📚 Pre-U WJEC Computer Science: Speaking and Listening Revision | Pre-U WJEC 计算机：口语/听力备考专项
In the Pre-U WJEC Computer Science specification, the concepts of “speaking” and “listening” are explored through the fascinating world of speech processing technologies. This field covers the way machines interpret human speech (listening, or automatic speech recognition) and produce spoken output (speaking, or speech synthesis). Understanding these systems requires a blend of acoustics, signal processing, machine learning, and natural language processing. This revision guide will walk you through the core theory, algorithms, and exam tips you need to master this topic.
在 Pre-U WJEC 计算机科学考试大纲中，“口语”和“听力”的概念通过引人入胜的语音处理技术领域展开。该领域涵盖了机器如何解读人类语音（听力，即自动语音识别）以及如何生成语音输出（口语，即语音合成）。理解这些系统需要声学、信号处理、机器学习和自然语言处理的综合知识。本备考指南将带你梳理核心理论、算法和应试要点，助你彻底掌握这一专题。
1. Speech Technology in the Pre-U Syllabus | Pre-U 大纲中的语音技术
WJEC Pre-U Computer Science includes speech interfaces as part of human-computer interaction and artificial intelligence. Candidates are expected to describe the components of a speech recognition pipeline and a text-to-speech system. You should be able to explain feature extraction, acoustic modelling, and language modelling. Exam questions often ask you to compare rule-based and statistical approaches, or to discuss the challenges of noisy environments and speaker variability.
WJEC Pre-U 计算机科学将语音接口作为人机交互与人工智能的一部分。考生需要描述语音识别流水线和文本转语音系统的组成部件。你应能够解释特征提取、声学建模和语言建模。考试题目往往要求你比较基于规则的方法与统计方法，或讨论噪声环境和说话人差异带来的挑战。
Both AS and A2 papers may include short-answer questions on speech technology, and occasionally an essay requiring evaluation of current applications such as virtual assistants. Memorising the typical block diagrams and understanding the role of each stage is essential for high marks.
AS 和 A2 试卷都可能包含关于语音技术的简答题，偶尔还会出现要求评估虚拟助手等当前应用的小论文。记住典型的结构框图并理解每个阶段的作用，是取得高分的基础。
2. Fundamentals of Speech Production and Acoustics | 语音产生与声学基础
Human speech is produced by air from the lungs passing through the vocal cords, which can vibrate to create voiced sounds. The vocal tract (throat, mouth, nasal cavity) acts as a filter, shaping the spectrum of the sound. Vowels are characterised by formants – peaks in the frequency spectrum – while consonants often involve transient bursts or frication noise. In computer science, we model speech as a signal that changes over time, carrying both acoustic and linguistic information.
人类语音是由来自肺部的气流通过声带产生的，声带振动产生浊音。声道（咽喉、口腔、鼻腔）充当滤波器，塑造声音的频谱。元音的特征是共振峰——频谱中的峰值——而辅音通常涉及瞬态爆破音或摩擦噪声。在计算机科学中，我们将语音建模为随时间变化的信号，携带着声学和语言信息。
The fundamental frequency (F0) determines the pitch, and its variation conveys intonation and stress. Understanding these acoustic properties helps in designing robust feature extraction methods. For Pre-U, you might be asked to explain how a spectrogram represents speech, and why formant frequencies are crucial for distinguishing vowels like /iː/ and /ɑː/.
基频 (F0) 决定音高，其变化传递语调与重音。理解这些声学特性有助于设计稳健的特征提取方法。对于 Pre-U，你可能会被要求解释语谱图如何表示语音，以及为什么共振峰频率对于区分像 /iː/ 和 /ɑː/ 这样的元音至关重要。
3. Feature Extraction: MFCCs and Beyond | 特征提取：MFCC 及更多
Raw audio waveforms are not directly usable for recognition; we need compact, discriminating features. Mel-Frequency Cepstral Coefficients (MFCCs) are the most widely used features in speech recognition. The process involves pre-emphasis, framing, windowing, FFT, Mel-filterbank energy computation, log scaling, and Discrete Cosine Transform (DCT). MFCCs mimic the human ear’s perception of pitch and loudness.
