Lesson 310
Speech Recognition & Synthesis
ASR · spectrograms · phonemes · CTC · vocoders
1:00How machines convert speech to text (ASR) and text back to speech (TTS) — from raw waveforms through spectrograms and MFCC features to neural vocoders.
By the end, you can
- Distinguish ASR from TTS and explain the "same bridge, opposite directions" symmetry.
- Trace the path from raw audio to MFCC features: waveform → spectrogram (FFT) → mel scale → MFCC.
- Explain why the mel scale and ~13 MFCC coefficients are used instead of raw FFT output.
- Order the four stages of the classic ASR pipeline and state what each stage produces.
- Explain why a language model is necessary and how it resolves homophones and ambiguous phrases.
- Describe how CTC labels frames, what the blank token represents, and how collapsing produces a transcript.
- Distinguish the three modern ASR families (CTC, attention, Transformers) by their alignment strategy and trade-offs.
- Order the four stages of a TTS pipeline and state the role of G2P and the vocoder.
- Explain how WaveNet generates audio sample-by-sample, and what Parallel WaveNet improved.
- Compute a simple WER given a reference and hypothesis, and interpret a MOS score.
Up next in AI, Machine Learning & Course Review
