Speech Recognition vs. Voice Recognition | 语音识别工作原理 | 模型训练 | 应用

注：机翻，未校。

Speech Recognition 与 Voice Recognition

剑桥词典
 
speech recognition，语音识别
voice recognition，声音识别

• Speech vs. Voice - What’s the Difference? | This vs. That
  https://thisvsthat.io/speech-vs-voice
• Voice vs Speech - What’s the difference? | WikiDiff
  https://wikidiff.com/voice/speech

---

Speech Recognition vs. Voice Recognition: In Depth Comparison 语音识别与声音识别：深度比较

Calendar12 July 2023

Have you ever stopped to think about how your voice magically turns into written words or how your smartphone recognizes your unique vocal identity? It’s mind-boggling, right?
你有没有停下来想过你的声音是如何神奇地变成书面文字的，或者你的智能手机是如何识别你独特的声音身份的？这真是令人难以置信，对吧？

Imagine this: you’re sitting in a room, jotting down notes for an important presentation. Instead of tediously typing every word, wouldn’t it be incredible if you could simply speak your thoughts and watch as they appear on the screen before your eyes? That’s the power of speech recognition! It’s like having your own personal stenographer, effortlessly transcribing your spoken words into written text.
想象一下：你坐在一个房间里，为一个重要的演示文稿记笔记。与其乏味地输入每个单词，不如简单地说出您的想法并看着它们出现在您眼前的屏幕上，这不是很不可思议吗？这就是语音识别的力量！这就像拥有自己的私人速记员，毫不费力地将您的口语转录成书面文本。

But hold on, that’s not all! Have you ever seen those spy movies where a secret agent’s voice unlocks a high-tech vault? Well, that’s voice recognition in action! It’s like having a superpower that allows you to open doors, access your digital devices, and even perform secure transactions, all with the sound of your voice.
但是等一下，这还不是全部！你有没有看过那些秘密特工的声音打开高科技金库的间谍电影？嗯，这就是声音识别的实际应用！这就像拥有一种超能力，可以让您打开门、访问您的数字设备，甚至执行安全交易，所有这些都可以通过您的声音来实现。

Now, you might be wondering, what’s the difference between speech recognition and voice recognition? Aren’t they the same thing? Ah, my curious friend, Not quite! While these terms are often used interchangeably, they actually refer to distinct technologies with their own unique abilities.
现在，您可能想知道，语音识别和声音识别有什么区别？它们不是一回事吗？啊，我好奇的朋友，不完全是！虽然这些术语经常互换使用，但它们实际上指的是具有自己独特能力的不同技术。

In this captivating article, we’ll unravel the secrets behind speech recognition and voice recognition, exploring their real-life applications, benefits, and most importantly, the intriguing differences between them.
在这篇引人入胜的文章中，我们将揭开语音识别和声音识别背后的秘密，探索它们在现实生活中的应用、好处，最重要的是，它们之间的有趣差异。

Understanding Speech Recognition 了解语音识别

Speech recognition, also known as automatic speech recognition (ASR), is a technological marvel that enables computers to convert spoken language into written text. It involves the process of analyzing audio input, extracting the spoken words, and transforming them into written form. Speech recognition systems utilize sophisticated algorithms and language models to achieve accurate transcription.
语音识别，也称为自动语音识别 （ASR），是一项技术奇迹，它使计算机能够将口语转换为书面文本。它涉及分析音频输入、提取口语单词并将其转换为书面形式的过程。语音识别系统利用复杂的算法和语言模型来实现准确的转录。

How does Speech Recognition Work? 语音识别的工作原理是什么？

The workings of speech recognition are quite fascinating. Let’s take a closer look at the underlying process:
语音识别的工作原理非常有趣。让我们仔细看看底层过程：

Audio Input: The speech recognition system receives audio input, typically through a microphone or other audio devices.
音频输入：语音识别系统通常通过麦克风或其他音频设备接收音频输入。

Pre-processing: The audio input undergoes pre-processing to eliminate background noise, enhance clarity, and normalize the audio signal.
预处理：音频输入经过预处理，以消除背景噪音、提高清晰度并使音频信号标准化。

Acoustic Modeling: The system employs acoustic modeling techniques to analyze and interpret the audio input. This involves breaking down the speech into smaller units known as phonemes and mapping them to corresponding linguistic representations.
声学建模：该系统采用声学建模技术来分析和解释音频输入。这涉及将语音分解为更小的单元，称为音素，并将它们映射到相应的语言表示。

Language Modeling:Language models play a crucial role in speech recognition by utilizing statistical patterns and grammar rules to predict and correct potential errors in transcription. They enhance the accuracy and contextuality of the converted text.
语言建模：语言模型通过利用统计模式和语法规则来预测和纠正转录中的潜在错误，在语音识别中发挥着至关重要的作用。它们提高了转换文本的准确性和上下文性。

