Applied Speech Recognition
Chapter 1: Foundations of Digital Audio and Speech
Introduction to Automatic Speech Recognition Systems
Properties of Human Speech: Phonemes and Allophones
Digital Audio Signals: Sampling, Quantization, and Encoding
Working with Audio Data in Python using Librosa
Time and Frequency Domain Analysis
Introduction to Spectrograms for Speech Visualization
Hands-on Practical: Loading and Visualizing Audio Waveforms
Chapter 2: Feature Extraction for Speech Recognition
The Role of Feature Extraction in ASR
Mel Frequency Cepstral Coefficients (MFCCs)
Calculating MFCCs Step-by-Step
Filter Banks and Log-Mel Spectrograms
Feature Normalization Techniques
Comparing MFCCs and Spectrograms as Input Features
Practice: Extracting and Normalizing Features from a Dataset
Chapter 3: Acoustic Modeling with Deep Neural Networks
Overview of Acoustic Models in ASR
Building Acoustic Models with Recurrent Neural Networks
Addressing Sequential Challenges with LSTMs and GRUs
Connectionist Temporal Classification (CTC) Loss
Implementing a CTC-based ASR Model
Hands-on Practical: Training a Simple LSTM Acoustic Model with CTC
Chapter 4: Advanced Acoustic Models and Architectures
Attention Mechanisms for Speech Recognition
Sequence-to-Sequence (Seq2Seq) Models for ASR
Listen, Attend, and Spell (LAS) Architecture
Introduction to Transformer Models for ASR
Conformer: Combining CNNs and Transformers
An Overview of Pre-trained ASR Models
Practice: Fine-tuning a Pre-trained ASR Model
Chapter 5: Language Modeling and Decoding
The Function of Language Models in ASR
N-gram Language Models
Building an N-gram Model with KenLM
Decoding Graphs for Model Integration
Decoding Algorithms: Greedy Search vs Beam Search
Implementing Beam Search with a Language Model
Hands-on Practical: Integrating a Language Model into a CTC Decoder
Chapter 6: Evaluating and Deploying ASR Systems
Metrics for ASR Performance: WER and CER
Calculating Word Error Rate
Common Data Augmentation Techniques for Speech
Using Hugging Face Pipelines for ASR
Building a Speech-to-Text Application with Gradio
Considerations for Real-time Streaming ASR
Practice: Evaluating and Building a Demo Application
Listen, Attend, and Spell (LAS) Architecture
Was this section helpful?
- Listen, Attend and Spell, William Chan, Navdeep Jaitly, Quoc V. Le, Oriol Vinyals, 2015 arXiv preprint arXiv:1508.01211 DOI: 10.48550/arXiv.1508.01211 - Introduces the end-to-end Listen, Attend, and Spell (LAS) model for speech recognition, detailing its encoder-decoder architecture with attention.
- Sequence to Sequence Learning with Neural Networks, Ilya Sutskever, Oriol Vinyals, Quoc V. Le, 2014 Advances in Neural Information Processing Systems 27 (NIPS 2014), Vol. 27 (Neural Information Processing Systems Foundation, Inc. (NeurIPS)) DOI: 10.48550/arXiv.1409.3215 - Presents the foundational sequence-to-sequence learning framework, which serves as the basis for the LAS model's approach to ASR.
- Neural Machine Translation by Jointly Learning to Align and Translate, Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio, 2014 International Conference on Learning Representations (ICLR 2015) DOI: 10.48550/arXiv.1409.0473 - Introduces the attention mechanism, a core component that allows the decoder to selectively focus on parts of the input sequence, adopted by LAS.
- Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks, Alex Graves, Santiago Fernández, Faustino Gomez, Jürgen Schmidhuber, 2006 Proceedings of the 23rd International Conference on Machine Learning (ICML '06) (Association for Computing Machinery) DOI: 10.1145/1150447.1150493 - Describes Connectionist Temporal Classification (CTC), providing important context for the development of LAS by highlighting earlier sequence modeling methods.