7), size="small")
            ax1.annotate(f"{seg.score:.2f}", (seg.start, i + 3), size="small")

    ax2.set_title("Label probability with and without repetation")
    xs, hs, ws = [], [], []
    for seg in segments:
        if seg.label != "|":
            xs.append((seg.end + seg.start) / 2 + 0.4)
            hs.append(seg.score)
            ws.append(seg.end - seg.start)
            ax2.annotate(seg.label, (seg.start + 0.8, -0.07))
    ax2.bar(xs, hs, width=ws, color="gray", alpha=0.5, edgecolor="black")

    xs, hs = [], []
    for p in path:
        label = transcript[p.token_index]
        if label != "|":
            xs.append(p.time_index + 1)
            hs.append(p.score)

    ax2.bar(xs, hs, width=0.5, alpha=0.5)
    ax2.axhline(0, color="black")
    ax2.grid(True, axis="y")
    ax2.set_ylim(-0.1, 1.1)
    fig.tight_layout()


plot_trellis_with_segments(trellis, segments, transcript)

Path, label and probability for each label, Label probability with and without repetation

看起来不错。

将片段合并为单词

现在让我们合并这些单词。Wav2Vec2 模型使用 '|' 作为单词边界，因此我们在每次出现 '|' 之前合并这些片段。

最后，我们将原始音频分割成多个片段，并逐一听取以确认分割是否正确。

# Merge words
def merge_words(segments, separator="|"):
    words = []
    i1, i2 = 0, 0
    while i1 < len(segments):
        if i2 >= len(segments) or segments[i2].label == separator:
            if i1 != i2:
                segs = segments[i1:i2]
                word = "".join([seg.label for seg in segs])
                score = sum(seg.score * seg.length for seg in segs) / sum(seg.length for seg in segs)
                words.append(Segment(word, segments[i1].start, segments[i2 - 1].end, score))
            i1 = i2 + 1
            i2 = i1
        else:
            i2 += 1
    return words


word_segments = merge_words(segments)
for word in word_segments:
    print(word)

I       (0.78): [   31,    35)
HAD     (0.84): [   37,    44)
THAT    (0.52): [   45,    53)
CURIOSITY       (0.89): [   56,    92)
BESIDE  (0.94): [   95,   116)
ME      (0.67): [  118,   124)
AT      (0.66): [  126,   129)
THIS    (0.70): [  131,   139)
MOMENT  (0.88): [  143,   157)

可视化

def plot_alignments(trellis, segments, word_segments, waveform, sample_rate=bundle.sample_rate):
    trellis_with_path = trellis.clone()
    for i, seg in enumerate(segments):
        if seg.label != "|":
            trellis_with_path[seg.start : seg.end, i] = float("nan")

    fig, [ax1, ax2] = plt.subplots(2, 1)

    ax1.imshow(trellis_with_path.T, origin="lower", aspect="auto")
    ax1.set_facecolor("lightgray")
    ax1.set_xticks([])
    ax1.set_yticks([])

    for word in word_segments:
        ax1.axvspan(word.start - 0.5, word.end - 0.5, edgecolor="white", facecolor="none")

    for i, seg in enumerate(segments):
        if seg.label != "|":
            ax1.annotate(seg.label, (seg.start, i - 0.7), size="small")
            ax1.annotate(f"{seg.score:.2f}", (seg.start, i + 3), size="small")

    # The original waveform
    ratio = waveform.size(0) / sample_rate / trellis.size(0)
    ax2.specgram(waveform, Fs=sample_rate)
    for word in word_segments:
        x0 = ratio * word.start
        x1 = ratio * word.end
        ax2.axvspan(x0, x1, facecolor="none", edgecolor="white", hatch="/")
        ax2.annotate(f"{word.score:.2f}", (x0, sample_rate * 0.51), annotation_clip=False)

    for seg in segments:
        if seg.label != "|":
            ax2.annotate(seg.label, (seg.start * ratio, sample_rate * 0.55), annotation_clip=False)
    ax2.set_xlabel("time [second]")
    ax2.set_yticks([])
    fig.tight_layout()


plot_alignments(
    trellis,
    segments,
    word_segments,
    waveform[0],
)

forced alignment tutorial

音频样本

def display_segment(i):
    ratio = waveform.size(1) / trellis.size(0)
    word = word_segments[i]
    x0 = int(ratio * word.start)
    x1 = int(ratio * word.end)
    print(f"{word.label} ({word.score:.2f}): {x0/bundle.sample_rate:.3f} - {x1/bundle.sample_rate:.3f} sec")
    segment = waveform[:, x0:x1]
    return IPython.display.Audio(segment.numpy(), rate=bundle.sample_rate)

# Generate the audio for each segment
print(transcript)
IPython.display.Audio(SPEECH_FILE)

|I|HAD|THAT|CURIOSITY|BESIDE|ME|AT|THIS|MOMENT|

display_segment(0)

I (0.78): 0.624 - 0.704 sec

display_segment(1)

HAD (0.84): 0.744 - 0.885 sec

display_segment(2)

THAT (0.52): 0.905 - 1.066 sec

display_segment(3)

CURIOSITY (0.89): 1.127 - 1.851 sec

display_segment(4)

BESIDE (0.94): 1.911 - 2.334 sec

display_segment(5)

ME (0.67): 2.374 - 2.495 sec

display_segment(6)

AT (0.66): 2.535 - 2.595 sec

display_segment(7)

THIS (0.70): 2.635 - 2.796 sec

display_segment(8)

MOMENT (0.88): 2.877 - 3.159 sec

结论

在本教程中，我们探讨了如何使用 torchaudio 的 Wav2Vec2 模型来执行 CTC 分割以实现强制对齐。

下载 Python 源代码: forced_alignment_tutorial.py

下载 Jupyter 笔记本: forced_alignment_tutorial.ipynb