注意
点击 这里 下载完整的示例代码
使用 CTC 解码器的 ASR 推理¶
作者: Caroline Chen
本教程展示了如何使用带有词典约束和 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 模型上演示这一点。
概述¶
束搜索解码通过迭代地扩展文本假设（束）来进行，并使用下一个可能的字符，并且仅在每个时间步保持得分最高的假设。语言模型可以被纳入评分计算中，添加词典约束会限制假设的下一个可能标记，以便只能生成来自词典的单词。
底层实现是从 Flashlight 的束搜索解码器移植的。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 博客 中找到。
使用带有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：可能出现的单词与其对应的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，我们导入必要的工具并获取我们要处理的数据
import torch
import torchaudio
print(torch.__version__)
print(torchaudio.__version__)
2.3.0
2.3.0
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
我们使用预训练的 Wav2Vec 2.0 基础模型，该模型在 LibriSpeech 数据集 的 10 分钟数据上进行了微调，可以使用 torchaudio.pipelines.WAV2VEC2_ASR_BASE_10M 加载。有关在 torchaudio 中运行 Wav2Vec 2.0 语音识别管道的更多详细信息，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]
3%|3 | 11.4M/360M [00:00<00:03, 119MB/s]
6%|6 | 22.8M/360M [00:00<00:03, 93.2MB/s]
9%|8 | 32.0M/360M [00:00<00:06, 55.8MB/s]
13%|#3 | 48.0M/360M [00:00<00:05, 62.6MB/s]
17%|#7 | 62.9M/360M [00:00<00:04, 76.8MB/s]
20%|#9 | 71.2M/360M [00:01<00:04, 69.3MB/s]
22%|##2 | 79.8M/360M [00:01<00:04, 67.9MB/s]
24%|##4 | 86.6M/360M [00:01<00:05, 55.4MB/s]
27%|##6 | 96.0M/360M [00:01<00:04, 61.3MB/s]
31%|### | 111M/360M [00:01<00:04, 62.0MB/s]
33%|###2 | 117M/360M [00:02<00:05, 48.5MB/s]
35%|###5 | 127M/360M [00:02<00:04, 51.0MB/s]
37%|###6 | 133M/360M [00:02<00:05, 43.1MB/s]
40%|###9 | 144M/360M [00:02<00:04, 51.4MB/s]
41%|####1 | 149M/360M [00:02<00:04, 50.2MB/s]
43%|####2 | 154M/360M [00:02<00:04, 48.5MB/s]
44%|####4 | 160M/360M [00:03<00:05, 37.0MB/s]
49%|####8 | 175M/360M [00:03<00:04, 44.0MB/s]
50%|####9 | 179M/360M [00:03<00:04, 41.8MB/s]
53%|#####3 | 192M/360M [00:03<00:03, 50.7MB/s]
58%|#####7 | 208M/360M [00:03<00:02, 58.7MB/s]
59%|#####9 | 213M/360M [00:04<00:03, 50.9MB/s]
62%|######1 | 223M/360M [00:04<00:02, 57.2MB/s]
63%|######3 | 229M/360M [00:04<00:03, 42.2MB/s]
67%|######6 | 240M/360M [00:04<00:02, 43.2MB/s]
68%|######7 | 244M/360M [00:04<00:03, 38.6MB/s]
71%|#######1 | 256M/360M [00:05<00:02, 43.7MB/s]
76%|#######5 | 272M/360M [00:05<00:01, 52.6MB/s]
80%|#######9 | 288M/360M [00:05<00:01, 59.3MB/s]
84%|########4 | 303M/360M [00:05<00:00, 70.6MB/s]
86%|########6 | 310M/360M [00:06<00:00, 58.8MB/s]
89%|########8 | 319M/360M [00:06<00:00, 56.4MB/s]
90%|######### | 325M/360M [00:06<00:00, 42.5MB/s]
91%|#########1| 329M/360M [00:06<00:00, 38.9MB/s]
93%|#########3| 336M/360M [00:06<00:00, 38.7MB/s]
98%|#########7| 352M/360M [00:07<00:00, 50.7MB/s]
100%|##########| 360M/360M [00:07<00:00, 53.2MB/s]
我们将从 LibriSpeech test-other 数据集加载一个样本。
speech_file = download_asset("tutorial-assets/ctc-decoding/1688-142285-0007.wav")
IPython.display.Audio(speech_file)
与该音频文件相对应的转录文本为
waveform, sample_rate = torchaudio.load(speech_file)
if sample_rate != bundle.sample_rate:
waveform = torchaudio.functional.resample(waveform, sample_rate, bundle.sample_rate)
解码器文件和数据¶
接下来，我们加载标记、词典和语言模型数据，这些数据由解码器用来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¶
标记是声学模型可以预测的可能符号，包括空白符号和静音符号。它可以作为文件传递，其中每行包含与同一索引对应的标记，也可以作为标记列表传递，每个标记映射到一个唯一的索引。
# tokens.txt
_
|
e
t
...
