注意

点击此处下载完整的示例代码

使用 CTC 解码器的 ASR 推理¶

**作者**：Caroline Chen

本教程演示了如何使用带词典约束和 KenLM 语言模型支持的 CTC 束搜索解码器执行语音识别推理。我们将在使用 CTC 损失训练的预训练 wav2vec 2.0 模型上演示此功能。

概述¶

束搜索解码通过迭代地扩展文本假设（束）并添加可能的下一个字符来工作，并在每个时间步长仅保留得分最高的假设。可以将语言模型整合到评分计算中，并且添加词典约束会限制假设的下一个可能标记，以便只能生成词典中的单词。

底层实现是从 Flashlight 的束搜索解码器移植而来。解码器优化的数学公式可以在 Wav2Letter 论文中找到，更详细的算法可以在这篇博文中找到。

使用带语言模型和词典约束的 CTC 束搜索解码器运行 ASR 推理需要以下组件

声学模型：从音频波形预测语音的模型
标记：声学模型可能预测的标记
词典：可能单词与其对应标记序列之间的映射
语言模型 (LM)：使用 KenLM 库训练的 N 元语言模型，或继承 CTCDecoderLM 的自定义语言模型

声学模型和设置¶

首先，我们导入必要的实用程序并获取我们要处理的数据

import torch
import torchaudio

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

2.5.0
2.5.0

import time
from typing import List

import IPython
import matplotlib.pyplot as plt
from torchaudio.models.decoder import ctc_decoder
from torchaudio.utils import download_asset

我们使用在 LibriSpeech 数据集的 10 分钟数据上微调的预训练 Wav2Vec 2.0 基础模型，可以使用 torchaudio.pipelines.WAV2VEC2_ASR_BASE_10M 加载。有关在 torchaudio 中运行 Wav2Vec 2.0 语音识别流水线的更多详细信息，请参阅本教程。

bundle = torchaudio.pipelines.WAV2VEC2_ASR_BASE_10M
acoustic_model = bundle.get_model()

Downloading: "https://download.pytorch.org/torchaudio/models/wav2vec2_fairseq_base_ls960_asr_ll10m.pth" to /root/.cache/torch/hub/checkpoints/wav2vec2_fairseq_base_ls960_asr_ll10m.pth

  0%|          | 0.00/360M [00:00<?, ?B/s]
  4%|4         | 16.0M/360M [00:00<00:05, 68.5MB/s]
  9%|8         | 32.0M/360M [00:00<00:06, 55.2MB/s]
 13%|#3        | 46.9M/360M [00:00<00:06, 51.9MB/s]
 14%|#4        | 52.0M/360M [00:01<00:07, 42.8MB/s]
 18%|#7        | 64.0M/360M [00:01<00:06, 50.6MB/s]
 22%|##2       | 80.0M/360M [00:01<00:04, 59.0MB/s]
 27%|##6       | 96.0M/360M [00:01<00:04, 66.3MB/s]
 31%|###1      | 112M/360M [00:01<00:03, 69.0MB/s]
 35%|###5      | 127M/360M [00:02<00:03, 77.1MB/s]
 37%|###7      | 135M/360M [00:02<00:03, 65.5MB/s]
 40%|###9      | 144M/360M [00:02<00:03, 58.2MB/s]
 42%|####1     | 150M/360M [00:02<00:03, 56.5MB/s]
 43%|####3     | 155M/360M [00:03<00:05, 35.9MB/s]
 44%|####4     | 160M/360M [00:03<00:06, 32.4MB/s]
 49%|####8     | 175M/360M [00:03<00:04, 40.8MB/s]
 50%|####9     | 179M/360M [00:03<00:04, 38.5MB/s]
 53%|#####3    | 192M/360M [00:03<00:03, 52.3MB/s]
 58%|#####7    | 208M/360M [00:04<00:03, 51.5MB/s]
 62%|######1   | 223M/360M [00:04<00:02, 64.2MB/s]
 64%|######3   | 230M/360M [00:04<00:02, 45.7MB/s]
 67%|######6   | 240M/360M [00:05<00:04, 29.1MB/s]
 68%|######7   | 244M/360M [00:05<00:04, 29.3MB/s]
 69%|######9   | 249M/360M [00:05<00:03, 31.7MB/s]
 71%|#######1  | 256M/360M [00:05<00:03, 33.7MB/s]
 72%|#######2  | 260M/360M [00:06<00:03, 27.9MB/s]
 76%|#######5  | 272M/360M [00:06<00:02, 38.5MB/s]
 80%|#######9  | 287M/360M [00:06<00:01, 55.5MB/s]
 82%|########1 | 294M/360M [00:06<00:01, 44.7MB/s]
 84%|########4 | 303M/360M [00:06<00:01, 47.1MB/s]
 86%|########5 | 308M/360M [00:06<00:01, 44.7MB/s]
 89%|########8 | 319M/360M [00:07<00:00, 49.0MB/s]
 90%|########9 | 324M/360M [00:07<00:00, 48.0MB/s]
 93%|#########2| 335M/360M [00:07<00:00, 60.7MB/s]
 95%|#########4| 341M/360M [00:07<00:00, 41.5MB/s]
 97%|#########7| 351M/360M [00:08<00:00, 34.0MB/s]
 99%|#########8| 355M/360M [00:08<00:00, 31.6MB/s]
100%|##########| 360M/360M [00:08<00:00, 45.7MB/s]

我们将从 LibriSpeech test-other 数据集中加载一个样本。

speech_file = download_asset("tutorial-assets/ctc-decoding/1688-142285-0007.wav")

IPython.display.Audio(speech_file)

与此音频文件对应的转录文本为

i really was very much afraid of showing him how much shocked i was at some parts of what he said

waveform, sample_rate = torchaudio.load(speech_file)

if sample_rate != bundle.sample_rate:
    waveform = torchaudio.functional.resample(waveform, sample_rate, bundle.sample_rate)

解码器的文件和数据¶

接下来，我们加载标记、词典和语言模型数据，解码器使用这些数据从声学模型输出预测单词。LibriSpeech 数据集的预训练文件可以通过 torchaudio 下载，或者用户可以提供自己的文件。

标记¶

标记是声学模型可能预测的符号，包括空白和静音符号。它可以作为文件传入，其中每一行包含对应于相同索引的标记，或者作为标记列表传入，每个标记对应唯一的索引。

# tokens.txt
_
|
e
t
...

tokens = [label.lower() for label in bundle.get_labels()]
print(tokens)

['-', '|', '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']

