Whisper Large V3 (Streaming)

1"""
2Streaming ASR — Minimal Example
3
4Stream audio from your laptop's microphone to the Baseten Streaming ASR
5WebSocket endpoint and print transcription results in real time.
6
7Requirements:
8    pip install websockets sounddevice numpy
9
10Usage:
11    export BASETEN_API_KEY="your-api-key"
12    python call.py --model-id <YOUR_MODEL_ID>
13"""
14
15import argparse
16import asyncio
17import json
18import os
19
20import numpy as np
21import sounddevice as sd
22import websockets
23
24# Audio config
25SAMPLE_RATE = 16000
26CHUNK_SIZE = 512
27CHANNELS = 1
28
29headers = {"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"}
30
31# Metadata sent as the first message — must match StreamingWhisperInput schema
32metadata = {
33    "whisper_params": {
34        "audio_language": "en",
35    },
36    "streaming_params": {
37        "encoding": "pcm_s16le",
38        "sample_rate": SAMPLE_RATE,
39        "enable_partial_transcripts": True,
40    },
41    "streaming_vad_config": {
42        "threshold": 0.5,
43        "min_silence_duration_ms": 300,
44        "speech_pad_ms": 30,
45    },
46}
47
48
49async def stream_microphone_audio(ws_url):
50    loop = asyncio.get_running_loop()
51    async with websockets.connect(ws_url, additional_headers=headers) as ws:
52        print("Connected to server")
53
54        await ws.send(json.dumps(metadata))
55        print("Sent metadata")
56
57        send_queue = asyncio.Queue()
58
59        def audio_callback(indata, frames, time_info, status):
60            if status:
61                print(f"Audio warning: {status}")
62            int16_data = (indata * 32767).astype(np.int16).tobytes()
63            loop.call_soon_threadsafe(send_queue.put_nowait, int16_data)
64
65        with sd.InputStream(
66            samplerate=SAMPLE_RATE,
67            blocksize=CHUNK_SIZE,
68            channels=CHANNELS,
69            dtype="float32",
70            callback=audio_callback,
71        ):
72            print("Streaming mic audio… (Ctrl+C to stop)\n")
73
74            async def send_audio():
75                while True:
76                    chunk = await send_queue.get()
77                    await ws.send(chunk)
78
79            async def receive_messages():
80                while True:
81                    response = await ws.recv()
82                    try:
83                        message = json.loads(response)
84                        msg_type = message.get("type")
85
86                        if msg_type == "transcription":
87                            is_final = message.get("is_final", False)
88                            segments = message.get("segments", [])
89                            text = " ".join(
90                                seg.get("text", "") for seg in segments
91                            ).strip()
92
93                            if is_final:
94                                print(f"[final]   {text}")
95                            else:
96                                print(f"[partial] {text}")
97
98                        elif msg_type == "end_audio":
99                            print("[end_audio]", message.get("body"))
100
101                        else:
102                            print(f"[{msg_type}]", message)
103
104                    except Exception as e:
105                        print("Parse error:", e, "| raw:", response[:200])
106
107            await asyncio.gather(send_audio(), receive_messages())
108
109
110if __name__ == "__main__":
111    parser = argparse.ArgumentParser()
112    parser.add_argument("--model-id", required=True, help="Baseten model ID")
113    args = parser.parse_args()
114
115    ws_url = f"wss://model-{args.model_id}.api.baseten.co/environments/production/websocket"
116    print(f"Connecting to: {ws_url}")
117
118    asyncio.run(stream_microphone_audio(ws_url))