['-', '|', '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的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，单词后面跟着用空格分隔的标记。
# lexcion.txt
a a |
able a b l e |
about a b o u t |
...
...
语言模型¶
语言模型可以在解码中使用，通过将表示序列可能性的语言模型分数纳入束搜索计算来改善结果。下面，我们概述了支持解码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¶
这是一个使用 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，因为它加载速度更快。
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 CTCDecoderLM 和 CTCDecoderLMState 在 Python 中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。
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：由于语言模型可能很大，此单元格可能需要几分钟才能运行。
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, 22.8MB/s]
100%|##########| 4.97M/4.97M [00:00<00:00, 26.1MB/s]
0%| | 0.00/57.0 [00:00<?, ?B/s]
100%|##########| 57.0/57.0 [00:00<00:00, 59.4kB/s]
0%| | 0.00/2.91G [00:00<?, ?B/s]
0%| | 14.9M/2.91G [00:00<00:32, 95.7MB/s]
1%| | 24.1M/2.91G [00:00<00:50, 61.4MB/s]
1%|1 | 32.0M/2.91G [00:00<01:00, 51.0MB/s]
2%|1 | 46.9M/2.91G [00:00<00:57, 53.2MB/s]
2%|1 | 52.1M/2.91G [00:01<01:00, 50.7MB/s]
2%|2 | 64.0M/2.91G [00:01<00:57, 53.0MB/s]
3%|2 | 79.8M/2.91G [00:01<00:41, 73.7MB/s]
3%|2 | 88.0M/2.91G [00:01<00:52, 58.2MB/s]
3%|3 | 94.9M/2.91G [00:01<00:53, 56.1MB/s]
3%|3 | 101M/2.91G [00:01<00:56, 53.5MB/s]
4%|3 | 112M/2.91G [00:02<00:57, 52.5MB/s]
4%|4 | 128M/2.91G [00:02<00:45, 66.4MB/s]
5%|4 | 135M/2.91G [00:02<00:48, 62.1MB/s]
5%|4 | 144M/2.91G [00:02<00:52, 56.2MB/s]
5%|5 | 159M/2.91G [00:03<01:03, 46.9MB/s]
5%|5 | 164M/2.91G [00:03<01:06, 44.7MB/s]
6%|5 | 175M/2.91G [00:03<00:55, 52.6MB/s]
6%|6 | 180M/2.91G [00:03<00:57, 51.2MB/s]
6%|6 | 192M/2.91G [00:03<00:51, 57.0MB/s]
7%|6 | 208M/2.91G [00:03<00:42, 68.6MB/s]
7%|7 | 214M/2.91G [00:03<00:48, 59.9MB/s]
8%|7 | 224M/2.91G [00:04<00:45, 62.9MB/s]
8%|7 | 230M/2.91G [00:04<00:50, 57.6MB/s]
8%|8 | 239M/2.91G [00:04<00:48, 59.3MB/s]
8%|8 | 245M/2.91G [00:04<00:56, 50.6MB/s]
9%|8 | 256M/2.91G [00:04<00:52, 54.4MB/s]
9%|9 | 271M/2.91G [00:04<00:42, 67.4MB/s]
9%|9 | 278M/2.91G [00:05<00:45, 62.1MB/s]
10%|9 | 284M/2.91G [00:05<00:50, 56.2MB/s]
10%|9 | 289M/2.91G [00:05<00:59, 47.3MB/s]
10%|# | 299M/2.91G [00:05<01:02, 44.9MB/s]
10%|# | 304M/2.91G [00:05<01:06, 42.2MB/s]
11%|# | 319M/2.91G [00:05<00:46, 59.8MB/s]
11%|# | 325M/2.91G [00:06<01:01, 45.2MB/s]
11%|#1 | 336M/2.91G [00:06<00:53, 51.9MB/s]
12%|#1 | 352M/2.91G [00:06<00:38, 71.2MB/s]
12%|#2 | 360M/2.91G [00:06<00:43, 62.8MB/s]
12%|#2 | 367M/2.91G [00:06<00:50, 54.3MB/s]
13%|#2 | 373M/2.91G [00:07<00:57, 47.1MB/s]
13%|#2 | 383M/2.91G [00:07<01:11, 38.1MB/s]
13%|#2 | 387M/2.91G [00:07<01:18, 34.5MB/s]