原始音频波形无法直接用于识别；我们需要紧凑且具有区分力的特征。梅尔频率倒谱系数 (MFCC) 是语音识别中使用最广泛的特征。其过程包括预加重、分帧、加窗、FFT、梅尔滤波器组能量计算、对数缩放及离散余弦变换 (DCT)。MFCC 模仿了人耳对音高和响度的感知。
Other features like delta and delta-delta coefficients capture temporal changes, adding velocity and acceleration information. For Pre-U, you should be able to list the steps of MFCC extraction and justify why log scale and DCT decorrelate the features. A typical exam question might provide a simplified pipeline and ask you to identify the purpose of each block.
其他特征如差分和二阶差分系数捕捉时间变化，添加速度和加速度信息。对于 Pre-U，你应该能够列出 MFCC 提取的步骤，并论证为何对数缩放和 DCT 能去除特征间的相关性。典型的考题可能会提供一个简化的流水线，要求你说明每个模块的用途。
4. Architecture of an Automatic Speech Recognition (ASR) System | 自动语音识别系统架构
A typical ASR system comprises an acoustic model, a pronunciation dictionary, a language model, and a decoder. The acoustic model maps acoustic features to phonemes or subword units. The pronunciation dictionary provides the phonetic transcription of words, while the language model estimates the probability of word sequences. The decoder searches for the most likely word sequence given the acoustic input, using all three knowledge sources.
典型的 ASR 系统包含声学模型、发音词典、语言模型和解码器。声学模型将声学特征映射到音素或子词单元。发音词典提供单词的音素转写，而语言模型估计单词序列的概率。解码器结合这三个知识源，在给定声学输入的情况下搜索最可能的单词序列。
Pre-U candidates should be comfortable with the formula derived from Bayes’ rule: P(W|O) ∝ P(O|W) × P(W), where W is the word sequence and O is the acoustic observation. P(O|W) is computed by the acoustic model and pronunciation dictionary; P(W) comes from the language model. This decomposition is fundamental to understanding why separate models are trained.
Pre-U 考生应熟悉基于贝叶斯规则的公式：P(W|O) ∝ P(O|W) × P(W)，其中 W 是词序列，O 是声学观察。P(O|W) 由声学模型和发音词典计算；P(W) 来自语言模型。这一分解对于理解为何要单独训练各个模型至关重要。
5. Hidden Markov Models in Speech Recognition | 语音识别中的隐马尔可夫模型
For decades, Hidden Markov Models (HMMs) were the backbone of acoustic modelling. An HMM models a phoneme as a sequence of hidden states, each emitting feature vectors according to a probability distribution (often Gaussian Mixture Models, GMMs). Transitions between states capture the temporal structure of speech. The three classic problems – evaluation, decoding, and training – are solved by the Forward algorithm, Viterbi algorithm, and Baum-Welch re-estimation respectively.
几十年来，隐马尔可夫模型 (HMM) 一直是声学建模的中坚力量。HMM 将音素建模为一系列隐藏状态，每个状态根据概率分布（常为高斯混合模型，GMM）发射特征向量。状态之间的转移捕捉语音的时间结构。三个经典问题——评估、解码和训练——分别由前向算法、维特比算法和 Baum-Welch 重估算法解决。
In the Pre-U exam, you may be asked to draw a left-to-right HMM topology for a word like “cat” (three phonemes, each with three emitting states). You should be able to explain how the Viterbi algorithm finds the most likely state sequence and how GMMs model the feature distribution of each state. Although deep learning has overtaken HMMs in industry, HMMs remain a key part of the syllabus due to their clear probabilistic framework.
