Front Matter · Why This Book Exists
5 entries- F1Why This Book ExistsAudio is everywhere, yet many practitioners treat it as a spectrogram-then-CNN black box; this book makes audio AI a first-class discipline.
- F2What This Book CoversThe complete arc from sound physics and DSP to foundation models, speech and music generation, voice agents, and production engineering.
- F3Who Should Read This BookStudents and engineers who want to build audio AI systems, not only call speech APIs.
- F4How to Use This BookReading paths for audio ML engineers, speech and generative-audio specialists, production builders, and researchers.
- F5Running Tools and DatasetsThe core Python stack and the representative datasets used throughout the labs.
Part I · Sound, Signals & Audio Data
5 chapters · 35 sectionsMake readers fluent in the physical, mathematical, and computational nature of audio before models enter the story.
-
1What Is Audio Intelligence? Audio AI turns continuous pressure waves into discrete representations, decisions, and generated sound; the whole book in miniature.
- 1.1 Sound as pressure, vibration, and information
- 1.2 Human hearing and machine hearing
- 1.3 Audio tasks: classification, transcription, separation, synthesis, retrieval, dialogue
- 1.4 Speech, music, environmental sound, bioacoustics, industrial acoustics
- 1.5 The audio AI pipeline: capture → representation → model → output
- 1.6 Latency, noise, privacy, and real-world constraints
- 1.7 Book map: hear, understand, speak, deploy
- Lab Load microphone recordings, plot waveforms, compute duration, loudness, clipping, silence ratio, and basic metadata.
-
2Digital Audio Foundations How a sound becomes a number: sampling, quantization, formats, and the hygiene every audio dataset depends on.
- 2.1 Sampling theorem and aliasing
- 2.2 Quantization, bit depth, dynamic range
- 2.3 Mono, stereo, multichannel, sample rates
- 2.4 File formats: WAV, FLAC, MP3, AAC, Opus
- 2.5 Resampling and sample-rate conversion
- 2.6 Normalization, clipping, gain, loudness units
- 2.7 Dataset hygiene and audio metadata
- Lab Implement a small audio loader/resampler and demonstrate aliasing by downsampling a chirp with and without anti-alias filtering.
-
3Time-Domain Signal Processing Before any transform, audio is a stream of samples; framing, filtering, and energy are the tools you reach for first.
- 3.1 Signals as arrays and streams
- 3.2 Convolution and filtering in time
- 3.3 FIR and IIR filters
- 3.4 Envelopes, zero crossings, RMS energy
- 3.5 Autocorrelation and periodicity
- 3.6 Windowing and frame-based processing
- 3.7 Streaming buffers and overlap-add
- Lab Build a frame-based energy detector and compare it with zero-crossing based voiced/unvoiced detection.
-
4Frequency-Domain Audio Every sound is a sum of sines; the STFT and mel spectrogram are the representations the rest of the book runs on.
- 4.1 Sine waves and Fourier intuition
- 4.2 DFT and FFT from first principles
- 4.3 STFT, spectrograms, and time-frequency tradeoff
- 4.4 Mel scale, filterbanks, MFCCs
- 4.5 Constant-Q transform and music representations
- 4.6 Phase, magnitude, and reconstruction
- 4.7 Spectrogram design choices for ML
- Lab Implement STFT and inverse STFT, then compare linear spectrograms, mel spectrograms, and MFCCs on speech, music, and environmental sounds.
-
5Audio Data Engineering Curation, segmentation, augmentation, and leakage-safe splits: the unglamorous work that decides whether a model is real.
- 5.1 Dataset curation and label noise
- 5.2 Segmentation and event boundaries
- 5.3 Silence trimming and voice activity detection
- 5.4 Augmentation: noise, reverb, speed, pitch, mixup
- 5.5 Train/test splits for speakers, rooms, and recording devices
- 5.6 Leakage in audio datasets
- 5.7 Reproducible audio pipelines
- Lab Create a leakage-safe split for an audio event dataset and show how random clip splitting overestimates performance.
Part II · Classical Audio Machine Learning
4 chapters · 28 sectionsEstablish durable baselines and interpretable methods before deep networks.
-
6Classical Features for Audio Recognition Spectral, tonal, and rhythmic features plus a shallow classifier still win when data is scarce and latency is tight.
