STT, short for speech-to-text, is the technology that converts spoken audio into a written transcript, enabling machines to understand and process human voice input. Modern STT systems are end-to-end neural models—most prominently OpenAI's Whisper—trained on vast multilingual audio corpora, allowing them to handle accents, background noise, and domain-specific vocabulary with high accuracy. The output can include punctuation, timestamps, and speaker labels from an optional diarization step.

Think of it like a court stenographer who has trained for years on thousands of hours of depositions, medical hearings, and technical conferences. They no longer consciously decode each sound; the mapping from spoken rhythm to written word has become automatic, just as it has for the neural model that succeeded them.

In practice, Whisper can be called with three lines: import whisper; model = whisper.load_model("base"); print(model.transcribe("recording.mp3")["text"]). For real-time voice interfaces, streaming ASR APIs process rolling 200 ms audio chunks and emit partial transcripts as the speaker continues, enabling low-latency applications like live captioning or voice command routing.

STT output feeds directly into Tokenization pipelines consumed by LLM systems; the acoustic encoder inside modern ASR models is itself a Transformer using Attention over mel-spectrogram frames, and STT is a core component of Multimodal systems that process audio alongside text and images. You will encounter it in voice assistants, meeting transcription services, accessibility software, podcast search indexing, and any spoken-language interface.