week03

Week 03

Automatic Speech Recognition I

Slides

• ASR I
  • Automatic speech recognition (ASR)
  • Metrics for ASR
  • CTC loss function
  • "Listen, Attend and Spell" archetecture
  • Beam-search

Prerequisites

• This lecture assumes you are familiar with the attention mechanism. If you're not feeling confident about this topic, we recommend that you read one of the following materials:
  • How Attention works in Deep Learning: understanding the attention mechanism in sequence models
    • Simple language. Easy to understand. Quite detailed. Not too technical. Surface knowledge.
  • Sequence to Sequence (seq2seq) and Attention
    • Nicely illustrated. Detailed math explanations. A bit bulk.

Practice & homework

• Seminar 1 Audio augmentations with code and examples
• Seminar 2 Practical excersises
  • Writing and testing WER-metric
  • Implementing CTC decoding
  • Implementing CTC beam-search
• (bonus!) Intro to PyCharm

Additional Materials

All links are provided on the last slide of the lecture

1. wer are we - track global ASR progress on various datasets
2. (paper) Librispeech: An ASR corpus based on public domain audio books (2015)
3. (paper) Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks [2006]
4. (blog post) Sequence Modeling With CTC
5. (paper) Deep Speech: Scaling up end-to-end speech recognition (2014)
6. (paper) Deep Speech 2: End-to-End Speech Recognition in English and Mandarin (2015)
7. (blog post) Speech Recognition — Deep Speech, CTC, Listen, Attend, and Spell
8. (paper) Show, Attend and Tell: Neural Image Caption Generation with Visual Attention (2015)
9. (paper) Listen, Attend and Spell (2015)
10. (paper) SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition (2019)
11. (site) http://kaldi-asr.org/
12. (paper) Show, Attend and Tell: Neural Image Caption Generation with Visual Attention