词典¶

词典是单词与其对应标记序列之间的映射，用于将解码器的搜索空间限制在词典中的单词。词典文件的预期格式是每行一个单词，单词后面是空格分隔的标记。

# lexcion.txt
a a |
able a b l e |
about a b o u t |
...
...

语言模型¶

解码中可以使用语言模型来改善结果，方法是将表示序列可能性的语言模型得分纳入束搜索计算中。下面，我们概述了解码器支持的不同形式的语言模型。

无语言模型¶

要创建没有语言模型的解码器实例，在初始化解码器时将 lm=None 设置为解码器。

KenLM¶

这是使用 KenLM 库训练的 N 元语言模型。可以使用 .arpa 或二进制 .bin LM，但建议使用二进制格式以加快加载速度。

本教程中使用的语言模型是使用 LibriSpeech 训练的 4 元 KenLM。

自定义语言模型¶

用户可以使用 Python 定义自己的自定义语言模型，无论是统计语言模型还是神经网络语言模型，可以使用 CTCDecoderLM 和 CTCDecoderLMState。

例如，以下代码创建了 PyTorch torch.nn.Module 语言模型的基本包装器。

from torchaudio.models.decoder import CTCDecoderLM, CTCDecoderLMState


class CustomLM(CTCDecoderLM):
    """Create a Python wrapper around `language_model` to feed to the decoder."""

    def __init__(self, language_model: torch.nn.Module):
        CTCDecoderLM.__init__(self)
        self.language_model = language_model
        self.sil = -1  # index for silent token in the language model
        self.states = {}

        language_model.eval()

    def start(self, start_with_nothing: bool = False):
        state = CTCDecoderLMState()
        with torch.no_grad():
            score = self.language_model(self.sil)

        self.states[state] = score
        return state

    def score(self, state: CTCDecoderLMState, token_index: int):
        outstate = state.child(token_index)
        if outstate not in self.states:
            score = self.language_model(token_index)
            self.states[outstate] = score
        score = self.states[outstate]

        return outstate, score

    def finish(self, state: CTCDecoderLMState):
        return self.score(state, self.sil)

下载预训练文件¶

可以使用 download_pretrained_files() 下载 LibriSpeech 数据集的预训练文件。

注意：此单元格可能需要几分钟才能运行，因为语言模型可能很大

from torchaudio.models.decoder import download_pretrained_files

files = download_pretrained_files("librispeech-4-gram")

print(files)

  0%|          | 0.00/4.97M [00:00<?, ?B/s]
 85%|########5 | 4.25M/4.97M [00:00<00:00, 19.0MB/s]
100%|##########| 4.97M/4.97M [00:00<00:00, 21.7MB/s]

  0%|          | 0.00/57.0 [00:00<?, ?B/s]
100%|##########| 57.0/57.0 [00:00<00:00, 119kB/s]

