Alibaba Cloud Model Studio:Real-time speech recognition - Qwen

Java

Install the SDK. Ensure that the DashScope SDK version is 2.22.5 or later.

import com.alibaba.dashscope.audio.omni.*;
import com.alibaba.dashscope.exception.NoApiKeyException;
import com.google.gson.JsonObject;
import org.slf4j.Logger;
import org.slf4j.LoggerFactory;

import javax.sound.sampled.LineUnavailableException;
import java.io.File;
import java.io.FileInputStream;
import java.util.Base64;
import java.util.Collections;
import java.util.concurrent.CountDownLatch;
import java.util.concurrent.atomic.AtomicReference;

public class Qwen3AsrRealtimeUsage {
    private static final Logger log = LoggerFactory.getLogger(Qwen3AsrRealtimeUsage.class);
    private static final int AUDIO_CHUNK_SIZE = 1024; // Audio chunk size in bytes
    private static final int SLEEP_INTERVAL_MS = 30;  // Sleep interval in milliseconds

    public static void main(String[] args) throws InterruptedException, LineUnavailableException {
        CountDownLatch finishLatch = new CountDownLatch(1);

        OmniRealtimeParam param = OmniRealtimeParam.builder()
                .model("qwen3-asr-flash-realtime")
                // The following URL is for the Singapore region. If you use a model in the Beijing region, replace the URL with wss://dashscope.aliyuncs.com/api-ws/v1/realtime.
                .url("wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime")
                // API keys for the Singapore and Beijing regions are different. To get an API key, see https://www.alibabacloud.com/help/zh/model-studio/get-api-key.
                // If you have not configured an environment variable, replace the following line with your Model Studio API key: .apikey("sk-xxx")
                .apikey(System.getenv("DASHSCOPE_API_KEY"))
                .build();

        OmniRealtimeConversation conversation = null;
        final AtomicReference<OmniRealtimeConversation> conversationRef = new AtomicReference<>(null);
        conversation = new OmniRealtimeConversation(param, new OmniRealtimeCallback() {
            @Override
            public void onOpen() {
                System.out.println("connection opened");
            }
            @Override
            public void onEvent(JsonObject message) {
                String type = message.get("type").getAsString();
                switch(type) {
                    case "session.created":
                        System.out.println("start session: " + message.get("session").getAsJsonObject().get("id").getAsString());
                        break;
                    case "conversation.item.input_audio_transcription.completed":
                        System.out.println("transcription: " + message.get("transcript").getAsString());
                        finishLatch.countDown();
                        break;
                    case "input_audio_buffer.speech_started":
                        System.out.println("======VAD Speech Start======");
                        break;
                    case "input_audio_buffer.speech_stopped":
                        System.out.println("======VAD Speech Stop======");
                        break;
                    case "conversation.item.input_audio_transcription.text":
                        System.out.println("transcription: " + message.get("text").getAsString());
                        break;
                    default:
                        break;
                }
            }
            @Override
            public void onClose(int code, String reason) {
                System.out.println("connection closed code: " + code + ", reason: " + reason);
            }
        });
        conversationRef.set(conversation);
        try {
            conversation.connect();
        } catch (NoApiKeyException e) {
            throw new RuntimeException(e);
        }

        OmniRealtimeTranscriptionParam transcriptionParam = new OmniRealtimeTranscriptionParam();
        transcriptionParam.setLanguage("zh");
        transcriptionParam.setInputAudioFormat("pcm");
        transcriptionParam.setInputSampleRate(16000);

        OmniRealtimeConfig config = OmniRealtimeConfig.builder()
                .modalities(Collections.singletonList(OmniRealtimeModality.TEXT))
                .transcriptionConfig(transcriptionParam)
                .build();
        conversation.updateSession(config);

        String filePath = "your_audio_file.pcm";
        File audioFile = new File(filePath);
        if (!audioFile.exists()) {
            log.error("Audio file not found: {}", filePath);
            return;
        }

        try (FileInputStream audioInputStream = new FileInputStream(audioFile)) {
            byte[] audioBuffer = new byte[AUDIO_CHUNK_SIZE];
            int bytesRead;
            int totalBytesRead = 0;

            log.info("Starting to send audio data from: {}", filePath);

