← AI Terminology
Speech-to-Text
Speech-to-text (STT), also called automatic speech recognition (ASR), is the technology that converts spoken audio into written text — using neural networks to model acoustic features, language patterns, and, in modern end-to-end systems, both simultaneously.
OpenAI's Whisper made high-accuracy multilingual STT freely available for the first time.
OpenAI's Whisper made high-accuracy multilingual STT freely available for the first time.
Why It Matters in AI
Voice is the most natural human interface — but computers only speak text. STT bridges this: it enables voice assistants (Siri, Alexa), real-time transcription (Otter.ai, Teams), call centre analytics, accessibility tools (captions for the deaf), and multilingual communication. Modern end-to-end transformers (Whisper, Seamless) achieve near-human transcription accuracy across 100+ languages — a capability that previously required expensive specialised systems or human transcriptionists.
Key Points
| Aspect | Description |
|---|---|
| Whisper | OpenAI, 2022 — 680K hours of multilingual audio; open-source; near-human accuracy on many languages |
| On-device | Whisper.cpp, Apple's on-device STT, Android SpeechRecognizer — private, offline transcription |
| Diarisation | Speaker diarisation: who spoke when — "Speaker A: …, Speaker B: …" — separate task |
| Architecture | CTC-based (early) → attention encoder-decoder (modern) — Whisper uses transformer encoder-decoder |
| Word Error Rate | Primary metric: WER = (S + D + I) / N — substitutions + deletions + insertions / total words |
| Real-time vs batch | Real-time: streaming ASR with low latency; batch: transcribe full recordings — different constraints |
Simple Analogy
A court stenographer who types every spoken word in real time: the stenographer (STT model) learned to associate sound patterns with words from millions of hours of transcribed audio, and now types as fast as people speak — in any language, any accent, any domain.
Common Usage Examples
import whisper; model = whisper.load_model("large"); result = model.transcribe("audio.mp3")- HuggingFace:
pipeline("automatic-speech-recognition", model="openai/whisper-large-v3")(audio_array) - Real-time:
faster_whisper— CTranslate2 implementation, 4× faster than original Whisper whisper.cpp:./main -m models/ggml-large-v3.bin -f audio.wav— CPU-only local transcription- AssemblyAI / Deepgram APIs: real-time streaming STT for production voice applications
Summary
In short: Speech-to-text converts spoken audio into written text — Whisper made accurate multilingual open-source STT freely available, enabling voice interfaces, real-time transcription, and accessibility tools across every language at near-human accuracy.