Streaming STT API
Voice-agent ready WebSocket transcription

How it works

1. Connect to wss://api.brainiall.com/v1/stt/stream with Bearer authentication.
2. Send binary frames containing 16 kHz int16 PCM mono audio chunks (any size). Server buffers internally.
3. Receive JSON text frames every ~1.5 s of accumulated audio: {"type":"partial","text":"...","offset_ms":0,"duration_ms":1500}
4. Send text frame "__END__" to flush remaining buffer. Receive final transcript: {"type":"final","text":"...","chunks_processed":N}

# Python client (websockets lib)
import asyncio, json, websockets, wave

async def stream(wav_path):
 async with websockets.connect(
 "wss://api.brainiall.com/v1/stt/stream",
 additional_headers=[("Authorization", "Bearer $BRAINIALL_API_KEY")],
 ) as ws:
 with wave.open(wav_path, "rb") as wf:
 data = wf.readframes(wf.getnframes)
 # Send in 200ms chunks
 chunk = 16000 * 2 // 5 # int16 @ 16kHz, 200ms
 for i in range(0, len(data), chunk):
 await ws.send(data[i:i+chunk])
 await asyncio.sleep(0.2)
 await ws.send("__END__")
 async for msg in ws:
 event = json.loads(msg)
 print(event["type"], event["text"])
 if event["type"] == "final":
 break

asyncio.run(stream("audio.wav"))

Use cases

Quality & price comparison

Quality on a 0–10 scale derived from published WER benchmarks (clean-speech benchmarks, multilingual benchmarks, CallHome). Price is per audio-minute in USD.

Where we're honest

• First-partial latency: ~1.5 s today (buffer-by-time MVP). Deepgram Nova-3 streaming is closer to ~300 ms first-partial. We're working on Silero VAD-based segmentation + 500 ms flush in Phase 2.
• Sliding window: Each partial is transcribed in isolation today (no overlap). The transcription engine occasionally hallucinates on short isolated segments. Phase 2 adds 500 ms overlap to fix this.
• Speaker diarization: Not in streaming mode yet (only batch). Our existing batch /transcribe endpoint includes Brainiall Speaker ID engine diarization.
• Languages: Auto-language detection (good for 99 languages). For tighter latency, future language= parameter will skip detection.

Pricing

Discount derived from quality position. 80% off at parity with the industry leaders on coverage and accuracy.