13%|#3 | 400M/2.91G [00:07<00:58, 46.4MB/s]
14%|#3 | 415M/2.91G [00:07<00:41, 64.6MB/s]
14%|#4 | 423M/2.91G [00:08<00:44, 59.9MB/s]
14%|#4 | 432M/2.91G [00:08<00:42, 62.6MB/s]
15%|#4 | 439M/2.91G [00:08<00:48, 54.7MB/s]
15%|#5 | 448M/2.91G [00:08<00:43, 61.6MB/s]
15%|#5 | 454M/2.91G [00:08<00:46, 56.7MB/s]
16%|#5 | 464M/2.91G [00:08<00:43, 60.9MB/s]
16%|#6 | 477M/2.91G [00:08<00:33, 78.5MB/s]
16%|#6 | 486M/2.91G [00:09<00:43, 59.9MB/s]
17%|#6 | 496M/2.91G [00:09<01:09, 37.3MB/s]
17%|#6 | 501M/2.91G [00:09<01:10, 36.7MB/s]
17%|#7 | 511M/2.91G [00:09<00:58, 44.2MB/s]
17%|#7 | 516M/2.91G [00:10<01:00, 42.9MB/s]
18%|#7 | 527M/2.91G [00:10<00:51, 49.8MB/s]
18%|#7 | 532M/2.91G [00:10<00:53, 47.7MB/s]
18%|#8 | 544M/2.91G [00:10<00:49, 51.5MB/s]
19%|#8 | 560M/2.91G [00:10<00:45, 56.1MB/s]
19%|#9 | 576M/2.91G [00:11<00:36, 68.9MB/s]
20%|#9 | 591M/2.91G [00:11<00:35, 71.0MB/s]
20%|## | 598M/2.91G [00:11<00:39, 63.0MB/s]
20%|## | 607M/2.91G [00:11<00:40, 61.7MB/s]
21%|## | 613M/2.91G [00:11<00:50, 49.3MB/s]
21%|## | 624M/2.91G [00:12<00:49, 50.2MB/s]
21%|##1 | 639M/2.91G [00:12<00:37, 65.2MB/s]
22%|##1 | 646M/2.91G [00:12<00:45, 53.8MB/s]
22%|##2 | 656M/2.91G [00:12<00:45, 53.6MB/s]
22%|##2 | 661M/2.91G [00:12<00:49, 49.6MB/s]
23%|##2 | 672M/2.91G [00:12<00:44, 54.9MB/s]
23%|##2 | 677M/2.91G [00:13<00:57, 42.3MB/s]
23%|##3 | 688M/2.91G [00:13<00:50, 47.3MB/s]
24%|##3 | 704M/2.91G [00:13<00:37, 63.0MB/s]
24%|##3 | 710M/2.91G [00:13<00:43, 54.5MB/s]
24%|##4 | 720M/2.91G [00:13<00:42, 55.4MB/s]
25%|##4 | 731M/2.91G [00:14<00:40, 57.9MB/s]
25%|##4 | 737M/2.91G [00:14<00:40, 57.7MB/s]
25%|##4 | 742M/2.91G [00:14<00:44, 53.0MB/s]
25%|##5 | 752M/2.91G [00:14<00:41, 55.9MB/s]
26%|##5 | 768M/2.91G [00:14<00:34, 67.3MB/s]
26%|##6 | 784M/2.91G [00:14<00:27, 84.1MB/s]
27%|##6 | 792M/2.91G [00:14<00:32, 70.9MB/s]
27%|##6 | 800M/2.91G [00:15<00:35, 64.2MB/s]
27%|##7 | 815M/2.91G [00:15<00:37, 59.9MB/s]
28%|##7 | 821M/2.91G [00:15<00:42, 53.4MB/s]
28%|##7 | 832M/2.91G [00:15<00:42, 52.6MB/s]
28%|##8 | 847M/2.91G [00:15<00:35, 63.5MB/s]
29%|##8 | 853M/2.91G [00:16<00:41, 53.3MB/s]
29%|##8 | 864M/2.91G [00:16<00:40, 54.9MB/s]
29%|##9 | 870M/2.91G [00:16<00:43, 50.9MB/s]
30%|##9 | 880M/2.91G [00:16<00:39, 55.9MB/s]
30%|### | 896M/2.91G [00:16<00:30, 71.2MB/s]
30%|### | 903M/2.91G [00:17<00:37, 58.7MB/s]
31%|### | 912M/2.91G [00:17<00:42, 51.2MB/s]
31%|### | 917M/2.91G [00:17<00:45, 48.0MB/s]
31%|###1 | 928M/2.91G [00:17<00:39, 54.8MB/s]
31%|###1 | 933M/2.91G [00:17<00:46, 46.4MB/s]
32%|###1 | 943M/2.91G [00:17<00:43, 49.5MB/s]
32%|###1 | 948M/2.91G [00:18<00:49, 43.3MB/s]
32%|###2 | 960M/2.91G [00:18<00:40, 52.8MB/s]
