Speech to Text

ASAS STT transcribes Arabic audio, including Gulf dialects and MSA. Short audio is transcribed synchronously; long audio becomes an async job whose result is delivered by webhook. A realtime WebSocket streams partial and final results as you speak.

Endpoints

Batch transcription

Upload an audio file (WAV/MP3/M4A). Short audio returns the transcript directly; long audio returns 202 with a job_id.

curl https://api.asas.flitc.tech/v1/stt/transcriptions \
  -H "Authorization: Bearer $ASAS_API_KEY" \
  -F file=@meeting.mp3 \
  -F language=ar \
  -F diarize=true

Synchronous response (short audio):

{
  "text": "النص الكامل للتفريغ…",
  "language": "ar",
  "duration_seconds": 63.4,
  "segments": [
    { "id": 0, "start": 0.0, "end": 4.2, "text": "…", "confidence": 0.94, "speaker": "spk_0" }
  ]
}

For long audio you get a job, then subscribe to the job.completed webhook (or poll GET /v1/stt/jobs/{id}) instead of holding a request open.

Parameters

Realtime WebSocket

Open a socket (authenticate with a ws-token), send a JSON start frame, then stream binary PCM audio. You receive partial (mutable) and final (committed) result frames.

// client → server
{ "type": "start", "encoding": "pcm_s16le", "sample_rate": 16000, "language": "ar", "interim_results": true }

// server → client
{ "type": "final", "text": "…", "start": 0.0, "end": 4.2, "duration_so_far": 4.2 }

Send { "action": "keepalive" } during silence, and { "action": "close_stream" } to finalize. The server replies with a completed frame carrying the authoritative duration_seconds.

Billing

STT bills by decoded audio_seconds, reported by the engine — never by the bytes or sample rate you declare. Streams are metered in checkpoints (roughly every 30–60 seconds of audio) with a final true-up on completion, so a dropped connection loses at most one checkpoint. Failed jobs and upstream errors are not billed.