5.1.5. Speech Recognition and Synthesis
Speech capabilities convert between spoken and written language, enabling applications to hear users and speak back to them. Think of it as giving your application ears and a voice.
Speech Recognition (Speech-to-Text):
- Converts spoken audio to written text
- Real-time or batch processing modes
- Can identify different speakers (speaker diarization)
- Supports multiple languages and accents
- Handles background noise and audio quality variations
Key scenarios for speech-to-text:
- Transcribing meetings and interviews automatically
- Creating closed captions for videos
- Enabling voice commands in applications
- Transcribing call center conversations for analysis
- Dictation and hands-free documentation
Speech Synthesis (Text-to-Speech):
- Converts written text to spoken audio
- Creates natural-sounding voices using neural networks
- Supports multiple voices, languages, and speaking styles
- Enables audio accessibility features for visually impaired users
- Can adjust speaking rate, pitch, and emphasis
Custom Neural Voice lets you create a unique synthetic voice from recordings of a specific person—useful for brand consistency or preserving someone's voice.
SSML (Speech Synthesis Markup Language) gives fine-grained control over pronunciation, pauses, emphasis, and speaking style in text-to-speech output.
Tokenization is fundamental to speech synthesis—it breaks text into individual words and phonemes so each can be assigned the correct sounds for pronunciation. Without proper tokenization, the system wouldn't know how to pronounce abbreviations, numbers, or uncommon words.
⚠️ Exam Tip: Speaker diarization identifies WHO is speaking (Speaker 1, Speaker 2). Speaker recognition verifies speaker IDENTITY (is this John?). These are different capabilities.