            // Read and send audio data in chunks
            while ((bytesRead = audioInputStream.read(audioBuffer)) != -1) {
                totalBytesRead += bytesRead;
                String audioB64 = Base64.getEncoder().encodeToString(audioBuffer);
                // Send audio chunk to conversation
                conversation.appendAudio(audioB64);

                // Add small delay to simulate real-time audio streaming
                Thread.sleep(SLEEP_INTERVAL_MS);
            }

            log.info("Finished sending audio data. Total bytes sent: {}", totalBytesRead);

        } catch (Exception e) {
            log.error("Error sending audio from file: {}", filePath, e);
        }

        // Send session.finish, wait for the session to finish, and then close the connection.
        conversation.endSession();
        log.info("Task finished");

        System.exit(0);
    }
}

Expected output:

connection opened
start session: <session-id>
======VAD Speech Start======
transcription: <intermediate text>
======VAD Speech Stop======
transcription: <final transcribed text>
connection closed code: 1000, reason: ...

Python

Install the SDK. Ensure that the DashScope SDK version is 1.25.6 or later.

import logging
import os
import base64
import signal
import sys
import time
import dashscope
from dashscope.audio.qwen_omni import *
from dashscope.audio.qwen_omni.omni_realtime import TranscriptionParams


def setup_logging():
    """Configure logging."""
    logger = logging.getLogger('dashscope')
    logger.setLevel(logging.DEBUG)
    handler = logging.StreamHandler(sys.stdout)
    handler.setLevel(logging.DEBUG)
    formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
    handler.setFormatter(formatter)
    logger.addHandler(handler)
    logger.propagate = False
    return logger


def init_api_key():
    """Initialize the API key."""
    # API keys for the Singapore and Beijing regions are different. To get an API key, see https://www.alibabacloud.com/help/zh/model-studio/get-api-key.
    # If you have not configured an environment variable, replace the following line with your Model Studio API key: dashscope.api_key = "sk-xxx"
    dashscope.api_key = os.environ.get('DASHSCOPE_API_KEY', 'YOUR_API_KEY')
    if dashscope.api_key == 'YOUR_API_KEY':
        print('[Warning] Using placeholder API key, set DASHSCOPE_API_KEY environment variable.')


class MyCallback(OmniRealtimeCallback):
    """Handle real-time recognition callbacks."""
    def __init__(self, conversation):
        self.conversation = conversation
        self.handlers = {
            'session.created': self._handle_session_created,
            'conversation.item.input_audio_transcription.completed': self._handle_final_text,
            'conversation.item.input_audio_transcription.text': self._handle_stash_text,
            'input_audio_buffer.speech_started': lambda r: print('======Speech Start======'),
            'input_audio_buffer.speech_stopped': lambda r: print('======Speech Stop======')
        }

    def on_open(self):
        print('Connection opened')

    def on_close(self, code, msg):
        print(f'Connection closed, code: {code}, msg: {msg}')

    def on_event(self, response):
        try:
            handler = self.handlers.get(response['type'])
            if handler:
                handler(response)
        except Exception as e:
            print(f'[Error] {e}')

    def _handle_session_created(self, response):
        print(f"Start session: {response['session']['id']}")

    def _handle_final_text(self, response):
        print(f"Final recognized text: {response['transcript']}")

    def _handle_stash_text(self, response):
        print(f"Got stash result: {response['stash']}")


def read_audio_chunks(file_path, chunk_size=3200):
    """Read the audio file in chunks."""
    with open(file_path, 'rb') as f:
        while chunk := f.read(chunk_size):
            yield chunk


def send_audio(conversation, file_path, delay=0.1):
    """Send audio data."""
    if not os.path.exists(file_path):
        raise FileNotFoundError(f"Audio file {file_path} does not exist.")

    print("Processing audio file... Press 'Ctrl+C' to stop.")
    for chunk in read_audio_chunks(file_path):
        audio_b64 = base64.b64encode(chunk).decode('ascii')
        conversation.append_audio(audio_b64)
        time.sleep(delay)

def main():
    setup_logging()
    init_api_key()

    audio_file_path = "./your_audio_file.pcm"
    conversation = OmniRealtimeConversation(
        model='qwen3-asr-flash-realtime',
        # The following URL is for the Singapore region. If you use a model in the Beijing region, replace the URL with wss://dashscope.aliyuncs.com/api-ws/v1/realtime.
        url='wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime',
        callback=MyCallback(conversation=None)  # Temporarily pass None and inject it later.
    )

