Speech-to-text (STT)
Speech-to-text (STT) is technology that converts spoken audio into written words that people or software can read, store, search, and analyze.
The process begins with an audio stream or recording. A speech-recognition system identifies speech sounds, maps them to likely words, and produces text, often with timestamps and punctuation. Depending on the application, the output may be a live partial transcript that changes as the speaker continues or a final transcript produced after more context is available.
Speech-to-text and automatic speech recognition are commonly treated as synonyms. When a distinction is useful, ASR names the recognition system or capability, while STT emphasizes the textual result. That result can appear directly as captions or a transcript, or it can serve as input to another system that detects intent, extracts details, summarizes a call, or chooses a response.
For an AI phone agent, STT creates the text representation of the caller’s side of the conversation. The agent can use it to follow the request and preserve context across turns. The transcript can also support later review and analysis. However, a transcript is an interpretation of the audio, not a perfect record. Names, numbers, specialized vocabulary, and speech in noisy conditions are common sources of error.
Good call design accounts for that uncertainty. The agent should confirm details that affect an action, such as a phone number, date, address, or account identifier. If the text is incomplete or contradictory, it should ask a focused follow-up question rather than filling the gap with an assumption. For sensitive workflows, teams should also decide whether a transcript should be created, who may access it, and how long it should be retained.
When evaluating STT for phone calls, use audio that resembles actual traffic. Include different speaking styles, accents, connection quality, background sounds, interruptions, and callers spelling unfamiliar words. Review whether errors change the meaning or outcome, not just the percentage of words transcribed correctly. A system can produce readable prose yet still fail on the few details the workflow needs most.