  0%|          | 0.00/2.91G [00:00<?, ?B/s]
  0%|          | 14.9M/2.91G [00:00<01:16, 40.6MB/s]
  1%|          | 18.8M/2.91G [00:00<01:30, 34.4MB/s]
  1%|          | 24.5M/2.91G [00:00<01:27, 35.4MB/s]
  1%|1         | 30.9M/2.91G [00:00<01:13, 41.9MB/s]
  1%|1         | 35.2M/2.91G [00:01<01:30, 34.0MB/s]
  2%|1         | 48.0M/2.91G [00:01<01:09, 44.0MB/s]
  2%|2         | 62.9M/2.91G [00:01<01:15, 40.6MB/s]
  2%|2         | 67.0M/2.91G [00:01<01:25, 35.6MB/s]
  3%|2         | 80.0M/2.91G [00:02<01:12, 42.0MB/s]
  3%|3         | 94.9M/2.91G [00:02<01:14, 40.7MB/s]
  3%|3         | 98.9M/2.91G [00:02<01:15, 40.1MB/s]
  4%|3         | 112M/2.91G [00:02<00:58, 51.2MB/s]
  4%|4         | 128M/2.91G [00:02<00:48, 62.0MB/s]
  5%|4         | 144M/2.91G [00:03<00:42, 70.0MB/s]
  5%|5         | 151M/2.91G [00:03<00:49, 60.5MB/s]
  5%|5         | 159M/2.91G [00:03<00:52, 56.5MB/s]
  6%|5         | 166M/2.91G [00:03<00:48, 60.4MB/s]
  6%|5         | 176M/2.91G [00:03<00:43, 67.7MB/s]
  6%|6         | 191M/2.91G [00:03<00:33, 86.5MB/s]
  7%|6         | 200M/2.91G [00:03<00:35, 81.4MB/s]
  7%|6         | 208M/2.91G [00:04<00:48, 59.5MB/s]
  8%|7         | 224M/2.91G [00:04<00:36, 78.4MB/s]
  8%|7         | 233M/2.91G [00:04<00:38, 75.7MB/s]
  8%|8         | 241M/2.91G [00:04<00:50, 56.7MB/s]
  9%|8         | 256M/2.91G [00:04<00:45, 63.2MB/s]
  9%|9         | 272M/2.91G [00:05<00:37, 75.0MB/s]
  9%|9         | 281M/2.91G [00:05<00:37, 75.7MB/s]
 10%|9         | 289M/2.91G [00:05<00:40, 69.6MB/s]
 10%|#         | 299M/2.91G [00:05<00:44, 63.6MB/s]
 10%|#         | 306M/2.91G [00:05<00:45, 61.4MB/s]
 11%|#         | 319M/2.91G [00:05<00:40, 69.7MB/s]
 11%|#         | 326M/2.91G [00:06<00:46, 59.8MB/s]
 11%|#1        | 336M/2.91G [00:06<00:50, 54.7MB/s]
 11%|#1        | 342M/2.91G [00:06<00:51, 53.4MB/s]
 12%|#1        | 352M/2.91G [00:06<00:51, 53.2MB/s]
 12%|#1        | 357M/2.91G [00:06<00:54, 50.4MB/s]
 12%|#2        | 367M/2.91G [00:06<00:50, 54.0MB/s]
 12%|#2        | 372M/2.91G [00:07<00:58, 46.4MB/s]
 13%|#2        | 383M/2.91G [00:07<00:55, 49.2MB/s]
 13%|#3        | 388M/2.91G [00:07<01:01, 44.0MB/s]
 13%|#3        | 400M/2.91G [00:07<00:51, 52.2MB/s]
 14%|#3        | 416M/2.91G [00:07<00:40, 66.9MB/s]
 14%|#4        | 432M/2.91G [00:08<00:40, 65.5MB/s]
 15%|#4        | 438M/2.91G [00:08<00:43, 61.3MB/s]
 15%|#5        | 448M/2.91G [00:08<00:41, 64.7MB/s]
 16%|#5        | 464M/2.91G [00:08<00:33, 79.0MB/s]
 16%|#5        | 472M/2.91G [00:08<00:38, 68.3MB/s]
 16%|#6        | 480M/2.91G [00:08<00:40, 64.3MB/s]
 17%|#6        | 496M/2.91G [00:09<00:37, 69.1MB/s]
 17%|#6        | 502M/2.91G [00:09<00:47, 55.2MB/s]
 17%|#7        | 511M/2.91G [00:09<00:47, 54.1MB/s]
 17%|#7        | 516M/2.91G [00:09<01:01, 41.7MB/s]
 18%|#7        | 527M/2.91G [00:09<00:56, 45.3MB/s]
 18%|#7        | 532M/2.91G [00:10<00:58, 43.7MB/s]
 18%|#8        | 544M/2.91G [00:10<00:54, 47.0MB/s]
 18%|#8        | 551M/2.91G [00:10<00:50, 50.4MB/s]
 19%|#8        | 559M/2.91G [00:10<00:50, 50.6MB/s]
 19%|#8        | 564M/2.91G [00:10<00:54, 46.9MB/s]
 19%|#9        | 575M/2.91G [00:10<00:41, 60.9MB/s]
 20%|#9        | 582M/2.91G [00:10<00:45, 55.9MB/s]
 20%|#9        | 591M/2.91G [00:11<00:49, 50.2MB/s]
 20%|##        | 596M/2.91G [00:11<00:49, 50.1MB/s]
 20%|##        | 608M/2.91G [00:11<00:42, 58.5MB/s]
 21%|##        | 623M/2.91G [00:11<00:36, 67.7MB/s]
 21%|##1       | 629M/2.91G [00:11<00:41, 59.1MB/s]
 21%|##1       | 639M/2.91G [00:12<00:48, 50.8MB/s]
 22%|##1       | 644M/2.91G [00:12<00:54, 44.9MB/s]
 22%|##2       | 656M/2.91G [00:12<00:47, 51.4MB/s]
 22%|##2       | 661M/2.91G [00:12<00:54, 44.5MB/s]
 23%|##2       | 672M/2.91G [00:12<00:49, 49.0MB/s]
 23%|##2       | 676M/2.91G [00:13<00:59, 40.6MB/s]
 23%|##3       | 688M/2.91G [00:13<00:50, 47.3MB/s]
 23%|##3       | 692M/2.91G [00:13<01:00, 39.6MB/s]
 24%|##3       | 704M/2.91G [00:13<00:47, 49.9MB/s]
 24%|##4       | 720M/2.91G [00:13<00:36, 64.7MB/s]
 25%|##4       | 734M/2.91G [00:13<00:31, 74.9MB/s]
 25%|##4       | 741M/2.91G [00:14<00:42, 54.7MB/s]
 25%|##5       | 752M/2.91G [00:14<00:40, 57.8MB/s]
 26%|##5       | 768M/2.91G [00:14<00:31, 74.7MB/s]
 26%|##6       | 784M/2.91G [00:14<00:30, 75.5MB/s]