- 6.1 Frame-level features
- 6.2 Spectral centroid, bandwidth, rolloff, flux
- 6.3 Chroma and tonal features
- 6.4 Rhythm and tempo features
- 6.5 Feature aggregation over clips
- 6.6 Scaling and feature selection
- 6.7 When classical features still win
- Lab Train a scikit-learn classifier on handcrafted audio features and compare feature importance across sound classes.
-
7Classical Speech Recognition Foundations HMMs, Gaussian mixtures, lexicons, and Viterbi: the ideas deep learning replaced but never erased.
- 7.1 Phones, phonemes, words, and language models
- 7.2 Hidden Markov models
- 7.3 Gaussian mixtures and acoustic modeling
- 7.4 Pronunciation lexicons
- 7.5 N-gram language models
- 7.6 Viterbi decoding
- 7.7 Why deep learning replaced but did not erase these ideas
- Lab Implement a toy HMM recognizer for a small command vocabulary.
-
8Audio Classification Baselines Framing, metrics, calibration, and error analysis by listening: baselines that survive review.
- 8.1 Problem framing: clip-level, frame-level, event-level
- 8.2 Logistic regression, SVMs, random forests
- 8.3 Imbalanced labels and multilabel audio
- 8.4 Metrics: accuracy, F1, mAP, event F-score
- 8.5 Calibration and thresholds
- 8.6 Error analysis with listening
- 8.7 Baseline reports that survive review
- Lab Build an ESC-50 baseline and produce a confusion matrix with representative audio examples for each error type.
-
9Source Separation and Beamforming Before Deep Learning The cocktail-party problem, solved classically with masking, NMF, ICA, and arrays, as an inductive bias for neural models.
- 9.1 The cocktail party problem
- 9.2 Spectral masking
- 9.3 Non-negative matrix factorization
- 9.4 Independent component analysis
- 9.5 Microphone arrays and beamforming
- 9.6 Wiener filtering
- 9.7 Classical separation as an inductive bias for neural models
- Lab Separate simple music mixtures using NMF and evaluate with SDR-like metrics.
Part III · Deep Learning for Audio Understanding
6 chapters · 42 sectionsBuild neural audio models from small CNNs to transformers and self-supervised encoders.
-
10Neural Audio Representations Waveform or spectrogram? Learnable filterbanks, receptive fields, and pooling decide what a network can hear.
- 10.1 Waveform input versus spectrogram input
- 10.2 1D convolution over waveforms
- 10.3 2D convolution over spectrograms
- 10.4 Learnable filterbanks
- 10.5 Receptive fields in audio
- 10.6 Pooling over time and frequency
- 10.7 Choosing representation by task and latency
- Lab Compare waveform CNN, mel-CNN, and MFCC-MLP on a keyword classification task.
-
11CNNs for Audio Classification Spectrograms as images, and where the analogy breaks; VGG- and ResNet-style tagging, transfer learning, and its failure modes.
- 11.1 Spectrograms as images, and where the analogy breaks
- 11.2 VGG-like audio CNNs
- 11.3 ResNet-style audio models
- 11.4 Efficient models for edge audio
- 11.5 Multilabel audio tagging
- 11.6 Transfer learning from pretrained audio CNNs
- 11.7 Failure modes: shortcuts, background bias, microphone bias
- Lab Fine-tune a pretrained audio tagging model on a custom environmental sound dataset.
-
12Sequence Models for Audio Audio is temporal; RNNs, TCNs, and Conformers, streaming versus non-streaming, and inference with cached state.
- 12.1 Why audio is temporal
- 12.2 RNNs, GRUs, LSTMs
- 12.3 Temporal convolutional networks
- 12.4 Conformer blocks
- 12.5 Streaming versus non-streaming models
- 12.6 Handling long recordings
- 12.7 Online inference and state caching
- Lab Build a streaming keyword spotter with chunked inference and latency measurement.
-
13Audio Transformers Attention over time-frequency patches: the Audio Spectrogram Transformer, long-context attention, and compute trade-offs.
- 13.1 Attention over frames and patches
- 13.2 Audio Spectrogram Transformer
- 13.3 Positional encoding for time-frequency patches
- 13.4 Long-context audio attention
- 13.5 Masking strategies for audio
- 13.6 Fine-tuning pretrained audio transformers
- 13.7 Compute and memory tradeoffs
- Lab Fine-tune an Audio Spectrogram Transformer and compare it to a CNN baseline.
