Speech-to-text, in real time.
Streaming transcription over WebSocket with interim and committed results, plus a synchronous upload route for files. Audio is never stored — the transcript is the deliverable.
Transcripts as the words arrive
Open a WebSocket, send raw PCM at 16 kHz, and read partial frames while the caller is still speaking — then the final when the utterance closes.
Everything Speech-to-Text API does
Transcribe live over a WebSocket, or POST a WAV and get the transcript back in one call.
Partial hypotheses arrive as they are formed; filter on final for text that will not be revised.
Each segment carries start and end in seconds, so a transcript can be aligned back to the audio.
Submitted audio is never persisted, on either transport. The recording is usually of a third party.
Accents, code-switching and background noise.
mono or stereo
Documented behaviour, not inference tricks.
Interim hypotheses, then committed text.
Takes what a phone line gives you
Stereo is downmixed and anything off 16 kHz is resampled server-side. Audio that ends mid-syllable still returns text — a second of silence is appended before end-of-audio, unbilled, so the recogniser commits.
Listen. Reason. Speak. In real time. One AI platform for every conversation.