 27%|##6       | 792M/2.91G [00:14<00:33, 67.7MB/s]
 27%|##6       | 800M/2.91G [00:15<00:36, 62.0MB/s]
 27%|##7       | 815M/2.91G [00:15<00:33, 67.7MB/s]
 28%|##7       | 822M/2.91G [00:15<00:40, 56.0MB/s]
 28%|##7       | 831M/2.91G [00:15<00:36, 61.8MB/s]
 28%|##8       | 837M/2.91G [00:15<00:40, 55.1MB/s]
 28%|##8       | 848M/2.91G [00:15<00:35, 63.6MB/s]
 29%|##8       | 864M/2.91G [00:16<00:31, 70.2MB/s]
 29%|##9       | 873M/2.91G [00:16<00:29, 74.2MB/s]
 30%|##9       | 880M/2.91G [00:16<00:44, 49.8MB/s]
 30%|###       | 896M/2.91G [00:16<00:37, 58.2MB/s]
 31%|###       | 912M/2.91G [00:17<00:45, 48.1MB/s]
 31%|###       | 917M/2.91G [00:17<00:50, 43.1MB/s]
 31%|###1      | 928M/2.91G [00:17<00:43, 50.0MB/s]
 31%|###1      | 933M/2.91G [00:17<00:47, 45.6MB/s]
 32%|###1      | 944M/2.91G [00:17<00:40, 52.8MB/s]
 32%|###2      | 960M/2.91G [00:17<00:31, 66.9MB/s]
 32%|###2      | 967M/2.91G [00:18<00:39, 53.1MB/s]
 33%|###2      | 975M/2.91G [00:18<00:38, 55.3MB/s]
 33%|###2      | 981M/2.91G [00:18<00:40, 52.1MB/s]
 33%|###3      | 992M/2.91G [00:18<00:36, 57.1MB/s]
 34%|###3      | 0.98G/2.91G [00:18<00:26, 78.4MB/s]
 34%|###4      | 0.99G/2.91G [00:19<00:33, 61.4MB/s]
 34%|###4      | 1.00G/2.91G [00:19<00:33, 61.8MB/s]
 35%|###4      | 1.02G/2.91G [00:19<00:26, 77.3MB/s]
 35%|###5      | 1.03G/2.91G [00:19<00:21, 93.5MB/s]
 36%|###5      | 1.04G/2.91G [00:19<00:31, 62.9MB/s]
 36%|###6      | 1.05G/2.91G [00:19<00:37, 52.8MB/s]
 36%|###6      | 1.06G/2.91G [00:20<00:29, 67.7MB/s]
 37%|###6      | 1.07G/2.91G [00:20<00:46, 42.1MB/s]
 37%|###7      | 1.08G/2.91G [00:20<00:46, 42.6MB/s]
 38%|###7      | 1.09G/2.91G [00:21<00:41, 47.3MB/s]
 38%|###7      | 1.10G/2.91G [00:21<00:48, 40.1MB/s]
 38%|###7      | 1.10G/2.91G [00:21<00:48, 39.9MB/s]
 38%|###8      | 1.11G/2.91G [00:21<00:45, 42.4MB/s]
 38%|###8      | 1.12G/2.91G [00:21<00:35, 53.6MB/s]
 39%|###8      | 1.13G/2.91G [00:21<00:41, 45.8MB/s]
 39%|###9      | 1.14G/2.91G [00:22<00:31, 59.8MB/s]
 39%|###9      | 1.15G/2.91G [00:22<00:33, 56.2MB/s]
 40%|###9      | 1.16G/2.91G [00:22<00:35, 53.0MB/s]
 40%|####      | 1.17G/2.91G [00:22<00:28, 65.0MB/s]
 41%|####      | 1.19G/2.91G [00:22<00:23, 79.5MB/s]
 41%|####1     | 1.20G/2.91G [00:22<00:24, 75.7MB/s]
 42%|####1     | 1.22G/2.91G [00:23<00:23, 77.9MB/s]
 42%|####2     | 1.23G/2.91G [00:23<00:22, 81.1MB/s]
 43%|####2     | 1.25G/2.91G [00:23<00:20, 86.3MB/s]
 43%|####3     | 1.27G/2.91G [00:23<00:18, 96.4MB/s]
 44%|####3     | 1.27G/2.91G [00:23<00:20, 83.9MB/s]
 44%|####4     | 1.28G/2.91G [00:23<00:23, 74.3MB/s]
 45%|####4     | 1.30G/2.91G [00:24<00:20, 82.9MB/s]
 45%|####4     | 1.30G/2.91G [00:24<00:22, 76.6MB/s]
 45%|####5     | 1.31G/2.91G [00:24<00:28, 60.2MB/s]
 45%|####5     | 1.32G/2.91G [00:24<00:33, 51.6MB/s]
 46%|####5     | 1.33G/2.91G [00:24<00:38, 44.1MB/s]
 46%|####5     | 1.33G/2.91G [00:25<00:37, 44.8MB/s]
 46%|####6     | 1.34G/2.91G [00:25<00:33, 51.0MB/s]
 46%|####6     | 1.35G/2.91G [00:25<00:39, 42.7MB/s]
 47%|####6     | 1.36G/2.91G [00:25<00:31, 53.2MB/s]
 47%|####6     | 1.36G/2.91G [00:25<00:38, 43.3MB/s]
 47%|####7     | 1.38G/2.91G [00:25<00:32, 50.0MB/s]
 48%|####7     | 1.39G/2.91G [00:26<00:28, 56.6MB/s]
 48%|####8     | 1.41G/2.91G [00:26<00:23, 68.2MB/s]
 49%|####8     | 1.41G/2.91G [00:26<00:24, 64.7MB/s]
 49%|####8     | 1.42G/2.91G [00:26<00:23, 68.5MB/s]
 49%|####9     | 1.44G/2.91G [00:26<00:20, 76.3MB/s]
 50%|####9     | 1.45G/2.91G [00:27<00:20, 77.0MB/s]
 50%|#####     | 1.46G/2.91G [00:27<00:25, 61.3MB/s]
 50%|#####     | 1.47G/2.91G [00:27<00:25, 60.8MB/s]
 51%|#####1    | 1.48G/2.91G [00:27<00:20, 75.7MB/s]
 51%|#####1    | 1.49G/2.91G [00:27<00:27, 55.8MB/s]
 52%|#####1    | 1.50G/2.91G [00:28<00:29, 50.8MB/s]
 52%|#####2    | 1.51G/2.91G [00:28<00:30, 48.5MB/s]
 52%|#####2    | 1.52G/2.91G [00:28<00:28, 51.9MB/s]
 53%|#####2    | 1.53G/2.91G [00:28<00:24, 60.2MB/s]
 53%|#####3    | 1.55G/2.91G [00:28<00:26, 55.1MB/s]
 53%|#####3    | 1.55G/2.91G [00:29<00:31, 46.0MB/s]
 54%|#####3    | 1.56G/2.91G [00:29<00:33, 42.9MB/s]
 54%|#####3    | 1.57G/2.91G [00:29<00:34, 41.4MB/s]
 54%|#####4    | 1.58G/2.91G [00:29<00:31, 45.6MB/s]
 55%|#####4    | 1.59G/2.91G [00:30<00:24, 58.2MB/s]