-
14Self-Supervised Audio Learning Labels are scarce; contrastive prediction, wav2vec 2.0, HuBERT, and masked spectrogram modeling learn from raw sound.
- 14.1 Why labels are scarce in audio
- 14.2 Contrastive predictive coding
- 14.3 wav2vec 2.0 style masked prediction
- 14.4 HuBERT and pseudo-labeling
- 14.5 BYOL-A and non-contrastive learning
- 14.6 AudioMAE and masked spectrogram modeling
- 14.7 Transfer, probing, and representation evaluation
- Lab Use a pretrained self-supervised encoder as a frozen feature extractor for downstream classification.
-
15Robustness and Generalization in Audio Models Noise, reverb, devices, and domains: room impulse responses, adaptation, open-set recognition, and stress testing.
- 15.1 Noise, reverberation, devices, and domains
- 15.2 Room impulse responses
- 15.3 Far-field and near-field audio
- 15.4 Domain adaptation
- 15.5 Adversarial and accidental perturbations
- 15.6 Open-set recognition
- 15.7 Stress testing audio systems
- Lab Create a robustness benchmark by adding controlled noise, reverb, compression, and device simulation.
Part IV · Speech Recognition, Speaker AI & Spoken Language
6 chapters · 42 sectionsThe core speech technologies behind dictation, transcription, diarization, and spoken interfaces.
-
16Automatic Speech Recognition: Problem and Metrics ASR as sequence transduction; targets, word error rate, alignment, and the streaming constraints that shape every design.
- 16.1 ASR as sequence transduction
- 16.2 Character, subword, and word targets
- 16.3 Word error rate and character error rate
- 16.4 Alignment, substitutions, insertions, deletions
- 16.5 Domain vocabulary and proper nouns
- 16.6 Streaming ASR constraints
- 16.7 Human transcription as a noisy reference
- Lab Evaluate an ASR model on clean, noisy, accented, and domain-specific audio.
-
17CTC Speech Recognition Alignment without labels: blanks, the forward-backward algorithm, greedy and beam decoding, and language-model fusion.
- 17.1 The alignment problem
- 17.2 CTC blanks and collapse rules
- 17.3 Forward-backward dynamic programming
- 17.4 Greedy decoding
- 17.5 Beam search and language model fusion
- 17.6 Training stability
- 17.7 CTC in modern ASR systems
- Lab Train a small CTC recognizer on a toy speech command corpus.
-
18Encoder-Decoder and Transducer ASR Attention, RNN-T, and Conformer encoders; timestamping, multilingual ASR, and Whisper-style weak supervision.
- 18.1 Attention-based encoder-decoder ASR
- 18.2 RNN-T and streaming transduction
- 18.3 Conformer encoders
- 18.4 Decoding with external language models
- 18.5 Timestamping and word alignment
- 18.6 Multilingual and code-switching ASR
- 18.7 Whisper-style weakly supervised ASR
- Lab Fine-tune a modern ASR model and build a transcript post-processing pipeline.
-
19Speaker Recognition and Verification Identity, not content: speaker embeddings, x-vectors and ECAPA, EER and DET curves, spoofing, and biometric risk.
- 19.1 Speaker identity versus speech content
- 19.2 Speaker embeddings
- 19.3 x-vectors and ECAPA-TDNN
- 19.4 Verification metrics: EER, DET curves
- 19.5 Spoofing and presentation attacks
- 19.6 Privacy and biometric risk
- 19.7 Speaker-aware applications
- Lab Build a speaker verification demo with enrollment, scoring, and threshold calibration.
-
20Diarization: Who Spoke When? Segment, embed, cluster; overlapped speech, neural diarization, and aligning speakers with the transcript.
- 20.1 Segmentation, embedding, clustering
- 20.2 Overlapped speech
- 20.3 Neural diarization
- 20.4 Diarization error rate
- 20.5 Meeting transcription pipelines
- 20.6 Speaker naming and identity linking
- 20.7 Diarization in production
- Lab Run diarization on a meeting recording and align speakers with ASR transcript segments.
-
21Spoken Language Understanding From speech to meaning: intent, slots, end-to-end SLU, prosody, emotion, and spoken dialogue state.
