STT (Speech-to-Text)
Converting audio to text — Whisper, Deepgram, AssemblyAI.
Simple explanation
STT (speech-to-text) transcribes audio into text. Modern models like Whisper handle multiple languages, accents and background noise. STT is the first step in any voice assistant.
Engineering explanation
STT models: Whisper (OpenAI, open-source), Deepgram Nova (streaming, low latency), AssemblyAI. Production: streaming STT for real-time (< 1.5s latency), VAD (voice activity detection) for speech boundaries, domain-tuning for accuracy on specific vocabulary/accents/noise.
Example
Audio: 'What's the weather?' → STT → text: 'What's the weather?' → LLM → response → TTS → audio response.
When to use it
Voice assistants, meeting transcription, accessibility, podcast indexing.
Common misconception
STT is NOT solved for all environments. Noisy environments, accents and domain jargon still need tuning. Clean audio = any model works; noisy = need large or domain-tuned.