 55%|#####5    | 1.61G/2.91G [00:30<00:19, 71.8MB/s]
 56%|#####5    | 1.62G/2.91G [00:30<00:21, 65.8MB/s]
 56%|#####5    | 1.62G/2.91G [00:30<00:21, 65.4MB/s]
 56%|#####6    | 1.64G/2.91G [00:30<00:22, 60.5MB/s]
 57%|#####6    | 1.65G/2.91G [00:31<00:39, 34.3MB/s]
 57%|#####6    | 1.65G/2.91G [00:31<00:39, 34.6MB/s]
 57%|#####6    | 1.66G/2.91G [00:31<00:44, 30.5MB/s]
 57%|#####7    | 1.67G/2.91G [00:31<00:31, 42.8MB/s]
 58%|#####7    | 1.68G/2.91G [00:32<00:36, 36.7MB/s]
 58%|#####7    | 1.69G/2.91G [00:32<00:28, 46.0MB/s]
 58%|#####8    | 1.69G/2.91G [00:32<00:27, 47.0MB/s]
 58%|#####8    | 1.70G/2.91G [00:32<00:25, 50.9MB/s]
 59%|#####8    | 1.70G/2.91G [00:32<00:25, 50.6MB/s]
 59%|#####9    | 1.72G/2.91G [00:32<00:17, 75.1MB/s]
 59%|#####9    | 1.73G/2.91G [00:32<00:18, 69.5MB/s]
 60%|#####9    | 1.73G/2.91G [00:33<00:24, 51.7MB/s]
 60%|#####9    | 1.74G/2.91G [00:33<00:26, 48.0MB/s]
 60%|#####9    | 1.75G/2.91G [00:33<00:40, 30.8MB/s]
 60%|######    | 1.75G/2.91G [00:33<00:43, 28.7MB/s]
 61%|######    | 1.76G/2.91G [00:34<00:27, 44.6MB/s]
 61%|######    | 1.77G/2.91G [00:34<00:31, 38.5MB/s]
 61%|######1   | 1.78G/2.91G [00:34<00:21, 56.6MB/s]
 62%|######1   | 1.80G/2.91G [00:34<00:17, 69.0MB/s]
 62%|######2   | 1.80G/2.91G [00:34<00:17, 68.2MB/s]
 62%|######2   | 1.81G/2.91G [00:34<00:16, 70.2MB/s]
 63%|######2   | 1.82G/2.91G [00:34<00:17, 65.3MB/s]
 63%|######2   | 1.83G/2.91G [00:35<00:20, 56.0MB/s]
 63%|######2   | 1.83G/2.91G [00:35<00:23, 49.1MB/s]
 63%|######3   | 1.84G/2.91G [00:35<00:20, 56.1MB/s]
 64%|######3   | 1.86G/2.91G [00:35<00:15, 71.1MB/s]
 64%|######4   | 1.87G/2.91G [00:35<00:20, 55.1MB/s]
 64%|######4   | 1.87G/2.91G [00:36<00:23, 47.8MB/s]
 65%|######4   | 1.88G/2.91G [00:36<00:29, 37.1MB/s]
 65%|######4   | 1.89G/2.91G [00:36<00:22, 47.8MB/s]
 65%|######5   | 1.90G/2.91G [00:36<00:24, 45.3MB/s]
 65%|######5   | 1.91G/2.91G [00:36<00:25, 43.1MB/s]
 66%|######5   | 1.91G/2.91G [00:37<00:28, 38.3MB/s]
 66%|######6   | 1.92G/2.91G [00:37<00:23, 45.6MB/s]
 67%|######6   | 1.94G/2.91G [00:37<00:17, 60.9MB/s]
 67%|######6   | 1.94G/2.91G [00:37<00:18, 55.5MB/s]
 67%|######7   | 1.95G/2.91G [00:37<00:21, 48.4MB/s]
 68%|######7   | 1.97G/2.91G [00:38<00:18, 55.8MB/s]
 68%|######7   | 1.97G/2.91G [00:38<00:19, 52.7MB/s]
 68%|######8   | 1.98G/2.91G [00:38<00:18, 55.3MB/s]
 68%|######8   | 1.99G/2.91G [00:38<00:16, 60.1MB/s]
 69%|######8   | 2.00G/2.91G [00:38<00:15, 62.1MB/s]
 69%|######8   | 2.00G/2.91G [00:38<00:23, 42.2MB/s]
 69%|######9   | 2.01G/2.91G [00:39<00:22, 42.8MB/s]
 69%|######9   | 2.02G/2.91G [00:39<00:22, 43.3MB/s]
 70%|######9   | 2.03G/2.91G [00:39<00:16, 58.1MB/s]
 70%|#######   | 2.04G/2.91G [00:39<00:17, 54.2MB/s]
 70%|#######   | 2.05G/2.91G [00:39<00:17, 51.9MB/s]
 70%|#######   | 2.05G/2.91G [00:40<00:23, 38.7MB/s]
 71%|#######   | 2.06G/2.91G [00:40<00:18, 48.7MB/s]
 71%|#######1  | 2.08G/2.91G [00:40<00:13, 64.8MB/s]
 72%|#######1  | 2.08G/2.91G [00:40<00:15, 57.9MB/s]
 72%|#######1  | 2.09G/2.91G [00:40<00:13, 63.3MB/s]
 72%|#######2  | 2.10G/2.91G [00:40<00:14, 59.4MB/s]
 72%|#######2  | 2.11G/2.91G [00:40<00:12, 67.6MB/s]
 73%|#######2  | 2.12G/2.91G [00:41<00:13, 64.5MB/s]
 73%|#######3  | 2.12G/2.91G [00:41<00:17, 48.1MB/s]
 74%|#######3  | 2.14G/2.91G [00:41<00:14, 56.6MB/s]
 74%|#######3  | 2.15G/2.91G [00:41<00:16, 49.7MB/s]
 74%|#######4  | 2.16G/2.91G [00:41<00:16, 48.8MB/s]
 74%|#######4  | 2.16G/2.91G [00:42<00:16, 49.1MB/s]
 74%|#######4  | 2.17G/2.91G [00:42<00:16, 48.1MB/s]
 75%|#######4  | 2.17G/2.91G [00:42<00:15, 49.7MB/s]
 75%|#######4  | 2.18G/2.91G [00:42<00:15, 51.2MB/s]
 75%|#######5  | 2.18G/2.91G [00:42<00:17, 45.7MB/s]
 75%|#######5  | 2.19G/2.91G [00:42<00:17, 43.2MB/s]
 75%|#######5  | 2.19G/2.91G [00:43<00:25, 30.7MB/s]
 76%|#######5  | 2.20G/2.91G [00:43<00:20, 36.8MB/s]
 76%|#######5  | 2.21G/2.91G [00:43<00:15, 49.0MB/s]
 76%|#######6  | 2.22G/2.91G [00:43<00:12, 59.0MB/s]
 77%|#######6  | 2.23G/2.91G [00:43<00:11, 64.9MB/s]
 77%|#######6  | 2.24G/2.91G [00:43<00:13, 55.1MB/s]
 77%|#######7  | 2.25G/2.91G [00:44<00:14, 50.3MB/s]
 77%|#######7  | 2.25G/2.91G [00:44<00:14, 47.5MB/s]
 78%|#######7  | 2.26G/2.91G [00:44<00:12, 54.0MB/s]
 78%|#######7  | 2.27G/2.91G [00:44<00:15, 45.0MB/s]