- 21.1 Intent detection from speech
- 21.2 Slot filling and semantic parsing
- 21.3 End-to-end spoken language understanding
- 21.4 Prosody and paralinguistic cues
- 21.5 Emotion and affect recognition
- 21.6 Spoken dialogue state tracking
- 21.7 Multimodal spoken interaction
- Lab Build a spoken command system that combines ASR, intent classification, and slot extraction.
Part V · Audio Events, Scenes & Monitoring
4 chapters · 28 sectionsNon-speech sound understanding across home, city, industry, bioacoustics, healthcare, and security.
-
22Environmental Sound Classification Weak labels and clip-level supervision over AudioSet-style taxonomies, multi-instance learning, and rare events.
- 22.1 Sound classes and acoustic scenes
- 22.2 Weak labels and clip-level supervision
- 22.3 AudioSet-style taxonomies
- 22.4 Multi-instance learning
- 22.5 Event localization in time
- 22.6 Class imbalance and rare events
- 22.7 Human-in-the-loop labeling
- Lab Train a multilabel environmental audio classifier and tune thresholds for different alert costs.
-
23Acoustic Scene Analysis Scene versus event: scene classification, sound event detection, temporal smoothing, and evaluation with collars.
- 23.1 Scene versus event
- 23.2 Background texture
- 23.3 Acoustic scene classification
- 23.4 Sound event detection
- 23.5 Temporal smoothing and post-processing
- 23.6 Evaluation with event collars
- 23.7 Deployment in noisy spaces
- Lab Build a sound event detector that outputs a timeline of events for a 10-minute recording.
-
24Industrial and Mechanical Audio AI Machines as sound sources: unsupervised anomaly detection, domain shift, and edge monitoring for motors, pumps, and bearings.
- 24.1 Machines as sound sources
- 24.2 Vibration-acoustic relation
- 24.3 Anomaly detection without labels
- 24.4 Predictive maintenance audio datasets
- 24.5 Domain shift across machines
- 24.6 Edge monitoring constraints
- 24.7 Case studies: motors, pumps, bearings, leaks
- Lab Build an unsupervised anomaly detector for machine sound using embeddings and reconstruction error.
-
25Bioacoustics and Healthcare Audio Birdsong, cough, and auscultation: weak labels, clinical validation, consent, and discovery from sound archives.
- 25.1 Animal vocalization analysis
- 25.2 Birdsong and species recognition
- 25.3 Cough, breathing, and auscultation signals
- 25.4 Medical label quality and clinical validation
- 25.5 Privacy and consent in health audio
- 25.6 Low-resource acoustic domains
- 25.7 Scientific discovery from sound archives
- Lab Construct a bioacoustic event detector and evaluate it under weak labels.
Part VI · Speech Synthesis & Generative Voice
5 chapters · 35 sectionsBuild the synthesis stack from vocoders through neural TTS and expressive, controllable speech.
-
26Speech Synthesis Foundations Text normalization, grapheme-to-phoneme, prosody, and the acoustic-model-plus-vocoder split every TTS system inherits.
- 26.1 Text normalization
- 26.2 Grapheme-to-phoneme conversion
- 26.3 Prosody, pitch, duration, energy
- 26.4 Concatenative and parametric synthesis
- 26.5 Neural acoustic models
- 26.6 Vocoders
- 26.7 Evaluating synthetic speech
- Lab Build a small text normalization and phoneme conversion pipeline.
-
27Neural Vocoders From spectrogram to waveform: Griffin-Lim, WaveNet, WaveRNN, GAN and diffusion vocoders, judged by listening.
- 27.1 From spectrogram to waveform
- 27.2 Griffin-Lim as a classical baseline
- 27.3 WaveNet and autoregressive vocoding
- 27.4 WaveRNN and efficiency
- 27.5 GAN vocoders: MelGAN, HiFi-GAN
- 27.6 Diffusion vocoders
- 27.7 Artifacts and listening tests
- Lab Compare Griffin-Lim and a pretrained neural vocoder on the same mel spectrograms.
-
28Text-to-Speech Models Tacotron, FastSpeech, and VITS; multispeaker and expressive TTS, voice cloning, and misuse prevention.