    # Inject self into the callback.
    conversation.callback.conversation = conversation

    def handle_exit(sig, frame):
        print('Ctrl+C pressed, exiting...')
        conversation.close()
        sys.exit(0)

    signal.signal(signal.SIGINT, handle_exit)

    conversation.connect()

    transcription_params = TranscriptionParams(
        language='zh',
        sample_rate=16000,
        input_audio_format="pcm"
    )

    conversation.update_session(
        output_modalities=[MultiModality.TEXT],
        enable_input_audio_transcription=True,
        transcription_params=transcription_params
    )

    try:
        send_audio(conversation, audio_file_path)
        # Send session.finish, wait for the session to finish, and then close the connection.
        conversation.end_session()
    except Exception as e:
        print(f"Error occurred: {e}")
    finally:
        conversation.close()
        print("Audio processing completed.")


if __name__ == '__main__':
    main()

Python

Install the required dependency:

pip uninstall websocket-client
pip uninstall websocket
pip install websocket-client

Do not name the sample code file websocket.py. Otherwise, the following error may occur: AttributeError: module 'websocket' has no attribute 'WebSocketApp'. Did you mean: 'WebSocket'?

# pip install websocket-client
import os
import time
import json
import threading
import base64
import websocket
import logging
import logging.handlers
from datetime import datetime

logger = logging.getLogger(__name__)
logger.setLevel(logging.DEBUG)

# API keys for the Singapore and Beijing regions are different. To get an API key, see https://www.alibabacloud.com/help/zh/model-studio/get-api-key.
# If you have not configured an environment variable, replace the following line with your Model Studio API key: API_KEY="sk-xxx"
API_KEY = os.environ.get("DASHSCOPE_API_KEY", "sk-xxx")
QWEN_MODEL = "qwen3-asr-flash-realtime"
# The following is the base URL for the Singapore region. If you use a model in the Beijing region, replace the base URL with wss://dashscope.aliyuncs.com/api-ws/v1/realtime.
baseUrl = "wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime"
url = f"{baseUrl}?model={QWEN_MODEL}"
print(f"Connecting to server: {url}")

# Note: If you are not in VAD mode, the cumulative duration of continuously sent audio should not exceed 60 seconds.
enableServerVad = True
is_running = True  # Add a running flag.

headers = [
    "Authorization: Bearer " + API_KEY,
    "OpenAI-Beta: realtime=v1"
]

def init_logger():
    formatter = logging.Formatter('%(asctime)s|%(levelname)s|%(message)s')
    f_handler = logging.handlers.RotatingFileHandler(
        "omni_tester.log", maxBytes=100 * 1024 * 1024, backupCount=3
    )
    f_handler.setLevel(logging.DEBUG)
    f_handler.setFormatter(formatter)

    console = logging.StreamHandler()
    console.setLevel(logging.DEBUG)
    console.setFormatter(formatter)

    logger.addHandler(f_handler)
    logger.addHandler(console)

def on_open(ws):
    logger.info("Connected to server.")

    # Session update event.
    event_manual = {
        "event_id": "event_123",
        "type": "session.update",
        "session": {
            "modalities": ["text"],
            "input_audio_format": "pcm",
            "sample_rate": 16000,
            "input_audio_transcription": {
                # Language identifier, optional. If you have clear language information, set it.
                "language": "zh"
            },
            "turn_detection": None
        }
    }
    event_vad = {
        "event_id": "event_123",
        "type": "session.update",
        "session": {
            "modalities": ["text"],
            "input_audio_format": "pcm",
            "sample_rate": 16000,
            "input_audio_transcription": {
                "language": "zh"
            },
            "turn_detection": {
                "type": "server_vad",
                "threshold": 0.0,
                "silence_duration_ms": 400
            }
        }
    }
    if enableServerVad:
        logger.info(f"Sending event: {json.dumps(event_vad, indent=2)}")
        ws.send(json.dumps(event_vad))
    else:
        logger.info(f"Sending event: {json.dumps(event_manual, indent=2)}")
        ws.send(json.dumps(event_manual))