Decoding: Using a process called decoding, the system matches the audio input against its extensive database of acoustic and language models to determine the most likely transcription.
解码：使用称为解码的过程，系统将音频输入与其广泛的声学和语言模型数据库进行匹配，以确定最可能的转录。

Text Output: Finally, the speech recognition system generates the written output, providing an accurate representation of the spoken words.
文本输出：最后，语音识别系统生成书面输出，准确表示口语。

How is The Speech Recognition Model Trained? 语音识别模型是如何训练的？

There are different types of speech datasets used to train a speech recognition model, which typically consist of paired audio and text samples. This means that for each audio segment, there is a corresponding transcription of the spoken words. The dataset needs to be diverse and representative of real-world speech patterns, encompassing different speakers, accents, languages, and recording conditions. Here’s an overview of the training process for a speech recognition model using such a dataset:
有不同类型的语音数据集用于训练语音识别模型，这些数据集通常由成对的音频和文本样本组成。这意味着对于每个音频片段，都有相应的口语转录。数据集需要多样化并代表现实世界的语音模式，包括不同的说话人、口音、语言和录制条件。以下是使用此类数据集的语音识别模型的训练过程概述：

Data Collection: Large amounts of audio data are collected from various sources, such as recorded speeches, interviews, lectures, custom collections, or publicly available datasets. The dataset should cover a wide range of topics and speakers to ensure generalization.
数据收集：从各种来源收集大量音频数据，例如录制的演讲、访谈、讲座、自定义集合或公开可用的数据集。数据集应涵盖广泛的主题和演讲者，以确保泛化。

Data Preprocessing: The collected audio data undergoes preprocessing steps to enhance its quality and normalize the audio signals. This may involve removing background noise, equalizing volume levels, and applying filters to improve clarity.
数据预处理：对采集的音频数据进行预处理，以提高音频质量并对音频信号进行归一化处理。这可能涉及消除背景噪音、均衡音量级别和应用滤镜以提高清晰度。

Transcription: Trained transcribers listen to the audio samples and manually transcribe the spoken words into written text. The transcriptions are carefully aligned with the corresponding audio segments to create the paired audio-text dataset.

Speech recognition models typically require large amounts of training data. For instance, OpenAI’s Whisper ASR system was trained on 680,000 hours of multilingual and multitask supervised data, making it one of the largest speech datasets ever created.
语音识别模型通常需要大量的训练数据。例如，OpenAI 的 Whisper ASR 系统在 680,000 小时的多语言和多任务监督数据上进行了训练，使其成为有史以来最大的语音数据集之一。

Dataset Split:The dataset is typically divided into three subsets: training, validation, and testing. The training subset, which is the largest, is used to train the model. The validation subset is used during training to monitor the model’s performance and adjust hyperparameters. The testing subset is used to evaluate the final model’s performance.
数据集拆分：数据集通常分为三个子集：训练、验证和测试。最大的训练子集用于训练模型。验证子集在训练期间用于监控模型的性能并调整超参数。测试子集用于评估最终模型的性能。

Feature Extraction: From the audio samples, various acoustic features are extracted. These features capture important characteristics of the audio, such as frequency content, duration, and intensity. Common features include Mel-frequency cepstral coefficients (MFCCs), spectrograms, and pitch information.
特征提取：从音频样本中提取各种声学特征。这些功能捕获音频的重要特征，例如频率内容、持续时间和强度。常见特征包括 Mel 频率倒谱系数 （MFCC）、频谱图和音高信息。

Language Modeling: Language models are trained on large textual datasets to learn statistical patterns, grammar rules, and linguistic contexts. These models provide additional contextual information during the training of the speech recognition model, improving its accuracy and contextuality.
语言建模：语言模型在大型文本数据集上进行训练，以学习统计模式、语法规则和语言上下文。这些模型在语音识别模型的训练过程中提供额外的上下文信息，从而提高其准确性和上下文性。

Training the Model:The speech recognition model is trained using the paired audio-text dataset and the extracted acoustic features. The model learns to associate the acoustic patterns with the corresponding textual representations. This involves using algorithms such as deep neural networks, recurrent neural networks (RNNs), or transformer-based models, which are trained using gradient-based optimization techniques.
训练模型：语音识别模型使用配对的音频文本数据集和提取的声学特征进行训练。该模型学习将声学模式与相应的文本表示相关联。这涉及使用深度神经网络、递归神经网络 （RNN） 或基于 transformer 的模型等算法，这些算法使用基于梯度的优化技术进行训练。