 78%|#######8  | 2.28G/2.91G [00:44<00:14, 47.1MB/s]
 79%|#######8  | 2.30G/2.91G [00:44<00:11, 57.3MB/s]
 79%|#######9  | 2.30G/2.91G [00:45<00:11, 55.4MB/s]
 79%|#######9  | 2.31G/2.91G [00:45<00:09, 67.2MB/s]
 80%|#######9  | 2.32G/2.91G [00:45<00:10, 58.1MB/s]
 80%|#######9  | 2.33G/2.91G [00:45<00:11, 52.3MB/s]
 80%|########  | 2.33G/2.91G [00:45<00:13, 46.2MB/s]
 81%|########  | 2.34G/2.91G [00:45<00:12, 48.2MB/s]
 81%|########  | 2.35G/2.91G [00:46<00:15, 37.8MB/s]
 81%|########  | 2.35G/2.91G [00:46<00:15, 38.4MB/s]
 81%|########1 | 2.36G/2.91G [00:46<00:15, 37.1MB/s]
 82%|########1 | 2.38G/2.91G [00:46<00:10, 53.5MB/s]
 82%|########2 | 2.39G/2.91G [00:46<00:08, 63.6MB/s]
 82%|########2 | 2.40G/2.91G [00:47<00:08, 64.3MB/s]
 83%|########2 | 2.41G/2.91G [00:47<00:07, 68.9MB/s]
 83%|########2 | 2.41G/2.91G [00:47<00:07, 69.4MB/s]
 83%|########3 | 2.42G/2.91G [00:47<00:08, 59.7MB/s]
 84%|########3 | 2.44G/2.91G [00:47<00:07, 64.6MB/s]
 84%|########3 | 2.44G/2.91G [00:47<00:08, 61.6MB/s]
 84%|########4 | 2.45G/2.91G [00:48<00:08, 59.4MB/s]
 85%|########4 | 2.47G/2.91G [00:48<00:07, 67.5MB/s]
 85%|########5 | 2.47G/2.91G [00:48<00:11, 41.7MB/s]
 85%|########5 | 2.48G/2.91G [00:48<00:11, 40.9MB/s]
 86%|########5 | 2.50G/2.91G [00:49<00:07, 56.8MB/s]
 86%|########6 | 2.51G/2.91G [00:49<00:08, 52.7MB/s]
 86%|########6 | 2.52G/2.91G [00:49<00:07, 54.6MB/s]
 87%|########6 | 2.53G/2.91G [00:49<00:06, 65.9MB/s]
 87%|########7 | 2.55G/2.91G [00:49<00:05, 69.2MB/s]
 88%|########7 | 2.55G/2.91G [00:49<00:06, 63.1MB/s]
 88%|########8 | 2.56G/2.91G [00:50<00:06, 60.5MB/s]
 89%|########8 | 2.58G/2.91G [00:50<00:05, 66.9MB/s]
 89%|########8 | 2.58G/2.91G [00:50<00:05, 60.9MB/s]
 89%|########9 | 2.59G/2.91G [00:50<00:05, 68.1MB/s]
 89%|########9 | 2.60G/2.91G [00:50<00:05, 55.9MB/s]
 90%|########9 | 2.61G/2.91G [00:50<00:06, 53.7MB/s]
 90%|########9 | 2.61G/2.91G [00:51<00:07, 43.6MB/s]
 90%|######### | 2.62G/2.91G [00:51<00:05, 54.1MB/s]
 90%|######### | 2.63G/2.91G [00:51<00:05, 50.8MB/s]
 91%|######### | 2.64G/2.91G [00:51<00:05, 50.7MB/s]
 91%|#########1| 2.66G/2.91G [00:51<00:04, 56.5MB/s]
 91%|#########1| 2.66G/2.91G [00:52<00:06, 44.0MB/s]
 92%|#########1| 2.67G/2.91G [00:52<00:07, 35.0MB/s]
 92%|#########1| 2.67G/2.91G [00:52<00:07, 36.2MB/s]
 92%|#########2| 2.69G/2.91G [00:52<00:04, 50.6MB/s]
 93%|#########2| 2.69G/2.91G [00:53<00:05, 42.8MB/s]
 93%|#########2| 2.70G/2.91G [00:53<00:05, 44.7MB/s]
 93%|#########3| 2.71G/2.91G [00:53<00:04, 44.4MB/s]
 93%|#########3| 2.72G/2.91G [00:53<00:03, 52.2MB/s]
 94%|#########3| 2.72G/2.91G [00:53<00:05, 40.0MB/s]
 94%|#########3| 2.73G/2.91G [00:53<00:04, 43.3MB/s]
 94%|#########3| 2.73G/2.91G [00:54<00:04, 41.5MB/s]
 94%|#########4| 2.74G/2.91G [00:54<00:05, 35.1MB/s]
 94%|#########4| 2.75G/2.91G [00:54<00:04, 42.8MB/s]
 95%|#########4| 2.75G/2.91G [00:54<00:04, 36.1MB/s]
 95%|#########5| 2.77G/2.91G [00:54<00:03, 45.5MB/s]
 95%|#########5| 2.77G/2.91G [00:54<00:03, 43.9MB/s]
 96%|#########5| 2.78G/2.91G [00:55<00:02, 51.5MB/s]
 96%|#########6| 2.80G/2.91G [00:55<00:01, 62.6MB/s]
 96%|#########6| 2.80G/2.91G [00:55<00:02, 54.6MB/s]
 96%|#########6| 2.81G/2.91G [00:55<00:02, 54.6MB/s]
 97%|#########6| 2.81G/2.91G [00:55<00:02, 44.1MB/s]
 97%|#########6| 2.82G/2.91G [00:56<00:02, 37.1MB/s]
 97%|#########7| 2.83G/2.91G [00:56<00:01, 52.0MB/s]
 97%|#########7| 2.83G/2.91G [00:56<00:02, 29.4MB/s]
 98%|#########7| 2.84G/2.91G [00:56<00:02, 32.1MB/s]
 98%|#########7| 2.84G/2.91G [00:56<00:02, 31.3MB/s]
 98%|#########7| 2.85G/2.91G [00:57<00:02, 24.9MB/s]
 98%|#########8| 2.86G/2.91G [00:57<00:02, 25.1MB/s]
 98%|#########8| 2.86G/2.91G [00:57<00:02, 21.4MB/s]
 99%|#########8| 2.88G/2.91G [00:58<00:01, 32.0MB/s]
 99%|#########9| 2.89G/2.91G [00:58<00:00, 48.8MB/s]
100%|#########9| 2.90G/2.91G [00:58<00:00, 48.1MB/s]
100%|#########9| 2.91G/2.91G [00:58<00:00, 51.0MB/s]
100%|##########| 2.91G/2.91G [00:58<00:00, 53.2MB/s]
PretrainedFiles(lexicon='/root/.cache/torch/hub/torchaudio/decoder-assets/librispeech-4-gram/lexicon.txt', tokens='/root/.cache/torch/hub/torchaudio/decoder-assets/librispeech-4-gram/tokens.txt', lm='/root/.cache/torch/hub/torchaudio/decoder-assets/librispeech-4-gram/lm.bin')

