📚 Year 13 WJEC Computer Science: Oral and Listening Revision Focus | 十三年级 WJEC 计算机：口语与听力备考专项
In this revision guide, we explore the core concepts behind spoken language technologies that enable computers to 'hear' and 'speak'. From speech recognition pipelines to natural language understanding, these topics are essential for WJEC A Level Computer Science students aiming to master artificial intelligence applications.
在本备考指南中，我们将深入探讨使计算机能够“听”和“说”的口语技术核心概念。从语音识别流水线到自然语言理解，这些主题对于希望掌握人工智能应用的 WJEC A Level 计算机科学学生至关重要。
1. Introduction to Speech Technologies | 语音技术简介
Speech technologies encompass automatic speech recognition (ASR), which converts spoken words into text, and text-to-speech (TTS), which generates spoken output from written text. These systems form the foundation of modern voice assistants like Siri, Alexa, and Google Assistant, and they exemplify the application of machine learning, signal processing, and computational linguistics.
语音技术包括自动语音识别（ASR），将口语转换为文字；以及文语转换（TTS），从书面文本生成语音输出。这些系统构成了 Siri、Alexa 和 Google 助手等现代语音助手的基础，体现了机器学习、信号处理和计算语言学的应用。
In WJEC Computer Science, you are expected to understand the broad architecture of such systems and the role of training data, algorithms, and evaluation in developing reliable speech interfaces.
在 WJEC 计算机科学中，你需要理解这类系统的总体架构，以及训练数据、算法和评估在开发可靠语音接口中的作用。
2. Sound and Audio Fundamentals | 声音与音频基础
Sound is a continuous analogue signal created by vibrations. To process speech digitally, an analogue-to-digital converter (ADC) samples the waveform at a specific rate, typically 16 kHz for wideband speech. Each sample is quantised to a discrete amplitude, yielding a stream of binary values.
声音是由振动产生的连续模拟信号。为了以数字方式处理语音，模数转换器（ADC）会以特定的速率对波形进行采样，宽带语音通常为 16 kHz。每个采样被量化为离散幅度，从而产生一连串的二进制值。
Important properties include the sampling rate (Nyquist theorem dictates it must be at least twice the highest frequency to avoid aliasing), bit depth, and the number of channels (mono or stereo). Understanding these basics helps you appreciate why features like MFCCs are derived from raw audio.
重要的属性包括采样率（奈奎斯特定理规定，其至少应是最高频率的两倍以避免混叠）、位深度以及声道数（单声道或立体声）。理解这些基础知识有助于你理解为何要从原始音频中提取 MFCC 等特征。
3. Speech Recognition Pipeline | 语音识别流水线
The ASR pipeline typically consists of several stages: pre-processing, feature extraction, acoustic modelling, language modelling, and decoding. Pre-processing may involve noise reduction and voice activity detection to isolate speech segments. Feature extraction transforms the audio into compact, meaningful representations.
ASR 流水线通常由几个阶段组成：预处理、特征提取、声学建模、语言建模和解码。预处理可能包括降噪和语音活动检测以分离语音片段。特征提取则将音频转换为紧凑而有意义的表示。
The acoustic model maps features to phonemes or subword units, while the language model predicts the probability of word sequences. A decoder combines these models to find the most likely transcription. This modular design allows each component to be optimised independently.
声学模型将特征映射到音素或子词单元，而语言模型则预测词语序列的概率。解码器结合这些模型以找出最可能的转录文本。这种模块化设计使得每个部分都可以独立优化。
4. Acoustic Modeling and HMMs | 声学模型与隐马尔可夫模型
Traditional acoustic models often use Hidden Markov Models (HMMs) to represent the temporal structure of speech. An HMM consists of hidden states that generate observable feature vectors. Each state is associated with an emission probability distribution, typically modelled by a Gaussian Mixture Model (GMM).
传统的声学模型通常使用隐马尔可夫模型（HMM）来表示语音的时间结构。HMM 由生成可观察特征向量的隐藏状态组成。每个状态都与一个发射概率分布相关联，通常由高斯混合模型（GMM）建模。
The model has transition probabilities between states, allowing it to capture the duration and sequence of phonemes. For example, the word 'cat' can be modelled as three states for /k/, /æ/, and /t/. Training uses the Baum-Welch algorithm to estimate parameters from labelled speech data.