32%|###2 | 965M/2.91G [00:18<00:45, 46.5MB/s]
33%|###2 | 976M/2.91G [00:18<00:40, 51.9MB/s]
33%|###3 | 992M/2.91G [00:18<00:35, 58.9MB/s]
33%|###3 | 998M/2.91G [00:19<00:37, 55.6MB/s]
34%|###3 | 0.98G/2.91G [00:19<00:38, 53.6MB/s]
34%|###3 | 0.98G/2.91G [00:19<00:41, 49.6MB/s]
34%|###4 | 0.99G/2.91G [00:19<00:33, 62.0MB/s]
34%|###4 | 1.00G/2.91G [00:19<00:31, 64.1MB/s]
35%|###4 | 1.01G/2.91G [00:19<00:24, 82.2MB/s]
35%|###5 | 1.02G/2.91G [00:20<00:44, 45.1MB/s]
35%|###5 | 1.03G/2.91G [00:20<00:45, 44.1MB/s]
36%|###5 | 1.04G/2.91G [00:20<00:38, 51.6MB/s]
36%|###5 | 1.05G/2.91G [00:20<00:40, 50.0MB/s]
37%|###6 | 1.06G/2.91G [00:20<00:27, 70.9MB/s]
37%|###6 | 1.07G/2.91G [00:20<00:33, 58.5MB/s]
37%|###7 | 1.08G/2.91G [00:20<00:32, 60.6MB/s]
37%|###7 | 1.08G/2.91G [00:21<00:33, 59.4MB/s]
38%|###7 | 1.09G/2.91G [00:21<00:32, 59.2MB/s]
38%|###7 | 1.10G/2.91G [00:21<00:32, 59.1MB/s]
38%|###8 | 1.11G/2.91G [00:21<00:38, 50.5MB/s]
38%|###8 | 1.11G/2.91G [00:21<00:39, 49.0MB/s]
39%|###8 | 1.12G/2.91G [00:21<00:37, 50.5MB/s]
39%|###9 | 1.14G/2.91G [00:22<00:27, 69.7MB/s]
39%|###9 | 1.15G/2.91G [00:22<00:29, 64.6MB/s]
40%|###9 | 1.16G/2.91G [00:22<00:30, 60.9MB/s]
40%|#### | 1.17G/2.91G [00:22<00:28, 65.9MB/s]
41%|#### | 1.19G/2.91G [00:22<00:24, 75.2MB/s]
41%|####1 | 1.20G/2.91G [00:22<00:22, 81.6MB/s]
42%|####1 | 1.21G/2.91G [00:23<00:20, 88.9MB/s]
42%|####2 | 1.22G/2.91G [00:23<00:31, 56.8MB/s]
42%|####2 | 1.23G/2.91G [00:23<00:32, 55.7MB/s]
43%|####2 | 1.25G/2.91G [00:23<00:26, 67.2MB/s]
43%|####3 | 1.25G/2.91G [00:23<00:29, 61.1MB/s]
43%|####3 | 1.27G/2.91G [00:24<00:28, 62.2MB/s]
44%|####4 | 1.28G/2.91G [00:24<00:24, 72.1MB/s]
44%|####4 | 1.29G/2.91G [00:24<00:25, 68.1MB/s]
45%|####4 | 1.30G/2.91G [00:24<00:25, 68.0MB/s]
45%|####4 | 1.30G/2.91G [00:24<00:32, 52.7MB/s]
45%|####5 | 1.31G/2.91G [00:24<00:28, 61.2MB/s]
45%|####5 | 1.32G/2.91G [00:25<00:31, 54.3MB/s]
46%|####5 | 1.33G/2.91G [00:25<00:36, 47.2MB/s]
46%|####5 | 1.33G/2.91G [00:25<00:38, 44.5MB/s]
46%|####6 | 1.34G/2.91G [00:25<00:31, 54.0MB/s]
46%|####6 | 1.35G/2.91G [00:25<00:31, 53.2MB/s]
47%|####6 | 1.36G/2.91G [00:25<00:25, 64.8MB/s]
47%|####6 | 1.37G/2.91G [00:26<00:35, 46.1MB/s]
47%|####7 | 1.37G/2.91G [00:26<00:35, 46.0MB/s]
47%|####7 | 1.38G/2.91G [00:26<00:42, 38.7MB/s]
48%|####7 | 1.39G/2.91G [00:26<00:29, 55.4MB/s]
48%|####7 | 1.40G/2.91G [00:26<00:31, 52.2MB/s]
48%|####8 | 1.41G/2.91G [00:26<00:30, 53.3MB/s]
49%|####8 | 1.42G/2.91G [00:27<00:24, 64.2MB/s]
49%|####9 | 1.43G/2.91G [00:27<00:27, 58.4MB/s]
49%|####9 | 1.44G/2.91G [00:27<00:27, 57.9MB/s]
50%|####9 | 1.45G/2.91G [00:27<00:23, 67.5MB/s]