构建解码器¶

在本教程中，我们构建了束搜索解码器和贪婪解码器以进行比较。

束搜索解码器¶

可以使用工厂函数 ctc_decoder() 构建解码器。除了前面提到的组件之外，它还接收各种束搜索解码参数和标记/单词参数。

此解码器也可以在没有语言模型的情况下运行，方法是将 None 传递到 lm 参数中。

LM_WEIGHT = 3.23
WORD_SCORE = -0.26

beam_search_decoder = ctc_decoder(
    lexicon=files.lexicon,
    tokens=files.tokens,
    lm=files.lm,
    nbest=3,
    beam_size=1500,
    lm_weight=LM_WEIGHT,
    word_score=WORD_SCORE,
)

贪婪解码器¶

class GreedyCTCDecoder(torch.nn.Module):
    def __init__(self, labels, blank=0):
        super().__init__()
        self.labels = labels
        self.blank = blank

    def forward(self, emission: torch.Tensor) -> List[str]:
        """Given a sequence emission over labels, get the best path
        Args:
          emission (Tensor): Logit tensors. Shape `[num_seq, num_label]`.

        Returns:
          List[str]: The resulting transcript
        """
        indices = torch.argmax(emission, dim=-1)  # [num_seq,]
        indices = torch.unique_consecutive(indices, dim=-1)
        indices = [i for i in indices if i != self.blank]
        joined = "".join([self.labels[i] for i in indices])
        return joined.replace("|", " ").strip().split()


greedy_decoder = GreedyCTCDecoder(tokens)

运行推理¶

现在我们有了数据、声学模型和解码器，就可以执行推理了。波束搜索解码器的输出类型为CTCHypothesis，包含预测的标记 ID、相应的单词（如果提供了词典）、假设分数以及与标记 ID 对应的时步。回想一下与波形对应的转录文本是

i really was very much afraid of showing him how much shocked i was at some parts of what he said

actual_transcript = "i really was very much afraid of showing him how much shocked i was at some parts of what he said"
actual_transcript = actual_transcript.split()

emission, _ = acoustic_model(waveform)

贪婪解码器给出以下结果。

greedy_result = greedy_decoder(emission[0])
greedy_transcript = " ".join(greedy_result)
greedy_wer = torchaudio.functional.edit_distance(actual_transcript, greedy_result) / len(actual_transcript)

print(f"Transcript: {greedy_transcript}")
print(f"WER: {greedy_wer}")

Transcript: i reily was very much affrayd of showing him howmuch shoktd i wause at some parte of what he seid
WER: 0.38095238095238093

使用波束搜索解码器

beam_search_result = beam_search_decoder(emission)
beam_search_transcript = " ".join(beam_search_result[0][0].words).strip()
beam_search_wer = torchaudio.functional.edit_distance(actual_transcript, beam_search_result[0][0].words) / len(
    actual_transcript
)

print(f"Transcript: {beam_search_transcript}")
print(f"WER: {beam_search_wer}")

Transcript: i really was very much afraid of showing him how much shocked i was at some part of what he said
WER: 0.047619047619047616

注意

如果未向解码器提供词典，则输出假设的words字段将为空。要检索无词典解码的转录文本，您可以执行以下操作以检索标记索引，将其转换为原始标记，然后将它们连接在一起。

tokens_str = "".join(beam_search_decoder.idxs_to_tokens(beam_search_result[0][0].tokens))
transcript = " ".join(tokens_str.split("|"))

我们看到，使用词典约束的波束搜索解码器得到的转录文本产生了更准确的结果，包含真实的单词，而贪婪解码器可能会预测拼写错误的单词，例如“affrayd”和“shoktd”。

增量解码¶

如果输入语音较长，可以增量方式解码发射。

您需要首先使用decode_begin()初始化解码器的内部状态。

beam_search_decoder.decode_begin()

然后，您可以将发射传递给decode_begin()。这里我们使用相同的发射，但一次一帧地将其传递给解码器。

for t in range(emission.size(1)):
    beam_search_decoder.decode_step(emission[0, t:t + 1, :])

最后，完成解码器的内部状态，并检索结果。

beam_search_decoder.decode_end()
beam_search_result_inc = beam_search_decoder.get_final_hypothesis()

增量解码的结果与批量解码相同。

beam_search_transcript_inc = " ".join(beam_search_result_inc[0].words).strip()
beam_search_wer_inc = torchaudio.functional.edit_distance(
    actual_transcript, beam_search_result_inc[0].words) / len(actual_transcript)

print(f"Transcript: {beam_search_transcript_inc}")
print(f"WER: {beam_search_wer_inc}")

assert beam_search_result[0][0].words == beam_search_result_inc[0].words
assert beam_search_result[0][0].score == beam_search_result_inc[0].score
torch.testing.assert_close(beam_search_result[0][0].timesteps, beam_search_result_inc[0].timesteps)

Transcript: i really was very much afraid of showing him how much shocked i was at some part of what he said
WER: 0.047619047619047616

时步对齐¶

回想一下，结果假设的组成部分之一是与标记 ID 对应的时步。

timesteps = beam_search_result[0][0].timesteps
predicted_tokens = beam_search_decoder.idxs_to_tokens(beam_search_result[0][0].tokens)

print(predicted_tokens, len(predicted_tokens))
print(timesteps, timesteps.shape[0])

['|', 'i', '|', 'r', 'e', 'a', 'l', 'l', 'y', '|', 'w', 'a', 's', '|', 'v', 'e', 'r', 'y', '|', 'm', 'u', 'c', 'h', '|', 'a', 'f', 'r', 'a', 'i', 'd', '|', 'o', 'f', '|', 's', 'h', 'o', 'w', 'i', 'n', 'g', '|', 'h', 'i', 'm', '|', 'h', 'o', 'w', '|', 'm', 'u', 'c', 'h', '|', 's', 'h', 'o', 'c', 'k', 'e', 'd', '|', 'i', '|', 'w', 'a', 's', '|', 'a', 't', '|', 's', 'o', 'm', 'e', '|', 'p', 'a', 'r', 't', '|', 'o', 'f', '|', 'w', 'h', 'a', 't', '|', 'h', 'e', '|', 's', 'a', 'i', 'd', '|', '|'] 99
tensor([  0,  31,  33,  36,  39,  41,  42,  44,  46,  48,  49,  52,  54,  58,
         64,  66,  69,  73,  74,  76,  80,  82,  84,  86,  88,  94,  97, 107,
        111, 112, 116, 134, 136, 138, 140, 142, 146, 148, 151, 153, 155, 157,
        159, 161, 162, 166, 170, 176, 177, 178, 179, 182, 184, 186, 187, 191,
        193, 198, 201, 202, 203, 205, 207, 212, 213, 216, 222, 224, 230, 250,
        251, 254, 256, 261, 262, 264, 267, 270, 276, 277, 281, 284, 288, 289,
        292, 295, 297, 299, 300, 303, 305, 307, 310, 311, 324, 325, 329, 331,
        353], dtype=torch.int32) 99

下面，我们将可视化相对于原始波形的标记时步对齐。

def plot_alignments(waveform, emission, tokens, timesteps, sample_rate):

    t = torch.arange(waveform.size(0)) / sample_rate
    ratio = waveform.size(0) / emission.size(1) / sample_rate

    chars = []
    words = []
    word_start = None
    for token, timestep in zip(tokens, timesteps * ratio):
        if token == "|":
            if word_start is not None:
                words.append((word_start, timestep))
            word_start = None
        else:
            chars.append((token, timestep))
            if word_start is None:
                word_start = timestep

    fig, axes = plt.subplots(3, 1)

    def _plot(ax, xlim):
        ax.plot(t, waveform)
        for token, timestep in chars:
            ax.annotate(token.upper(), (timestep, 0.5))
        for word_start, word_end in words:
            ax.axvspan(word_start, word_end, alpha=0.1, color="red")
        ax.set_ylim(-0.6, 0.7)
        ax.set_yticks([0])
        ax.grid(True, axis="y")
        ax.set_xlim(xlim)