该模型具有状态之间的转移概率，使其能够捕捉音素的持续时间和序列。例如，单词 'cat' 可被建模为 /k/、/æ/ 和 /t/ 这三个状态。训练过程中使用 Baum-Welch 算法从带标签的语音数据中估计参数。
5. Feature Extraction: MFCCs | 特征提取：MFCC
Mel-Frequency Cepstral Coefficients (MFCCs) are the most popular features for speech recognition. The process applies a Fourier transform to short frames of audio, maps the spectrum onto the mel scale (which mimics human pitch perception), and takes the logarithm and then a discrete cosine transform.
梅尔频率倒谱系数（MFCC）是最常用的语音识别特征。这一过程先对短时语音帧进行傅里叶变换，将频谱映射到梅尔尺度（模拟人类音高感知），再取对数，然后进行离散余弦变换。
The resulting coefficients capture the vocal tract shape while discarding irrelevant details like pitch. Typically, 12–13 MFCCs plus their delta and delta-delta coefficients are used to represent each frame of speech. This compact representation greatly reduces dimensionality for the acoustic model.
得到的倒谱系数能够捕捉声道形状而丢弃如基频等无关细节。通常，每个语音帧使用 12 到 13 个 MFCC 及其一阶和二阶差分系数来表示。这种紧凑的表示显著降低了声学模型的维度。
6. Language Models and N-grams | 语言模型与 N 元语法
A language model assigns a probability to a sequence of words P(w₁, w₂, …, wₙ). N-gram models approximate this by conditioning each word only on the previous N−1 words. For example, a bigram model computes P(wₙ | wₙ₋₁), and a trigram model uses P(wₙ | wₙ₋₂, wₙ₋₁).
语言模型为一个单词序列 P(w₁, w₂, …, wₙ) 赋予一个概率。N 元语法模型通过将每个词的条件限定在前 N−1 个词来对其进行近似。例如，二元模型计算 P(wₙ | wₙ₋₁)，而三元模型使用 P(wₙ | wₙ₋₂, wₙ₋₁)。
Probabilities are estimated from large text corpora using maximum likelihood estimation with smoothing techniques (e.g., Laplace smoothing) to handle unseen word combinations. Perplexity is a common metric to evaluate language models; lower perplexity indicates a better model.
概率可以利用带平滑技术（例如拉普拉斯平滑）的最大似然估计从大规模文本语料库中估计得出，以处理未见过的词语组合。困惑度是评估语言模型的常用度量指标；困惑度越低，表明模型越好。
7. Natural Language Understanding | 自然语言理解
After ASR converts speech to text, natural language understanding (NLU) extracts meaning and intent. This often involves tokenisation, part-of-speech tagging, named entity recognition, and syntactic parsing. NLU enables a system to interpret a user's command rather than just transcribe it literally.
在 ASR 将语音转换成文字之后，自然语言理解（NLU）负责提取含义和意图。这通常涉及分词、词性标注、命名实体识别和句法分析。NLU 使系统能够理解用户的指令，而不仅仅是将其逐字转写。
Intent classification maps the utterance to a predefined action, while slot filling extracts parameters like dates, locations, or product names. Modern systems use deep learning models such as recurrent neural networks or transformers, but the underlying principle of structured interpretation remains crucial for exam questions.
意图分类将话语映射到一个预定义的动作，而槽位填充则提取日期、地点或产品名称等参数。现代系统使用循环神经网络或变换器之类的深度学习模型，但结构化解释的基本原理对于考试而言仍然至关重要。
8. Speech Synthesis and TTS | 语音合成与文语转换
Text-to-speech synthesis reverses the process: it converts written text into intelligible and natural-sounding speech. Early systems used concatenative synthesis, which joined pre-recorded speech units. Modern parametric and neural TTS models generate waveforms directly from linguistic features.
文语转换反转了这一过程：它将书面文字转换为清晰且自然的语音。早期的系统使用拼接合成，即将预录的语音单元拼接起来。现代的参量和神经 TTS 模型则直接从语言特征生成波形。
Key components include text normalisation (expanding abbreviations, numbers), grapheme-to-phoneme conversion, and prosody prediction (pitch, duration, volume). Neural networks like WaveNet produce highly realistic speech, but understanding simpler approaches helps you answer WJEC questions on TTS principles.
