| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """This script demonstrates how to use the vLLM Realtime WebSocket API to perform |
| 4 | audio transcription by uploading an audio file. |
| 5 | |
| 6 | Before running this script, you must start the vLLM server with a realtime-capable |
| 7 | model, for example: |
| 8 | |
| 9 | vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --enforce-eager |
| 10 | |
| 11 | Requirements: |
| 12 | - vllm with audio support |
| 13 | - websockets |
| 14 | - numpy |
| 15 | |
| 16 | The script: |
| 17 | 1. Connects to the Realtime WebSocket endpoint |
| 18 | 2. Converts an audio file to PCM16 @ 16kHz |
| 19 | 3. Sends audio chunks to the server |
| 20 | 4. Receives and prints transcription as it streams |
| 21 | """ |
| 22 | |
| 23 | import argparse |
| 24 | import asyncio |
| 25 | import json |
| 26 | |
| 27 | import numpy as np |
| 28 | import pybase64 as base64 |
| 29 | import websockets |
| 30 | |
| 31 | from vllm.assets.audio import AudioAsset |
| 32 | from vllm.multimodal.media.audio import load_audio |
| 33 | |
| 34 | |
| 35 | def audio_to_pcm16_base64(audio_path: str) -> str: |
| 36 | """Load an audio file and convert it to base64-encoded PCM16 @ 16kHz.""" |
| 37 | # Load audio and resample to 16kHz mono |
| 38 | audio, _ = load_audio(audio_path, sr=16000, mono=True) |
| 39 | # Convert to PCM16 |
| 40 | pcm16 = (audio * 32767).astype(np.int16) |
| 41 | # Encode as base64 |
| 42 | return base64.b64encode(pcm16.tobytes()).decode("utf-8") |
| 43 | |
| 44 | |
| 45 | async def realtime_transcribe(audio_path: str, host: str, port: int, model: str): |
| 46 | """Connect to the Realtime API and transcribe an audio file.""" |
| 47 | uri = f"ws://{host}:{port}/v1/realtime" |
| 48 | |
| 49 | async with websockets.connect(uri) as ws: |
| 50 | # Wait for session.created |
| 51 | response = json.loads(await ws.recv()) |
| 52 | if response["type"] == "session.created": |
| 53 | print(f"Session created: {response['id']}") |
| 54 | else: |
| 55 | print(f"Unexpected response: {response}") |
| 56 | return |
| 57 | |
| 58 | # Validate model |
| 59 | await ws.send(json.dumps({"type": "session.update", "model": model})) |
| 60 | |
| 61 | # Signal ready to start |
| 62 | await ws.send(json.dumps({"type": "input_audio_buffer.commit"})) |
| 63 | |
| 64 | # Convert audio file to base64 PCM16 |
| 65 | print(f"Loading audio from: {audio_path}") |
| 66 | audio_base64 = audio_to_pcm16_base64(audio_path) |
| 67 | |
| 68 | # Send audio in chunks (4KB of raw audio = ~8KB base64) |
| 69 | chunk_size = 4096 |
| 70 | audio_bytes = base64.b64decode(audio_base64) |
| 71 | total_chunks = (len(audio_bytes) + chunk_size - 1) // chunk_size |
| 72 | |
| 73 | print(f"Sending {total_chunks} audio chunks...") |
| 74 | for i in range(0, len(audio_bytes), chunk_size): |
| 75 | chunk = audio_bytes[i : i + chunk_size] |
| 76 | await ws.send( |
| 77 | json.dumps( |
| 78 | { |
| 79 | "type": "input_audio_buffer.append", |
| 80 | "audio": base64.b64encode(chunk).decode("utf-8"), |
| 81 | } |
| 82 | ) |
| 83 | ) |
| 84 | |
| 85 | # Signal all audio is sent |
| 86 | await ws.send(json.dumps({"type": "input_audio_buffer.commit", "final": True})) |
| 87 | print("Audio sent. Waiting for transcription...\n") |
| 88 | |
| 89 | # Receive transcription |
| 90 | print("Transcription: ", end="", flush=True) |
| 91 | while True: |
| 92 | response = json.loads(await ws.recv()) |
| 93 | if response["type"] == "transcription.delta": |
| 94 | print(response["delta"], end="", flush=True) |
| 95 | elif response["type"] == "transcription.done": |
| 96 | print(f"\n\nFinal transcription: {response['text']}") |
| 97 | if response.get("usage"): |
| 98 | print(f"Usage: {response['usage']}") |
| 99 | break |
| 100 | elif response["type"] == "error": |
| 101 | print(f"\nError: {response['error']}") |
| 102 | break |
| 103 | |
| 104 | |
| 105 | def main(args): |
| 106 | if args.audio_path: |
| 107 | audio_path = args.audio_path |
| 108 | else: |
| 109 | # Use default audio asset |
| 110 | audio_path = str(AudioAsset("mary_had_lamb").get_local_path()) |
| 111 | print(f"No audio path provided, using default: {audio_path}") |
| 112 | |
| 113 | asyncio.run(realtime_transcribe(audio_path, args.host, args.port, args.model)) |
| 114 | |
| 115 | |
| 116 | if __name__ == "__main__": |
| 117 | parser = argparse.ArgumentParser( |
| 118 | description="Realtime WebSocket Transcription Client" |
| 119 | ) |
| 120 | parser.add_argument( |
| 121 | "--model", |
| 122 | type=str, |
| 123 | default="mistralai/Voxtral-Mini-4B-Realtime-2602", |
| 124 | help="Model that is served and should be pinged.", |
| 125 | ) |
| 126 | parser.add_argument( |
| 127 | "--audio_path", |
| 128 | type=str, |
| 129 | default=None, |
| 130 | help="Path to the audio file to transcribe.", |
| 131 | ) |
| 132 | parser.add_argument( |
| 133 | "--host", |
| 134 | type=str, |
| 135 | default="localhost", |
| 136 | help="vLLM server host (default: localhost)", |
| 137 | ) |
| 138 | parser.add_argument( |
| 139 | "--port", |
| 140 | type=int, |
| 141 | default=8000, |
| 142 | help="vLLM server port (default: 8000)", |
| 143 | ) |
| 144 | args = parser.parse_args() |
| 145 | main(args) |