Speech-to-Text
Speech-to-text (STT), also called automatic speech recognition (ASR), is AI technology that converts spoken audio into written text: you talk, and the software writes down what you said. It powers phone dictation, live captions, voice assistants and voice chatbots, and it is the first step whenever an AI system has to understand speech.
Understanding Speech-to-Text
Speech-to-text has advanced dramatically with deep learning. Modern ASR systems like OpenAI's Whisper achieve human-level transcription accuracy across accents, languages, and acoustic conditions. This accuracy has made voice input viable for professional use cases beyond simple commands. Meeting transcription, voice note capture, voice-commanded task creation, and voice-driven AI assistants all depend on reliable STT. The combination of STT with LLM understanding creates truly natural voice interfaces where you speak naturally and the AI understands intent rather than parsing rigid voice commands.
How speech-to-text works: modern speech recognition runs in three stages. First the audio is cut into short frames (typically 10 to 25 milliseconds each) and turned into a spectrogram, a picture of which frequencies are loud at each moment. Then a neural network, usually a transformer or conformer trained on hundreds of thousands of hours of transcribed speech, maps those frames to likely sounds and words. Finally a decoder picks the most probable sentence, using what it learned about language to tell "recognize speech" from "wreck a nice beach". Older systems kept separate acoustic, pronunciation and language models; today's end-to-end models learn all three together.
Is speech-to-text generative AI? It is usually classed as recognition, not generation: its job is to reproduce what was said, not to create new content. The line has blurred, though. Models such as Whisper use the same encoder-decoder transformer design as generative language models and write the transcript token by token, which is also why they occasionally "hallucinate" words nobody said, especially during silence or background noise. Text-to-speech, the reverse direction, is the generative half of a voice system.
Speech-to-text in chatbots: in a voice chatbot or voice assistant, STT is one link in a chain. Speech-to-text turns your words into text, a large language model reads that text and decides what to say or do, and text-to-speech reads the answer aloud. Streaming STT sends partial transcripts while you are still talking, and endpointing detects when you have finished, so the reply can start within about a second. Some newer speech-to-speech models skip the text step, but most production assistants keep the three-part pipeline because each part can be swapped and checked. "Speech to intent" goes one step further than transcription: "remind me at 5" becomes a reminder, not just a sentence.
Use cases and limits: dictating messages and documents, captions for video and live events, call-center analytics, medical and legal dictation, voice search, and hands-free control while driving or cooking. Accuracy is measured by word error rate (WER), the share of words inserted, dropped or substituted. Clean dictation into a close microphone often scores in the low single digits, while crosstalk, strong accents, jargon and phone-quality audio push it much higher. For a verbatim record of a meeting with speaker labels, use a dedicated transcription service; for an assistant that acts on what you say, STT plus an LLM is the right design.
How GAIA Uses Speech-to-Text
GAIA's voice agent component uses speech-to-text to enable hands-free interaction. You can dictate tasks, ask questions about your schedule, and issue commands verbally.
Related Concepts
Text-to-Speech
Text-to-speech (TTS) is the technology that converts written text into synthesized spoken audio, enabling computers and AI systems to communicate verbally through natural-sounding voices.
Natural Language Processing (NLP)
Natural Language Processing (NLP) is a branch of artificial intelligence that focuses on enabling computers to understand, interpret, generate, and respond to human language in a meaningful way.
Intent Recognition
Intent recognition is the process by which an AI system identifies the underlying goal or purpose of a user's input, enabling it to select the appropriate response or action rather than responding only to surface-level phrasing.
Multimodal AI
Multimodal AI refers to artificial intelligence systems that can process and generate multiple types of data, such as text, images, audio, and video, within a single model or integrated pipeline.