50%|##### | 1.46G/2.91G [00:27<00:26, 59.2MB/s]
50%|##### | 1.47G/2.91G [00:28<00:27, 55.5MB/s]
51%|##### | 1.47G/2.91G [00:28<00:28, 54.0MB/s]
51%|#####1 | 1.48G/2.91G [00:28<00:28, 54.1MB/s]
51%|#####1 | 1.49G/2.91G [00:28<00:29, 50.9MB/s]
51%|#####1 | 1.49G/2.91G [00:28<00:34, 44.4MB/s]
52%|#####1 | 1.50G/2.91G [00:28<00:40, 37.5MB/s]
52%|#####1 | 1.50G/2.91G [00:29<00:40, 37.7MB/s]
52%|#####2 | 1.52G/2.91G [00:29<00:31, 46.9MB/s]
52%|#####2 | 1.52G/2.91G [00:29<00:30, 48.8MB/s]
53%|#####2 | 1.53G/2.91G [00:29<00:26, 56.4MB/s]
53%|#####2 | 1.54G/2.91G [00:29<00:27, 53.8MB/s]
53%|#####3 | 1.55G/2.91G [00:29<00:26, 55.7MB/s]
54%|#####3 | 1.56G/2.91G [00:29<00:21, 68.5MB/s]
54%|#####3 | 1.57G/2.91G [00:30<00:23, 62.3MB/s]
54%|#####4 | 1.58G/2.91G [00:30<00:22, 64.3MB/s]
54%|#####4 | 1.58G/2.91G [00:30<00:22, 62.4MB/s]
55%|#####4 | 1.59G/2.91G [00:30<00:20, 67.4MB/s]
55%|#####4 | 1.60G/2.91G [00:30<00:22, 61.8MB/s]
55%|#####5 | 1.61G/2.91G [00:30<00:24, 58.0MB/s]
56%|#####5 | 1.62G/2.91G [00:31<00:20, 66.1MB/s]
56%|#####6 | 1.64G/2.91G [00:31<00:19, 70.2MB/s]
57%|#####6 | 1.65G/2.91G [00:31<00:19, 70.1MB/s]
57%|#####6 | 1.65G/2.91G [00:31<00:24, 54.2MB/s]
57%|#####7 | 1.66G/2.91G [00:31<00:26, 51.2MB/s]
57%|#####7 | 1.67G/2.91G [00:31<00:20, 65.5MB/s]
58%|#####7 | 1.68G/2.91G [00:32<00:23, 55.5MB/s]
58%|#####7 | 1.69G/2.91G [00:32<00:23, 56.8MB/s]
58%|#####8 | 1.69G/2.91G [00:32<00:23, 56.4MB/s]
59%|#####8 | 1.70G/2.91G [00:32<00:23, 55.7MB/s]
59%|#####9 | 1.72G/2.91G [00:32<00:20, 63.3MB/s]
60%|#####9 | 1.73G/2.91G [00:32<00:16, 74.9MB/s]
60%|#####9 | 1.74G/2.91G [00:33<00:19, 64.3MB/s]
60%|###### | 1.75G/2.91G [00:33<00:22, 55.9MB/s]
60%|###### | 1.75G/2.91G [00:33<00:25, 48.7MB/s]
61%|###### | 1.76G/2.91G [00:33<00:20, 61.4MB/s]
61%|###### | 1.77G/2.91G [00:33<00:25, 48.4MB/s]
61%|######1 | 1.78G/2.91G [00:33<00:22, 53.2MB/s]
62%|######1 | 1.80G/2.91G [00:34<00:19, 60.7MB/s]
62%|######2 | 1.81G/2.91G [00:34<00:15, 74.4MB/s]
63%|######2 | 1.82G/2.91G [00:34<00:20, 57.2MB/s]
63%|######2 | 1.83G/2.91G [00:34<00:23, 49.9MB/s]
63%|######3 | 1.83G/2.91G [00:34<00:23, 49.1MB/s]
63%|######3 | 1.84G/2.91G [00:35<00:23, 48.6MB/s]
63%|######3 | 1.85G/2.91G [00:35<00:26, 42.9MB/s]
64%|######3 | 1.86G/2.91G [00:35<00:22, 49.8MB/s]
64%|######4 | 1.87G/2.91G [00:35<00:17, 64.2MB/s]
65%|######4 | 1.88G/2.91G [00:35<00:20, 54.8MB/s]
65%|######4 | 1.89G/2.91G [00:36<00:19, 56.2MB/s]
66%|######5 | 1.91G/2.91G [00:36<00:16, 63.6MB/s]
66%|######6 | 1.92G/2.91G [00:36<00:14, 70.8MB/s]
67%|######6 | 1.94G/2.91G [00:36<00:14, 72.2MB/s]
67%|######7 | 1.95G/2.91G [00:36<00:13, 76.5MB/s]