- 28.1 Tacotron-style sequence-to-sequence TTS
- 28.2 FastSpeech and non-autoregressive TTS
- 28.3 VITS and end-to-end TTS
- 28.4 Multispeaker TTS
- 28.5 Voice cloning and adaptation
- 28.6 Expressive TTS and style control
- 28.7 Responsible TTS and misuse prevention
- Lab Fine-tune a small TTS model on a speaker dataset and evaluate pronunciation, prosody, and artifacts.
-
29Voice Conversion and Speech Editing Disentangling content, speaker, and style; enhancement, inpainting, accent conversion, watermarking, and provenance.
- 29.1 Content, speaker, and style disentanglement
- 29.2 Any-to-one and any-to-any voice conversion
- 29.3 Speech enhancement and denoising
- 29.4 Speech inpainting and editing
- 29.5 Accent conversion and pronunciation correction
- 29.6 Watermarking and provenance
- 29.7 Evaluation beyond MOS
- Lab Build a voice conversion demo using pretrained content and speaker embeddings.
-
30Real-Time Conversational Speech Output Streaming and chunked synthesis under a latency budget, with barge-in, dialogue prosody, and safety filters.
- 30.1 Streaming TTS
- 30.2 Chunked synthesis
- 30.3 Latency budgets
- 30.4 Interruptibility and barge-in
- 30.5 Prosody for dialogue
- 30.6 Safety filters for generated speech
- 30.7 TTS serving architecture
- Lab Build a streaming TTS service and measure first-audio latency, total latency, and quality tradeoffs.
Part VII · Music, Sound Design & Audio Generation
5 chapters · 35 sectionsGenerative audio beyond speech: symbolic music, waveform generation, diffusion, audio-text models, and sound design.
-
31Music AI Foundations Notes, pitch, rhythm, timbre, and harmony; symbolic versus audio representations and the music-information-retrieval toolkit.
- 31.1 Notes, pitch, rhythm, timbre, harmony
- 31.2 Symbolic versus audio music representations
- 31.3 MIDI and piano rolls
- 31.4 Chroma and key estimation
- 31.5 Beat tracking and onset detection
- 31.6 Music information retrieval
- 31.7 Evaluation of music models
- Lab Build MIDI parsing, piano-roll visualization, and simple melody statistics.
-
32Symbolic Music Generation From Markov melodies to Transformer tokenizations: structure, conditioning, human editing, and copyright.
- 32.1 Markov models for melody
- 32.2 RNN music models
- 32.3 Transformer tokenizations for music
- 32.4 Structure, repetition, and long-range coherence
- 32.5 Conditioning on chord, style, and form
- 32.6 Human editing loops
- 32.7 Copyright and dataset issues
- Lab Train a small Transformer to continue symbolic melodies.
-
33Neural Audio Generation Waveform, spectrogram, and latent generation; autoregressive, GAN, diffusion, and flow matching, with guidance and control.
- 33.1 Waveform generation
- 33.2 Spectrogram generation
- 33.3 Autoregressive audio models
- 33.4 GAN-based audio synthesis
- 33.5 Diffusion and flow matching for audio
- 33.6 Latent audio generation
- 33.7 Sampling, guidance, and controllability
- Lab Use a pretrained text-to-audio model and analyze how prompt changes affect generated spectrograms and waveforms.
-
34Text-to-Audio and Audio-to-Audio Models CLAP-style embeddings, audio captioning, text-guided generation, style transfer, and semantic-alignment evaluation.
- 34.1 Audio-language pairs
- 34.2 CLAP-style contrastive embeddings
- 34.3 Audio captioning
- 34.4 Text-guided audio generation
- 34.5 Audio style transfer
- 34.6 Sound effect generation
- 34.7 Evaluation of semantic alignment
- Lab Build an audio-text retrieval system and compare retrieval using text queries, audio queries, and captions.
-
35Controllable and Interactive Sound Design Latent sliders, timbre transfer, prompt-to-sound, and inpainting for game, film, and human-in-the-loop creation.
- 35.1 Control signals and latent sliders
- 35.2 Timbre transfer
- 35.3 Prompt-to-sound workflows
- 35.4 Sound inpainting
- 35.5 Game and film sound generation
- 35.6 Human-in-the-loop creative systems
- 35.7 Dataset and rights management
- Lab Create a prompt-conditioned sound design notebook for short sound effects with versioned generations.