关键组成部分包括文本规范化（展开缩写、数字）、字形到音素转换以及韵律预测（音高、时长、音量）。像 WaveNet 这样的神经网络可以生成极其逼真的语音，但理解更简单的方法有助于你回答有关 TTS 原理的 WJEC 问题。
9. Dialogue Systems and Voice Assistants | 对话系统与语音助手
A spoken dialogue system combines ASR, NLU, dialogue management, and TTS to conduct conversations. The dialogue manager maintains the state of the interaction, decides the next system action, and handles clarification or confirmation when confidence is low.
口语对话系统结合了 ASR、NLU、对话管理和 TTS 来进行对话。对话管理器负责维护交互状态，决定系统的下一步动作，并在置信度较低时处理澄清或确认。
Typical applications include booking systems, customer service bots, and personal assistants. Designing such systems requires understanding turn-taking, error recovery, and mixed-initiative interactions. These human-computer interaction aspects are increasingly relevant in WJEC Computer Science exams.
典型的应用包括预订系统、客服机器人和个人助手。设计这类系统需要理解话轮转换、错误恢复和混合主导交互。这些人机交互方面的内容在 WJEC 计算机科学考试中日益重要。
10. Evaluation and Ethical Issues | 评估与伦理问题
The performance of speech systems is measured using word error rate (WER) for ASR, and Mean Opinion Score (MOS) for TTS. WER is calculated as (Substitutions + Insertions + Deletions) / Reference Words. These metrics provide objective ways to compare different models.
语音系统的性能使用 ASR 的词错误率（WER）和 TTS 的平均意见分（MOS）来衡量。WER 的计算公式为 (替换 + 插入 + 删除) / 参考词数。这些指标为比较不同模型提供了客观的方法。
Ethical considerations include privacy risks (voice data can be misused), bias in training data leading to poor performance for certain accents or dialects, and the potential for deep-fake audio. WJEC expects you to critically discuss these challenges alongside technical solutions.
伦理方面的考量包括隐私风险（语音数据可能被滥用）、训练数据中的偏见导致对某些口音或方言的性能不佳，以及深度伪造音频的潜在可能。WJEC 要求你在讨论这些技术解决方案的同时，也批判性地探讨这些挑战。
11. Comparison of ASR Approaches | ASR 方法对比
The field has evolved from HMM-GMM systems to deep neural networks (DNNs) and end-to-end models like Connectionist Temporal Classification (CTC) or attention-based sequence-to-sequence architectures. Each approach balances accuracy, training data requirements, and computational cost.
该领域已从 HMM-GMM 系统发展到深度神经网络（DNN）以及端到端模型，如联结时序分类（CTC）和基于注意力的序列到序列架构。每种方法都在准确率、训练数据需求和计算成本之间有所权衡。
The table below summarises key differences that can be useful for analytical questions in your exam.
下表总结了关键差异，这对考试中的分析性问题很有帮助。
| Approach | Strengths | Limitations |
|---|---|---|
| HMM-GMM | Well-understood, efficient with moderate data | Assumes independence, limited by Gaussian mixtures |
| DNN-HMM hybrid | Better discrimination, can exploit large datasets | Requires frame-level alignments |
| End-to-end (CTC/Attention) | Simplifies pipeline, directly optimises transcription | Needs very large corpora, harder to interpret |
When answering exam questions, explain why a particular architecture might be chosen for a given scenario.
在回答考试问题时，要解释为何针对特定场景选择某种架构。
12. Practical Tips for the Exam | 考试实用建议
WJEC questions on speech and language processing often require you to describe the stages of a pipeline, compare models, or evaluate the social impact of voice technology. Use precise terminology and, where possible, relate concepts to real-world examples like smart speakers or dictation software.
WJEC 关于语音和语言处理的问题通常要求你描述流水线的各个阶段、比较模型或评价语音技术的社会影响。要使用准确的术语，并尽可能将概念与智能音箱或听写软件等实际案例相联系。
Remember to balance technical detail with high-level understanding. Practice drawing labelled diagrams of ASR architectures, and be ready to write short algorithms for feature extraction or n-gram probability calculation using pseudo-code.
切记要在技术细节与高层理解之间取得平衡。练习绘制带标注的 ASR 架构图，并准备好用伪代码编写特征提取或 N 元概率计算的简短算法。
Published by TutorHao | WJEC Computer Science Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
或截图后用「扫一扫 → 相册」选择本图