67%|######7 | 1.96G/2.91G [00:37<00:15, 66.8MB/s]
68%|######7 | 1.97G/2.91G [00:37<00:14, 68.6MB/s]
68%|######7 | 1.97G/2.91G [00:37<00:22, 44.4MB/s]
68%|######8 | 1.99G/2.91G [00:37<00:15, 63.2MB/s]
69%|######8 | 2.00G/2.91G [00:38<00:20, 48.6MB/s]
69%|######9 | 2.01G/2.91G [00:38<00:15, 62.4MB/s]
69%|######9 | 2.02G/2.91G [00:38<00:16, 58.7MB/s]
70%|######9 | 2.03G/2.91G [00:38<00:21, 44.3MB/s]
70%|######9 | 2.03G/2.91G [00:38<00:25, 36.9MB/s]
70%|####### | 2.05G/2.91G [00:39<00:21, 43.5MB/s]
70%|####### | 2.05G/2.91G [00:39<00:20, 44.6MB/s]
71%|####### | 2.06G/2.91G [00:39<00:15, 57.8MB/s]
71%|#######1 | 2.07G/2.91G [00:39<00:17, 50.6MB/s]
71%|#######1 | 2.08G/2.91G [00:39<00:15, 57.5MB/s]
72%|#######1 | 2.08G/2.91G [00:39<00:15, 58.0MB/s]
72%|#######1 | 2.09G/2.91G [00:39<00:14, 61.1MB/s]
72%|#######2 | 2.10G/2.91G [00:40<00:15, 56.5MB/s]
72%|#######2 | 2.11G/2.91G [00:40<00:16, 53.5MB/s]
73%|#######2 | 2.11G/2.91G [00:40<00:20, 40.9MB/s]
73%|#######3 | 2.12G/2.91G [00:40<00:16, 51.4MB/s]
74%|#######3 | 2.14G/2.91G [00:40<00:12, 66.9MB/s]
74%|#######3 | 2.15G/2.91G [00:41<00:15, 51.7MB/s]
74%|#######4 | 2.16G/2.91G [00:41<00:16, 50.4MB/s]
74%|#######4 | 2.16G/2.91G [00:41<00:14, 55.2MB/s]
75%|#######4 | 2.17G/2.91G [00:41<00:15, 52.1MB/s]
75%|#######4 | 2.17G/2.91G [00:41<00:17, 44.7MB/s]
75%|#######4 | 2.18G/2.91G [00:41<00:19, 41.3MB/s]
75%|#######5 | 2.19G/2.91G [00:41<00:14, 52.2MB/s]
75%|#######5 | 2.19G/2.91G [00:42<00:15, 51.3MB/s]
76%|#######5 | 2.20G/2.91G [00:42<00:12, 61.8MB/s]
76%|#######5 | 2.21G/2.91G [00:42<00:13, 56.9MB/s]
76%|#######6 | 2.22G/2.91G [00:42<00:13, 53.5MB/s]
77%|#######6 | 2.23G/2.91G [00:42<00:11, 63.5MB/s]
77%|#######6 | 2.24G/2.91G [00:42<00:11, 60.8MB/s]
77%|#######7 | 2.25G/2.91G [00:43<00:12, 54.6MB/s]
77%|#######7 | 2.25G/2.91G [00:43<00:13, 51.5MB/s]
78%|#######7 | 2.27G/2.91G [00:43<00:11, 60.6MB/s]
78%|#######8 | 2.27G/2.91G [00:43<00:13, 50.0MB/s]
78%|#######8 | 2.28G/2.91G [00:43<00:12, 55.8MB/s]
79%|#######8 | 2.29G/2.91G [00:43<00:13, 50.9MB/s]
79%|#######8 | 2.30G/2.91G [00:43<00:10, 61.2MB/s]
79%|#######9 | 2.30G/2.91G [00:44<00:12, 52.6MB/s]
79%|#######9 | 2.31G/2.91G [00:44<00:13, 48.2MB/s]
80%|#######9 | 2.33G/2.91G [00:44<00:11, 56.1MB/s]
80%|######## | 2.33G/2.91G [00:44<00:13, 45.0MB/s]
81%|######## | 2.34G/2.91G [00:45<00:11, 52.1MB/s]
81%|######## | 2.35G/2.91G [00:45<00:14, 42.5MB/s]
81%|######## | 2.35G/2.91G [00:45<00:13, 45.8MB/s]
81%|########1 | 2.36G/2.91G [00:45<00:13, 44.4MB/s]
81%|########1 | 2.36G/2.91G [00:45<00:13, 43.5MB/s]
82%|########1 | 2.38G/2.91G [00:45<00:09, 57.6MB/s]
82%|########2 | 2.39G/2.91G [00:45<00:07, 78.2MB/s]