Iterative Training: The model is trained iteratively, where batches of data are fed to the model, and the model’s parameters are adjusted based on the prediction errors. The training process aims to minimize the difference between the predicted transcriptions and the ground truth transcriptions in the dataset.
迭代训练：对模型进行迭代训练，其中将批量数据馈送到模型，并根据预测误差调整模型的参数。训练过程旨在最小化数据集中预测的转录和 Ground Truth 转录之间的差异。

Hyperparameter Tuning: During training, hyperparameters (parameters that control the learning process) are adjusted to optimize the model’s performance. This includes parameters related to network architecture, learning rate, regularization techniques, and optimization algorithms.
超参数优化：在训练期间，会调整超参数（控制学习过程的参数）以优化模型的性能。这包括与网络架构、学习率、正则化技术和优化算法相关的参数。

Validation and Testing:Throughout the training process, the model’s performance is evaluated on the validation subset to monitor its progress and prevent overfitting. Once training is complete, the final model is evaluated on the testing subset to assess its accuracy, word error rate, and other relevant metrics.
验证和测试：在整个训练过程中，在验证子集上评估模型的性能，以监控其进度并防止过度拟合。训练完成后，将在测试子集上评估最终模型，以评估其准确性、单词错误率和其他相关指标。

Fine-tuning and Optimization: After the initial training, the model can undergo further fine-tuning and optimization to improve its performance. This may involve incorporating additional training data, adjusting model architecture, or using advanced optimization techniques.
微调和优化：在初始训练之后，模型可以进行进一步的微调和优化，以提高其性能。这可能涉及合并额外的训练数据、调整模型架构或使用高级优化技术。

By training on a diverse and extensive dataset of paired audio and text samples, speech recognition models can learn to accurately transcribe spoken words, enabling applications such as transcription services, virtual assistants, and more. The training process involves leveraging the power of machine learning algorithms and optimizing model parameters to achieve high accuracy and robustness in recognizing and transcribing speech.
通过在配对音频和文本样本的多样化和广泛的数据集上进行训练，语音识别模型可以学习准确转录口语，从而支持转录服务、虚拟助手等应用程序。训练过程包括利用机器学习算法的强大功能和优化模型参数，以实现识别和转录语音的高精度和稳健性。

Applications of Speech Recognition 语音识别的应用

Speech recognition technology has revolutionized numerous industries, transforming the way we interact with devices and systems. Here are some prominent applications:
语音识别技术已经彻底改变了许多行业，改变了我们与设备和系统的交互方式。以下是一些突出的应用：

Transcription Services
Speech recognition has streamlined the transcription process, making it faster and more efficient. It has become an invaluable tool for medical, legal, and business professionals, saving hours of manual effort.
转录服务 语音识别简化了转录过程，使其更快、更高效。它已成为医疗、法律和商业专业人士的宝贵工具，可节省数小时的手动工作。

Voice Assistants
Virtual assistants like Apple’s Siri, Amazon’s Alexa, and Google Assistant employ speech recognition to understand and respond to user commands. They can perform tasks, answer queries, and control various devices using voice commands.
语音助手 Apple 的 Siri、Amazon 的 Alexa 和 Google Assistant 等虚拟助手使用语音识别来理解和响应用户命令。他们可以使用语音命令执行任务、回答查询和控制各种设备。

Accessibility
Speech recognition has significantly improved accessibility for individuals with disabilities. It allows people with motor impairments or visual impairments to interact with computers, smartphones, and other devices using their voices.
可及性 语音识别显著改善了残障人士的辅助功能。它允许有运动障碍或视力障碍的人使用他们的声音与计算机、智能手机和其他设备进行交互。

Call Centers
Many call centers leverage speech recognition technology to enhance customer service. It enables automated call routing, voice authentication, and real-time speech-to-text conversion for call transcripts.
呼叫中心 许多呼叫中心利用语音识别技术来增强客户服务。它支持自动呼叫路由、语音身份验证和通话记录的实时语音到文本转换。

Dictation Software
Speech recognition has made dictation effortless and accurate. Professionals in various fields, such as writers, journalists, and students, benefit from dictation software that converts spoken words into written text.
听写软件 语音识别使听写变得轻松而准确。各个领域的专业人士，例如作家、记者和学生，都受益于将口语转换为书面文本的听写软件。

Benefits of Speech Recognition 语音识别的优势

Speech recognition offers several advantages that make it a powerful technology:
语音识别具有多项优势，使其成为一项强大的技术：

Increased Productivity
Speech recognition enables faster and more efficient data entry, transcription, and command execution, enhancing productivity in various domains.
提高生产力 语音识别支持更快、更高效的数据输入、转录和命令执行，从而提高各个领域的生产力。

