Skip to navigation

Speech to Text

Stream live transcripts or transcribe finished audio files.

Kotoba’s Speech-to-Text (ASR) lets you turn audio into text in either of two complementary modes:

  • Live (WebSocket) — push PCM16 chunks as they arrive, read transcript deltas back the same connection. Built for microphones and any latency-sensitive pipeline where the first words matter before the utterance ends. Deployments with a diarization service also attribute the live transcript to speakers (on fal: kotoba-streaming-stt).
  • File (REST) — POST an audio file and get the transcript back in the response, optionally with per-segment timestamps and speaker attribution. Best for offline workflows and anything that doesn’t need partial results. A long file goes in the same request: the server transcribes it in pieces and joins the text. Private endpoints add an asynchronous job API.

Supported input languages:

ModeLanguages
Live (WebSocket)Japanese (ja)
File (REST)Japanese (ja), English (en), Korean (ko), Chinese (zh)

Pick a transport

You want…Use
Mic in, text out, while audio is flowingLive (WebSocket)
File in, full transcript outFile (REST)
Per-segment timestampsREST with with_timestamps=True (either mode)
Who said what (speaker diarization)REST with diarization=True (deployments whose model carries speaker tokens)
Who is speaking, liveWebSocket with diarization=True (deployments with a diarization service, kotoba-sdk >= 0.3.2)
Long files (meetings, interviews)File (REST): audio over 120 s is transcribed in pieces, up to the upload limit (500 MiB by default)
Asynchronous transcription jobsREST job API (batch=True, private endpoints)

Where to go next