For AI agents: a documentation index is available at /llms.txt. Markdown versions of all pages can be requested by appending `.md` to the URL, or by setting the `Accept` header to `text/markdown`.
Skip to main content
Speech to Text

Understanding interaction patterns

Speech to Text is described in two ways, for two different readers, and both are correct.

An interaction pattern describes how you present audio for processing. There are three: pre-recorded, streaming, and agent STT. This is the vocabulary used for onboarding and explanation.

A processing mode describes the transport you use to reach the service. There are two: Batch and Realtime. This is the vocabulary used in the API reference and the SDKs, including client names such as BatchClient.

The two systems are not a subset of one another. Two interaction patterns share the Realtime processing mode, which is why "Realtime" alone does not tell you which pattern you are using.

Pre-recorded transcription​

You submit a complete audio file and receive a complete transcript. Processing is asynchronous by default: you submit a job, monitor its progress, and retrieve the result when it finishes. To block for the result in a single request instead, use synchronous transcription.

Pre-recorded transcription suits work where the audio already exists and accuracy matters more than latency:

  • Transcribe recorded meetings or interviews
  • Caption on-demand videos and podcasts
  • Generate searchable transcripts for media archives
  • Process customer service recordings for compliance or insights
  • Automate subtitles across large video libraries

What is a job?​

Each pre-recorded transcription request is handled as a job, a self-contained unit representing a single transcription task.

A job includes:

  • The audio or video file to be transcribed
  • Configuration settings such as language, formatting, and diarization
  • Metadata and status tracking
  • The resulting transcript or transcripts

You submit a job to the API, monitor its progress, and retrieve results once it is complete. You can create a job by uploading a file directly or by referencing a URL.

Streaming transcription​

You open a session, send audio as it is captured, and receive transcripts while the audio is still arriving. Results arrive word by word: interim partial transcripts are refined until a final transcript is emitted.

Streaming transcription suits work where the audio is live and latency matters:

  • Transcribe live broadcasts or events
  • Caption webinars, meetings, or podcasts as they happen
  • Monitor contact center calls in real time
  • Build accessibility features such as live captions

Agent STT​

You open a session and send live audio, as with streaming, but results arrive as complete speaker turns rather than as a word-by-word stream. Turn detection runs on the server, so the transcript you receive is already punctuated, speaker-labeled, and segmented at turn boundaries.

Agent STT suits conversational applications that hand each turn to a language model and need to know when a speaker has finished talking.

Agent STT provides the speech-to-text layer only. It is not a conversational pipeline and does not include a language model or speech synthesis.

How patterns map to the APIs​

This crosswalk is the authoritative mapping between the two naming systems.

Interaction patternProcessing modeTransportsession_type
pre-recordedBatchREST APIbatch
streamingRealtimeWebSocket APIrealtime
agent STTRealtimeWebSocket APIagent

Processing mode, transport, and session_type are fixed by your choice of interaction pattern rather than selected separately.

session_type is a field on the usage endpoint. Its values name interaction patterns, not processing modes: realtime means streaming only, and agent STT reports as agent. Do not read realtime as covering all Realtime traffic.

Next steps​