Part VIII · Audio-Language Models & Voice Agents
4 chapters · 28 sectionsConnect audio AI with LLMs, multimodal models, and production conversational agents.
-
36Audio Embeddings and Cross-Modal Retrieval Joint audio-text spaces for semantic search, indexing long recordings, and multimodal RAG over meetings and media.
- 36.1 Joint audio-text embedding spaces
- 36.2 Contrastive learning for audio-language
- 36.3 Audio search and semantic retrieval
- 36.4 Indexing long recordings
- 36.5 Multimodal RAG over meetings and media
- 36.6 Retrieval evaluation
- 36.7 Cost and latency of audio retrieval
- Lab Build semantic search over a podcast or lecture archive using audio embeddings and transcript chunks.
-
37Audio Captioning and Question Answering From classification to captioning and QA, grounding answers in time, and handling hallucination and uncertainty.
- 37.1 From sound classification to captioning
- 37.2 Encoder-decoder captioning
- 37.3 Audio question answering datasets
- 37.4 Grounding answers in time
- 37.5 Hallucination and uncertainty
- 37.6 Long-audio reasoning
- 37.7 Evaluation with humans and LLM judges
- Lab Build an audio QA prototype that answers questions about a short recording with cited time intervals.
-
38Realtime Voice Agents Full-duplex interaction: VAD and endpointing, the ASR-LLM-TTS cascade versus native speech-to-speech, turn-taking, and safety.
- 38.1 Full-duplex voice interaction
- 38.2 Voice activity detection and endpointing
- 38.3 ASR → LLM → TTS cascade
- 38.4 Native speech-to-speech models
- 38.5 Turn-taking, interruption, and barge-in
- 38.6 Memory and context in voice systems
- 38.7 Safety, consent, and identity
- Lab Build a real-time voice assistant with VAD, streaming ASR, LLM response generation, streaming TTS, and interruption handling.
-
39Meeting, Lecture, and Media Intelligence Ingest, transcribe, diarize, summarize, and search organizational audio, with privacy built into the architecture.
- 39.1 Recording ingestion
- 39.2 Transcription and diarization
- 39.3 Summarization and chaptering
- 39.4 Action items and decisions
- 39.5 Audio-grounded search
- 39.6 Privacy in organizational audio
- 39.7 Product architecture
- Lab Build a meeting intelligence app that produces speaker-attributed notes, action items, and searchable audio segments.
Part IX · Spatial, Multichannel & Edge Audio
3 chapters · 21 sectionsHandle audio in physical spaces and on constrained devices.
-
40Spatial Audio and Microphone Arrays Sound in rooms: interaural cues, direction of arrival, beamforming revisited, ambisonics, and neural spatial audio.
- 40.1 Sound propagation and rooms
- 40.2 Interaural time and level differences
- 40.3 Direction of arrival
- 40.4 Beamforming revisited
- 40.5 Binaural audio and ambisonics
- 40.6 Neural spatial audio
- 40.7 Applications in AR, robotics, meetings, and surveillance
- Lab Simulate a microphone array and estimate direction of arrival under noise.
-
41Audio on Edge Devices TinyML for audio: keyword spotting on microcontrollers, quantization and pruning, wake words, and battery and privacy trade-offs.
- 41.1 TinyML for audio
- 41.2 Keyword spotting on microcontrollers
- 41.3 Quantization and pruning
- 41.4 Streaming memory budgets
- 41.5 ONNX, TFLite, and embedded deployment
- 41.6 Wake words and always-on listening
- 41.7 Battery, thermal, and privacy tradeoffs
- Lab Quantize a keyword spotting model and measure size, speed, and accuracy.
-
42Real-Time Audio Systems Engineering Buffers, callbacks, jitter, and underruns; WebRTC, streaming protocols, GPU batching, observability, and graceful degradation.
- 42.1 Audio I/O stacks
- 42.2 Buffers, callbacks, jitter, and underruns
- 42.3 WebRTC and browser audio
- 42.4 Server-side streaming protocols
- 42.5 GPU batching versus latency
- 42.6 Observability for audio streams
- 42.7 Failure handling and graceful degradation
- Lab Build a browser-to-server streaming audio pipeline with latency instrumentation.
