Speech-to-text (STT) converts spoken audio into written text. A speech-to-text API accepts an audio stream or a recorded file and returns a transcript, usually billed per minute of audio processed. It is the same core capability as automatic speech recognition, and vendors use the two names interchangeably.
How speech-to-text works
A speech-to-text system slices the audio into short frames, typically 10 to 25 milliseconds each. It extracts acoustic features from every frame, a compact numerical summary of the sound, then feeds that sequence to a model that predicts the most likely words.
Classical systems split the job in two: an acoustic model that maps sound to phonemes, and a separate language model that weighs likely word sequences, so the system prefers "recognize speech" over "wreck a nice beach."
Modern systems fold both into a single end-to-end neural network trained on hundreds of thousands of hours of transcribed audio. Top models now sit under 7 percent word error rate on clean speech.
A brief history of speech-to-text
Speech recognition has been a research goal since Bell Labs built "Audrey" in 1952, a machine that recognized spoken digits. IBM's Shoebox followed in 1962, and DARPA funded major speech-understanding work through the 1970s, including Carnegie Mellon's Harpy.
The 1980s and 1990s were the era of statistical systems built on hidden Markov models. Deep learning reset accuracy in the 2010s, and end-to-end neural networks, capped by OpenAI's Whisper in 2022, brought accurate multilingual speech-to-text within reach of any developer.
Streaming versus batch transcription
Providers expose STT in two modes, and the choice shapes cost, latency, and how you build. Streaming transcription opens a live connection and returns words within a few hundred milliseconds as the person speaks, which is what live captioning, voice AI agents, and call monitoring need.
Batch transcription takes a finished file and returns the full transcript in one pass, trading immediacy for slightly higher accuracy and simpler engineering. Streaming almost always costs more per minute, and some providers price the two modes as separate products.
How speech-to-text is priced
Speech-to-text is usually billed per minute, but that unit is not universal, and comparing on a single number will mislead you. Some providers bill per hour, some per second, and audio-native language models bill per token.
Batch tiers often run far cheaper than the streaming rate, while multilingual or higher-accuracy models can cost more. Always check which model and mode a quoted price refers to before you compare.
What separates speech-to-text providers
Every vendor in this directory does the same core task, so the real differences sit underneath the label. They are accuracy measured by word error rate, latency, supported languages, price, and features like speaker diarization and automatic punctuation.
Knowing that "speech-to-text API," "transcription API," and "speech recognition" all name the same product keeps you comparing on those axes rather than on marketing. For a broader survey of the field, see the overview of speech recognition on Wikipedia.
Frequently asked questions
What is speech-to-text? Speech-to-text is technology that converts spoken words into written text, usually delivered as an API that takes audio and returns a transcript. It is also called voice-to-text or automatic speech recognition.
How does speech-to-text work? The system breaks audio into short frames, extracts numerical features from each, and runs them through a neural model that predicts the word sequence. It then adds punctuation and formatting before returning the text.
Is speech-to-text the same as speech recognition? For practical purposes, yes. "Speech-to-text," "speech recognition," and "automatic speech recognition" name the same task. Watch only for "voice recognition," which can mean identifying who is speaking rather than what they said.
How is speech-to-text priced? Most APIs bill per minute of audio, though some charge per second, per hour, or per token for audio-native models. Streaming usually costs more than batch, so confirm the mode behind any quoted rate.