Streaming transcription

WSwss://api.lucidweights.ai/v1/stt/stream

Real-time streaming transcription. The server detects end-of-speech and emits a final transcript per utterance. You do not need to signal turn boundaries.

Connect

wss://api.lucidweights.ai/v1/stt/stream?api_key=KEY&sample_rate=16000&languages=en

Query parameters

Field	Type	Default	Description
api_key	string	required	Required in the query string, not a header
sample_rate	int	16000	Sample rate of the PCM you send
backend	string	ultra	ultra, fast, or normal
model_name	string	large	For fast/normal backends
languages	string	""	Comma-separated ISO codes
enable_interim_transcript	bool	false	Emit non-final partial results
silence_threshold	int (ms)	800	Trailing silence that ends a turn
vad_threshold	float	0.6	Speech-detection sensitivity (higher = stricter)
enable_turn_taker	bool	false	Semantic end-of-turn model. Ends turns sooner than silence alone
turn_taker_threshold	float	0.55	Confidence to end a turn early
wait_end_digit	int (ms)	500	Extra hold when transcript ends in a number
audio_gain	float	1	Linear gain before processing, clipped at ±1.0
auto_detect_sample_rate	bool	false	Infer rate from packet timing

Sending audio

JSON text frames only. Binary frames are not read.

Field	Type	Default	Description
audio_data	string	required	Base64 of 16-bit signed LE PCM, mono, at sample_rate
client_send_timestamp	int	-	Client epoch ms. Required for meaningful ctos_latency on finals

{
  "audio_data": "<base64 int16 PCM>",
  "client_send_timestamp": 1721470000000
}

There is no end-of-stream message. Close the socket. Pending audio is flushed after silence_threshold ms of inactivity.

Receiving transcripts

Interim (when enable_interim_transcript=true):

{ "transcription": "hello wor", "final": false }

Final:

{
  "transcription": "hello world",
  "final": true,
  "send_timestamp_ms": 1721470001234,
  "ctos_latency": 42,
  "vad_run_count": 0,
  "total_vad_proccesing_time": 0
}

NOTE

ctos_latency appears only on turn-end finals, not inactivity flushes. Empty transcripts are never sent.

WARNING

No µ-law mode on the stream. G.711 callers must expand to PCM16 client-side. Mono 16-bit signed integer only. Three consecutive failed utterances end the stream with close code 1011.

Example

import asyncio, base64, json, time, websockets

URL = (f"wss://api.lucidweights.ai/v1/stt/stream"
       f"?api_key=YOUR_KEY&sample_rate=16000&languages=en")

async def main():
    async with websockets.connect(URL) as ws:
        async def send():
            for chunk in pcm16_chunks():  # bytes, 16-bit LE mono
                await ws.send(json.dumps({
                    "audio_data": base64.b64encode(chunk).decode(),
                    "client_send_timestamp": int(time.time() * 1000),
                }))

        async def recv():
            async for msg in ws:
                d = json.loads(msg)
                if d.get("final"):
                    print("FINAL:", d["transcription"])

        await asyncio.gather(send(), recv())

asyncio.run(main())