Accessibility and Inclusivity
By allowing individuals with disabilities to interact with devices using their voices, speech recognition promotes inclusivity and equal access to technology.
可访问性和包容性 语音识别允许残障人士使用他们的声音与设备交互，从而促进了包容性和平等使用技术。

Hands-Free Operation
With speech recognition, users can perform tasks without the need for manual input, making it ideal for situations where hands-free operation is necessary or convenient.
免提操作 通过语音识别，用户无需手动输入即可执行任务，非常适合需要或方便免提操作的情况。

Multilingual Support
Advanced speech recognition systems can recognize and transcribe multiple languages, facilitating communication in diverse linguistic contexts.
多语言支持 先进的语音识别系统可以识别和转录多种语言，从而促进不同语言环境中的交流。

Understanding Voice Recognition 了解声音识别

Voice recognition, also known as speaker recognition or voice authentication, is a technology that focuses on identifying and verifying the unique characteristics of an individual’s voice. It aims to determine the identity of the speaker, rather than convert speech into text.
声音识别，也称为说话人识别或语音身份验证，是一种专注于识别和验证个人声音独特特征的技术。它旨在确定说话者的身份，而不是将语音转换为文本。

How does Voice Recognition Work? 语音识别如何工作？

Voice recognition systems employ sophisticated algorithms and machine learning techniques to analyze various vocal features, such as pitch, tone, rhythm, and pronunciation. Let’s explore the process:
声音识别系统采用复杂的算法和机器学习技术来分析各种声音特征，例如音高、语气、节奏和发音。让我们来探索一下这个过程：

Enrollment:In the enrollment phase, the system records a sample of the user’s voice, capturing their unique vocal characteristics.
注册：在注册阶段，系统会录制用户的声音样本，捕捉他们独特的声音特征。

Feature Extraction:The system extracts specific features from the recorded voice sample, analyzing factors like pitch, speech rate, and spectral patterns.
特征提取：系统从录制的语音样本中提取特定特征，分析音高、语速和频谱模式等因素。

Voiceprint Creation:Using the extracted features, the system creates a unique voiceprint, which serves as a reference for future authentication.
声纹创建：使用提取的特征，系统创建一个唯一的声纹，作为未来身份验证的参考。

Authentication: When a user attempts to authenticate, their voice is compared to the stored voiceprint. The system assesses the similarity and determines whether the speaker’s identity matches the enrolled voiceprint.
身份验证：当用户尝试进行身份验证时，他们的语音将与存储的声纹进行比较。系统会评估相似性并确定说话人的身份是否与已注册的声纹匹配。

Decision: Based on the comparison results, the voice recognition system makes a decision, either granting or denying access.
决策：根据比较结果，声音识别系统做出决策，授予或拒绝访问权限。

How is The Voice Recognition Model Trained? 语音识别模型是如何训练的？

Training a voice recognition model requires a dataset that encompasses audio samples from different individuals, capturing the unique vocal characteristics that differentiate one person from another. The dataset used for training a voice recognition model typically consists of the following components:
训练声音识别模型需要一个数据集，其中包含来自不同人的音频样本，以捕获区分一个人与另一个人的独特声音特征。用于训练声音识别模型的数据集通常由以下组件组成：

Enrolled Voice Samples:The dataset includes voice samples from individuals who voluntarily enroll in the system. These samples serve as the reference or template for each individual’s voiceprint. The enrollment process involves recording a set of voice samples from each person.
已注册的语音样本：该数据集包括来自自愿注册系统的个人的语音样本。这些样本用作每个人声纹的参考或模板。注册过程包括录制每个人的一组语音样本。

Test Voice Samples:Along with enrolled voice samples, the dataset also includes separate voice samples for testing and evaluation purposes. These samples are used to assess the model’s accuracy and performance in recognizing and verifying the identity of speakers.
测试语音样本：除了注册的语音样本外，数据集还包括用于测试和评估目的的单独语音样本。这些样本用于评估模型在识别和验证说话人身份方面的准确性和性能。

The training process for a voice recognition model involves the following steps:
声音识别模型的训练过程包括以下步骤：

Feature Extraction:From the enrolled voice samples, specific features are extracted to capture the unique vocal characteristics of each individual. These features may include pitch, speech rate, formant frequencies, spectral patterns, and other relevant acoustic properties.
特征提取：从注册的语音样本中提取特定特征以捕获每个人独特的声音特征。这些特征可能包括音高、语速、共振峰频率、频谱模式和其他相关的声学特性。

Voiceprint Creation: Using the extracted features, a voiceprint or voice template is created for each individual. The voiceprint represents a unique representation of an individual’s voice characteristics.
声纹创建