Part X · Evaluation, Trust, Deployment & Future Directions
6 chapters · 42 sectionsTurn models into measurable, reliable, responsible systems.
-
43Evaluation of Audio AI Systems Task metrics, listening tests, subgroup evaluation, calibration, and regression tests that gate a production launch.
- 43.1 Task-specific metrics
- 43.2 Listening tests and MOS
- 43.3 Dataset bias and subgroup evaluation
- 43.4 Robustness benchmarks
- 43.5 Calibration and abstention
- 43.6 Regression tests for audio systems
- 43.7 Evaluation reports for production launch
- Lab Create a unified evaluation harness for ASR, classification, and TTS components.
-
44Privacy, Security, and Responsible Audio AI Audio as sensitive data: consent, biometrics, deepfake speech, watermarking, secure pipelines, and governance.
- 44.1 Audio as sensitive data
- 44.2 Consent and recording policy
- 44.3 Biometric identifiers
- 44.4 Deepfake speech and misuse
- 44.5 Watermarking and provenance
- 44.6 Secure audio pipelines
- 44.7 Governance and documentation
- Lab Write a model card and data card for a voice AI system, including privacy and misuse sections.
-
45Deploying Audio AI Batch, streaming, and real-time serving; scaling ASR and TTS, caching, drift monitoring, and cost-latency-reliability trade-offs.
- 45.1 Batch, streaming, and real-time deployments
- 45.2 Model serving choices
- 45.3 Scaling ASR and TTS
- 45.4 Caching and reuse
- 45.5 Monitoring quality drift
- 45.6 Human review workflows
- 45.7 Cost, latency, and reliability tradeoffs
- Lab Deploy a small audio inference service with monitoring and a regression test suite.
-
46Applications and Case Studies Call centers, accessibility, in-cabin sensing, healthcare, industry, creative tools, and education, end to end.
- 46.1 Call centers and customer support
- 46.2 Accessibility and assistive technology
- 46.3 Automotive and in-cabin sensing
- 46.4 Healthcare and remote monitoring
- 46.5 Industrial monitoring
- 46.6 Creative tools and games
- 46.7 Education, lectures, and language learning
- Lab Design an end-to-end case study blueprint for one chosen application, including data, model, evaluation, deployment, and risk controls.
-
47Frontier Research in Audio AI Speech foundation models, universal encoders, audio-language-action models, and real-time speech-to-speech reasoning.
- 47.1 Speech foundation models
- 47.2 Universal audio encoders
- 47.3 Audio-language-action models
- 47.4 Generative audio at long duration
- 47.5 Real-time speech-to-speech reasoning
- 47.6 Multimodal agents with hearing
- 47.7 Open research problems
- Lab Replicate a recent audio foundation model result at small scale using frozen encoders and lightweight adapters.
-
48Capstone: An End-to-End Audio AI System Choose a problem, plan the data, build a baseline and a modern model, evaluate honestly, deploy, and present.
- 48.1 Problem selection
- 48.2 Data plan and recording protocol
- 48.3 Baseline and modern model
- 48.4 Evaluation harness
- 48.5 Real-time or batch deployment
- 48.6 Responsible use review
- 48.7 Final report and demo
Appendices · Reference & Pedagogy
7 appendices- AMathematical Foundations for Audio AIComplex numbers, Fourier transforms, convolution, probability, entropy, cross-correlation, dynamic programming, and optimization.
- BPyTorch and Torchaudio PrimerTensors, datasets, dataloaders, transforms, GPU training, mixed precision, streaming inference, and checkpointing.
- CAudio Datasets and BenchmarksDataset cards, licenses, task mapping, common pitfalls, leakage warnings, and benchmark protocols.
- DEvaluation Metrics ReferenceWER, CER, PER, DER, EER, SDR, SI-SDR, PESQ, STOI, FAD, CLAPScore, mAP, event F-score, and MOS protocols.
- EMicrophones, Audio Interfaces, and Edge HardwareMicrophone types, polar patterns, arrays, preamps, ADCs, noise floors, embedded boards, and inference trade-offs.
- FCourse SyllabiUndergraduate and graduate audio and speech tracks, a professional voice-AI bootcamp, and a generative-audio seminar.
- GSolutions to Selected ExercisesWorked solutions for mathematical derivations, coding exercises, and system-design questions.