Chapter 9

CV, ASR & TTS

Same Transformer — different modality

Overview

Modality table

Patch · mel frame · latent → predict next unit

ViT

Vision Transformer

224×224 → 196 patches → Transformer → class

Code

ViT inference

HuggingFace AutoModelForImageClassification

SAM / SD

Segment & generate

SAM: click → mask (IoU) · SD: text → denoise (FID)

Whisper

Speech → text

whisper.transcribe() · encoder-decoder

WER

Word Error Rate

$(S+D+I)/N$ · jiwer · <5% = strong

TTS

Text → speech

Encoder → mel → vocoder · MOS eval

VLM

GPT-4o / LLaVA

Image + text → fused tokens → answer

Eval

Multimodal metrics

IoU · FID · WER · MOS · MMMU

Lab

5-step pipeline

ViT + Whisper + TTS + VLM + agent tools

Course complete

Scroll: course-full-multimodal.html#multimodal-practice

Evaluation Guide · Hub

Intro
1 / 1 Hub Space F