在 Pre-U 考试中，你可能会被要求绘制一个像 “cat” 这样的词的从左到右的 HMM 拓扑结构（三个音素，每个有三个发射状态）。你应该能够解释维特比算法如何找到最可能的状态序列，以及 GMM 如何对每个状态的特征分布进行建模。尽管深度学习在工业界已超越 HMM，但 HMM 因其清晰的概率框架，仍然是考纲的重要组成部分。
6. Deep Learning Approaches for Speech | 深度学习方法在语音中的应用
Modern ASR systems use deep neural networks (DNNs) to replace GMMs within the HMM framework, or to directly map acoustics to characters (end-to-end models). Recurrent neural networks (RNNs), particularly Long Short-Term Memory (LSTM) networks, capture long-range temporal dependencies. Connectionist Temporal Classification (CTC) allows training on unsegmented sequence data, while attention-based models have pushed state-of-the-art accuracy even further.
现代 ASR 系统使用深度神经网络 (DNN) 取代 HMM 框架内的 GMM，或者直接建立从声学到字符的映射（端到端模型）。循环神经网络 (RNN)，特别是长短期记忆 (LSTM) 网络，能够捕获长距离的时间依赖关系。联结时序分类 (CTC) 允许在未分段的序列数据上进行训练，而基于注意力的模型则将准确率推向了更高的前沿。
For Pre-U, you need to contrast traditional HMM-GMM systems with deep learning methods, highlighting advantages such as feature learning, better handling of speaker variation, and scalability. However, you should also discuss challenges like the need for huge training datasets and the “black box” nature of DNNs. A high-mark essay might compare CTC with attention-based encoders-decoders in terms of alignment and latency.
对于 Pre-U，你需要对比传统 HMM-GMM 系统与深度学习方法，突出特征学习、更好地处理说话人差异和可扩展性等优点。但同时，你也应该讨论挑战，比如需要海量训练数据和 DNN 的“黑箱”特性。一篇高分作文可能会从对齐和延迟的角度比较 CTC 与基于注意力的编码器-解码器。
7. Speech Synthesis and Text-to-Speech (TTS) | 语音合成与文本转语音
A TTS pipeline typically begins with text normalisation, converting symbols and numbers into full words. Then, letter-to-sound rules or a pronunciation dictionary assign phonemes. Prosody generation adds pitch, duration, and pause information. Finally, a speech waveform is synthesised, either by concatenating pre-recorded units (concatenative synthesis) or by generating the signal from parameters (parametric synthesis). Recent neural TTS models like Tacotron and WaveNet produce remarkably natural speech.
TTS 流水线通常从文本规范化开始，将符号和数字转换为完整单词。然后，通过字母到声音的规则或发音词典赋予音素。韵律生成添加工高、时长和停顿信息。最后，通过拼接预录制单元（拼接合成）或从参数生成信号（参数合成）来合成语音波形。近期如 Tacotron 和 WaveNet 这样的神经 TTS 模型能够生成极其自然的语音。
Pre-U candidates should be able to describe the differences between concatenative and parametric synthesis. Concatenative synthesis can sound natural but may have discontinuities at boundaries, while parametric synthesis is more flexible but can sound buzzy. Neural vocoders like WaveNet directly model waveforms autoregressively, producing high-fidelity audio. Expect questions about the evaluation of synthetic speech, such as the Mean Opinion Score (MOS).
Pre-U 考生应能描述拼接合成与参数合成的区别。拼接合成听起来自然，但边界处可能出现不连贯；参数合成更灵活，但可能带有“嗡嗡”声。像 WaveNet 这样的神经声码器以自回归方式直接建模波形，生成高保真音频。预计会有关于合成语音评估的题目，如平均意见得分 (MOS)。
8. Evaluation Metrics for Speech Systems | 语音系统的评估指标
Word Error Rate (WER) is the standard metric for ASR, calculated as (Substitutions + Insertions + Deletions) / Total Words in reference. Lower WER indicates better accuracy. For TTS, subjective listening tests like the Mean Opinion Score (MOS) ask listeners to rate naturalness on a scale of 1–5. Objective measures such as Mel-Cepstral Distortion (MCD) quantify spectral distance, but they do not always correlate well with human perception.