82%|########2 | 2.40G/2.91G [00:45<00:07, 72.6MB/s]
83%|########2 | 2.41G/2.91G [00:46<00:08, 64.4MB/s]
83%|########2 | 2.41G/2.91G [00:46<00:09, 59.3MB/s]
83%|########3 | 2.42G/2.91G [00:46<00:07, 66.5MB/s]
84%|########3 | 2.44G/2.91G [00:46<00:06, 76.5MB/s]
84%|########4 | 2.45G/2.91G [00:46<00:07, 64.4MB/s]
84%|########4 | 2.46G/2.91G [00:47<00:08, 56.5MB/s]
85%|########4 | 2.47G/2.91G [00:47<00:08, 53.9MB/s]
85%|########5 | 2.48G/2.91G [00:47<00:06, 65.8MB/s]
86%|########5 | 2.49G/2.91G [00:47<00:07, 58.1MB/s]
86%|########5 | 2.50G/2.91G [00:47<00:07, 56.6MB/s]
86%|########6 | 2.51G/2.91G [00:48<00:06, 67.0MB/s]
87%|########6 | 2.52G/2.91G [00:48<00:06, 65.8MB/s]
87%|########6 | 2.53G/2.91G [00:48<00:05, 70.9MB/s]
88%|########7 | 2.55G/2.91G [00:48<00:04, 83.6MB/s]
88%|########8 | 2.56G/2.91G [00:48<00:04, 84.6MB/s]
89%|########8 | 2.58G/2.91G [00:48<00:03, 99.6MB/s]
89%|########8 | 2.59G/2.91G [00:48<00:03, 88.8MB/s]
89%|########9 | 2.60G/2.91G [00:49<00:06, 52.7MB/s]
90%|########9 | 2.61G/2.91G [00:49<00:05, 60.9MB/s]
90%|########9 | 2.62G/2.91G [00:49<00:05, 58.5MB/s]
90%|######### | 2.62G/2.91G [00:49<00:05, 59.6MB/s]
91%|######### | 2.64G/2.91G [00:49<00:04, 70.5MB/s]
91%|#########1| 2.66G/2.91G [00:50<00:03, 77.4MB/s]
92%|#########1| 2.66G/2.91G [00:50<00:04, 53.8MB/s]
92%|#########1| 2.67G/2.91G [00:50<00:05, 47.4MB/s]
92%|#########1| 2.68G/2.91G [00:50<00:05, 45.5MB/s]
92%|#########2| 2.69G/2.91G [00:51<00:04, 50.9MB/s]
93%|#########2| 2.70G/2.91G [00:51<00:03, 67.2MB/s]
93%|#########3| 2.71G/2.91G [00:51<00:03, 64.0MB/s]
93%|#########3| 2.72G/2.91G [00:51<00:02, 69.5MB/s]
94%|#########3| 2.73G/2.91G [00:51<00:03, 59.1MB/s]
94%|#########4| 2.74G/2.91G [00:51<00:03, 50.8MB/s]
95%|#########4| 2.75G/2.91G [00:52<00:03, 55.7MB/s]
95%|#########5| 2.77G/2.91G [00:52<00:02, 60.7MB/s]
96%|#########5| 2.78G/2.91G [00:52<00:01, 69.9MB/s]
96%|#########6| 2.80G/2.91G [00:52<00:01, 82.0MB/s]
96%|#########6| 2.80G/2.91G [00:52<00:01, 62.9MB/s]
97%|#########6| 2.81G/2.91G [00:52<00:01, 67.3MB/s]
97%|#########7| 2.82G/2.91G [00:53<00:01, 67.7MB/s]
97%|#########7| 2.83G/2.91G [00:53<00:01, 58.0MB/s]
98%|#########7| 2.84G/2.91G [00:53<00:01, 55.1MB/s]
98%|#########7| 2.84G/2.91G [00:53<00:01, 48.7MB/s]
98%|#########7| 2.85G/2.91G [00:53<00:01, 43.1MB/s]
98%|#########8| 2.85G/2.91G [00:54<00:01, 37.5MB/s]
98%|#########8| 2.86G/2.91G [00:54<00:01, 36.0MB/s]
99%|#########8| 2.87G/2.91G [00:54<00:00, 49.5MB/s]
99%|#########8| 2.88G/2.91G [00:54<00:00, 40.5MB/s]
99%|#########9| 2.89G/2.91G [00:54<00:00, 46.2MB/s]
99%|#########9| 2.89G/2.91G [00:55<00:00, 36.4MB/s]
