What is speech to text (STT)?

Speech to text, also called automatic speech recognition (ASR), is the technology that listens to audio and converts what it hears into written words. If you've ever dictated a message on your phone, used a voice assistant, or seen automatic captions appear on a video, you've already used STT.

How it works

Modern speech recognition models are trained on enormous datasets of audio paired with verified transcripts. The model learns to map acoustic patterns—the way specific phonemes sound, how they vary by speaker, accent, and noise level—to the words they represent.

At inference time, the system takes raw audio waveforms and runs them through an encoder that extracts acoustic features, then a language model component that predicts the most likely sequence of words given both the audio signal and the surrounding linguistic context. This two-part design is why modern STT is robust to background noise and thick accents: the language model can fill in ambiguous sounds using knowledge of how language actually works.

Word error rate (WER) is the standard benchmark for STT accuracy. Leading models now achieve single-digit WER on clean speech, meaning fewer than one word in ten is wrong. On noisy audio or strongly accented speech the error rate climbs, but continues to improve each year.

Where it's used

Speech to text unlocks a wide range of practical applications:

• Meeting transcription: Automatically capture what was said in a call, so participants can focus on the conversation instead of taking notes.
• Video captions: Generate accurate subtitles for video content, making it accessible to deaf or hard-of-hearing audiences and searchable by text.
• Voice-to-document: Dictate long-form content—reports, clinical notes, legal briefs—faster than typing.
• Customer intelligence: Transcribe support calls and sales conversations to surface trends, coaching opportunities, and compliance issues.
• Podcast search: Convert audio episodes into text so search engines can index the spoken content.

STT versus TTS

STT and text to speech are opposite transformations. TTS converts text into audio; STT converts audio back into text. In many real-world pipelines they appear together: a video is first transcribed with STT, the transcript is edited, and a new audio track is generated with TTS—a workflow common in dubbing and localization.

Speaker awareness

Raw transcription tells you what was said, but not who said it. For multi-speaker recordings—interviews, panel discussions, call center recordings—you often want to know which speaker said which words. That problem is handled by a related technique called speaker diarization, which can be run on top of an STT transcript to add speaker labels to each turn.

Accuracy factors

Several variables affect how accurate a transcription will be: audio quality, speaker accent, domain vocabulary, and the presence of overlapping speakers. Providing a domain-specific vocabulary list or a custom language model can dramatically improve accuracy for specialist content—medical dictation, legal proceedings, or technical support calls all benefit from tuning.