词错误率 (WER) 是 ASR 的标准指标，计算公式为 (替换 + 插入 + 删除) / 参考总词数。WER 越低，准确率越高。对于 TTS，主观听力测试如平均意见得分 (MOS) 要求听者按 1–5 分评价自然度。客观指标如梅尔倒谱失真 (MCD) 量化频谱距离，但它们与人的主观感受并不总是高度相关。
In Pre-U, you should be able to compute WER from a given hypothesis and reference transcript. You might also need to discuss the limitations of WER (e.g., equal weight to all errors) and the importance of task-specific evaluation. For synthesis, a common discussion point is why MOS remains the gold standard despite its cost and subjectivity.
在 Pre-U 中，你应该能够根据给定的识别结果和参考文本计算 WER。你可能还需要讨论 WER 的局限性（如对所有错误赋予同等权重）以及任务特定评估的重要性。对于合成，常见的讨论点是为何尽管 MOS 成本高且具主观性，它仍然是黄金标准。
9. Pre-U Exam Strategy and Common Pitfalls | Pre-U 考试策略与常见误区
When tackling speech-related questions, always relate your answers to the computer science principles rather than pure linguistics. Use precise terminology: “acoustic model” not “sound matcher”, “language model” not “grammar checker”. If asked to draw a spectrogram or HMM, label axes and states clearly. In essays, structure your arguments around a small number of well-explained examples, such as a virtual assistant handling a noisy command.
回答语音相关问题时，务必将答案与计算机科学原理相联系，而非纯粹的语言学。使用精准术语：“声学模型”而非“声音匹配器”，“语言模型”而非“语法检查器”。如果要求绘制语谱图或 HMM，请清晰标注坐标轴和状态。在论文中，围绕少数几个解释清楚的例子组织论点，例如虚拟助手处理含噪指令的场景。
A common mistake is to confuse the roles of the acoustic model and language model. The acoustic model says how likely a sound is for a given phoneme; the language model says how likely a word sequence is in the language. Another pitfall is neglecting to mention that MFCCs discard phase information, which is acceptable because human hearing is relatively insensitive to phase. Be ready to justify design choices.
一个常见的错误是混淆声学模型和语言模型的作用。声学模型说的是给定音素下某个声音的概率；语言模型说的是某个词序列在语言中的概率。另一个误区是忘记提到 MFCC 丢弃了相位信息，这是可以接受的，因为人耳对相位相对不敏感。准备好论证设计选择的原因。
10. Future Directions and Ethical Considerations | 未来方向与伦理考量
Speech technology is advancing rapidly with self-supervised models like wav2vec and HuBERT, which learn representations from vast unlabeled data. Multilingual and code-switching ASR, emotion recognition, and speaker diarisation are increasingly integrated into real-world applications. For Pre-U, you might be asked to discuss the societal impact: voice cloning raises concerns about consent and misinformation, while voice assistants in the home provoke privacy debates.
语音技术正随着 wav2vec 和 HuBERT 等自监督模型快速发展，这些模型从海量无标签数据中学习表征。多语言和语码切换 ASR、情感识别以及说话人分离正越来越多地集成到实际应用中。对于 Pre-U，你可能会被要求讨论社会影响：声音克隆引发了关于同意和虚假信息的担忧，而家庭语音助手则引发了隐私争议。
It is important to connect these trends to topics you have studied, such as bias in training data (accents, dialects) that can lead to unequal performance. A strong answer will show awareness of both technical limitations and the responsibility of computer scientists. Revising real-world case studies will enrich your essay and demonstrate higher-order thinking.
将这些趋势与已学专题联系起来非常重要，例如训练数据中的偏见（口音、方言），可能导致性能不均衡。一个出色的回答应展现对技术局限和计算机科学家责任的双重认识。温习真实世界的案例研究会丰富你的论文，并展示较高层次的思维能力。
Published by TutorHao | Computer Science Revision Series | aleveler.com
更多咨询请联系16621398022（同微信）
屏轩国际教育cambridge primary/secondary checkpoint, cat4, ukiset,ukcat,igcse,alevel,PAT,STEP,MAT, ibdp,ap,ssat,sat,sat2课程辅导，国外大学本科硕士研究生博士课程论文辅导Cancel reply