100%|#########9| 2.91G/2.91G [00:55<00:00, 48.5MB/s]
100%|##########| 2.91G/2.91G [00:55<00:00, 56.4MB/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，以便进行比较。
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，它还接受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，就可以进行推理了。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 相对应的时步。回想一下与波形对应的转录文本是
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，将它们转换为原始标记，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，使用词典约束的束搜索解码器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¶
如果输入语音很长，可以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，但将其一次一帧地传递给解码器。
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处理解码相同。
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 相对应的时步。
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，我们将可视化标记时间步与原始波形的对齐情况。
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)
束搜索解码器参数¶
在本节中，我们将更深入地探讨一些不同的参数和权衡。有关可定制参数的完整列表，请参阅documentation。
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属性。例如，通过在之前构建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)
束大小¶
beam_size参数决定在每个解码步骤后要保存的最大最佳假设数量。使用更大的束大小允许探索更大范围的可能假设，这可以产生具有更高分数的假设，但它在计算上更昂贵，并且在一定程度上不会提供额外的收益。
在下面的示例中，我们看到解码质量随着我们从 1 增加到 5 再到 50 增加束大小而得到改善，但请注意，使用 500 的束大小如何提供与 50 的束大小相同的输出，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; 0.0475 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; 0.0519 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; 0.1656 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; 0.5422 secs)
束大小标记¶
beam_size_token参数对应于在解码步骤中考虑用于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，但会以计算为代价。
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; 0.1602 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; 0.1807 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; 0.1997 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; 0.2320 secs)
束阈值¶
beam_threshold参数用于在每个解码步骤中修剪存储的假设集，删除分数比最高得分假设高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; 0.0287 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; 0.0497 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; 0.2134 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; 0.2353 secs)
语言模型权重¶
参数 lm_weight 是分配给语言模型分数的权重，该分数将与声学模型分数累加，以确定总分数。较大的权重鼓励模型根据语言模型预测下一个词，而较小的权重则更多地考虑声学模型分数。
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; 0.2594 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; 0.2673 secs)
lm weight 15 : was there in his was at some of what he said (score: 2918.99; 0.2436 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：添加的未知词出现分数sil_score：添加的静音出现分数log_add：是否使用对数加法进行词典 Trie 扩散
脚本总运行时间： ( 2 分钟 50.590 秒)