    _plot(axes[0], (0.3, 2.5))
    _plot(axes[1], (2.5, 4.7))
    _plot(axes[2], (4.7, 6.9))
    axes[2].set_xlabel("time (sec)")
    fig.tight_layout()


plot_alignments(waveform[0], emission, predicted_tokens, timesteps, bundle.sample_rate)

波束搜索解码器参数¶

在本节中，我们将更深入地了解一些不同的参数和权衡。有关可自定义参数的完整列表，请参阅文档。

辅助函数¶

def print_decoded(decoder, emission, param, param_value):
    start_time = time.monotonic()
    result = decoder(emission)
    decode_time = time.monotonic() - start_time

    transcript = " ".join(result[0][0].words).lower().strip()
    score = result[0][0].score
    print(f"{param} {param_value:<3}: {transcript} (score: {score:.2f}; {decode_time:.4f} secs)")

nbest¶

此参数指示要返回的最佳假设的数量，这是贪婪解码器无法实现的特性。例如，通过在前面构建波束搜索解码器时设置nbest=3，我们现在可以访问具有前 3 个分数的假设。

for i in range(3):
    transcript = " ".join(beam_search_result[0][i].words).strip()
    score = beam_search_result[0][i].score
    print(f"{transcript} (score: {score})")

i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.824109642502)
i really was very much afraid of showing him how much shocked i was at some parts of what he said (score: 3697.858373688456)
i reply was very much afraid of showing him how much shocked i was at some part of what he said (score: 3695.0157600045172)

波束大小¶

beam_size参数确定在每个解码步骤后要保留的最大最佳假设数量。使用更大的波束大小可以探索更大范围的可能假设，从而产生得分更高的假设，但它在计算上更昂贵，并且在达到一定程度后不会带来额外收益。

在下面的示例中，我们看到解码质量随着我们将波束大小从 1 增加到 5 再到 50 而得到改善，但请注意，使用 500 的波束大小与使用 50 的波束大小产生的输出相同，同时增加了计算时间。

beam_sizes = [1, 5, 50, 500]

for beam_size in beam_sizes:
    beam_search_decoder = ctc_decoder(
        lexicon=files.lexicon,
        tokens=files.tokens,
        lm=files.lm,
        beam_size=beam_size,
        lm_weight=LM_WEIGHT,
        word_score=WORD_SCORE,
    )

    print_decoded(beam_search_decoder, emission, "beam size", beam_size)

beam size 1  : i you ery much afra of shongut shot i was at some arte what he sad (score: 3144.93; 0.0444 secs)
beam size 5  : i rely was very much afraid of showing him how much shot i was at some parts of what he said (score: 3688.02; 0.0496 secs)
beam size 50 : i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.1621 secs)
beam size 500: i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.5372 secs)

波束大小标记¶

beam_size_token参数对应于在解码步骤中考虑扩展每个假设的标记数量。探索更多可能的下一个标记会增加潜在假设的范围，但代价是计算量增加。

num_tokens = len(tokens)
beam_size_tokens = [1, 5, 10, num_tokens]

for beam_size_token in beam_size_tokens:
    beam_search_decoder = ctc_decoder(
        lexicon=files.lexicon,
        tokens=files.tokens,
        lm=files.lm,
        beam_size_token=beam_size_token,
        lm_weight=LM_WEIGHT,
        word_score=WORD_SCORE,
    )

    print_decoded(beam_search_decoder, emission, "beam size token", beam_size_token)

beam size token 1  : i rely was very much affray of showing him hoch shot i was at some part of what he sed (score: 3584.80; 0.1564 secs)
beam size token 5  : i rely was very much afraid of showing him how much shocked i was at some part of what he said (score: 3694.83; 0.1773 secs)
beam size token 10 : i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3696.25; 0.1958 secs)
beam size token 29 : i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.2272 secs)

波束阈值¶

beam_threshold参数用于在每个解码步骤修剪存储的假设集，删除得分比最高得分假设高beam_threshold的假设。在选择较小的阈值以修剪更多假设并减少搜索空间，以及选择足够大的阈值以确保不会修剪合理的假设之间存在平衡。

beam_thresholds = [1, 5, 10, 25]

for beam_threshold in beam_thresholds:
    beam_search_decoder = ctc_decoder(
        lexicon=files.lexicon,
        tokens=files.tokens,
        lm=files.lm,
        beam_threshold=beam_threshold,
        lm_weight=LM_WEIGHT,
        word_score=WORD_SCORE,
    )

    print_decoded(beam_search_decoder, emission, "beam threshold", beam_threshold)

beam threshold 1  : i ila ery much afraid of shongut shot i was at some parts of what he said (score: 3316.20; 0.0277 secs)
beam threshold 5  : i rely was very much afraid of showing him how much shot i was at some parts of what he said (score: 3682.23; 0.0602 secs)
beam threshold 10 : i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.2091 secs)
beam threshold 25 : i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.2304 secs)

语言模型权重¶

lm_weight参数是分配给语言模型分数的权重，用于与声学模型分数累加以确定总体分数。较大的权重鼓励模型根据语言模型预测下一个单词，而较小的权重则更多地依赖声学模型分数。

lm_weights = [0, LM_WEIGHT, 15]

for lm_weight in lm_weights:
    beam_search_decoder = ctc_decoder(
        lexicon=files.lexicon,
        tokens=files.tokens,
        lm=files.lm,
        lm_weight=lm_weight,
        word_score=WORD_SCORE,
    )

    print_decoded(beam_search_decoder, emission, "lm weight", lm_weight)

lm weight 0  : i rely was very much affraid of showing him ho much shoke i was at some parte of what he seid (score: 3834.05; 0.2525 secs)
lm weight 3.23: i really was very much afraid of showing him how much shocked i was at some part of what he said (score: 3699.82; 0.2589 secs)
lm weight 15 : was there in his was at some of what he said (score: 2918.99; 0.2394 secs)

其他参数¶

可以优化的其他参数包括：

word_score：单词结束时添加的分数
unk_score：添加的未知单词出现分数
sil_score：添加的静音出现分数
log_add：是否对词典 Trie 平滑使用对数加法

脚本的总运行时间：（2 分 57.274 秒）

由 Sphinx-Gallery 生成的图库