def on_message(ws, message):
    global is_running
    try:
        data = json.loads(message)
        logger.info(f"Received event: {json.dumps(data, ensure_ascii=False, indent=2)}")
        if data.get("type") == "session.finished":
            logger.info(f"Final transcript: {data.get('transcript')}")
            logger.info("Closing WebSocket connection after session finished...")
            is_running = False  # Stop the audio sending thread.
            ws.close()
    except json.JSONDecodeError:
        logger.error(f"Failed to parse message: {message}")

def on_error(ws, error):
    logger.error(f"Error: {error}")

def on_close(ws, close_status_code, close_msg):
    logger.info(f"Connection closed: {close_status_code} - {close_msg}")

def send_audio(ws, local_audio_path):
    time.sleep(3)  # Wait for the session update to complete.
    global is_running

    with open(local_audio_path, 'rb') as audio_file:
        logger.info(f"Start reading the file: {datetime.now().strftime('%Y-%m-%d %H:%M:%S.%f')[:-3]}")
        while is_running:
            audio_data = audio_file.read(3200)  # ~0.1 second of PCM16/16 kHz audio.
            if not audio_data:
                logger.info(f"Finished reading the file: {datetime.now().strftime('%Y-%m-%d %H:%M:%S.%f')[:-3]}")
                if ws.sock and ws.sock.connected:
                    if not enableServerVad:
                        commit_event = {
                            "event_id": "event_789",
                            "type": "input_audio_buffer.commit"
                        }
                        ws.send(json.dumps(commit_event))
                    finish_event = {
                        "event_id": "event_987",
                        "type": "session.finish"
                    }
                    ws.send(json.dumps(finish_event))
                break

            if not ws.sock or not ws.sock.connected:
                logger.info("The WebSocket is closed. Stop sending audio.")
                break

            encoded_data = base64.b64encode(audio_data).decode('utf-8')
            eventd = {
                "event_id": f"event_{int(time.time() * 1000)}",
                "type": "input_audio_buffer.append",
                "audio": encoded_data
            }
            ws.send(json.dumps(eventd))
            logger.info(f"Sending audio event: {eventd['event_id']}")
            time.sleep(0.1)  # Simulate real-time collection.

# Initialize the logger.
init_logger()
logger.info(f"Connecting to WebSocket server at {url}...")

local_audio_path = "your_audio_file.pcm"
ws = websocket.WebSocketApp(
    url,
    header=headers,
    on_open=on_open,
    on_message=on_message,
    on_error=on_error,
    on_close=on_close
)

thread = threading.Thread(target=send_audio, args=(ws, local_audio_path))
thread.start()
ws.run_forever()

Java

Install the Java-WebSocket dependency.

<dependency>
    <groupId>org.java-websocket</groupId>
    <artifactId>Java-WebSocket</artifactId>
    <version>1.5.6</version>
</dependency>

implementation 'org.java-websocket:Java-WebSocket:1.5.6'

import org.java_websocket.client.WebSocketClient;
import org.java_websocket.handshake.ServerHandshake;
import org.json.JSONObject;

import java.net.URI;
import java.nio.file.Files;
import java.nio.file.Paths;
import java.util.Base64;
import java.util.concurrent.atomic.AtomicBoolean;
import java.util.logging.*;

public class QwenASRRealtimeClient {

    private static final Logger logger = Logger.getLogger(QwenASRRealtimeClient.class.getName());
    // API keys for the Singapore and Beijing regions are different. To get an API key, see https://www.alibabacloud.com/help/zh/model-studio/get-api-key.
    // If you have not configured an environment variable, replace the following line with your Model Studio API key: private static final String API_KEY = "sk-xxx"
    private static final String API_KEY = System.getenv().getOrDefault("DASHSCOPE_API_KEY", "sk-xxx");
    private static final String MODEL = "qwen3-asr-flash-realtime";

    // Controls whether to use VAD mode.
    private static final boolean enableServerVad = true;

    private static final AtomicBoolean isRunning = new AtomicBoolean(true);
    private static WebSocketClient client;

    public static void main(String[] args) throws Exception {
        initLogger();

        // The following is the base URL for the Singapore region. If you use a model in the Beijing region, replace the base URL with wss://dashscope.aliyuncs.com/api-ws/v1/realtime.
        String baseUrl = "wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime";
        String url = baseUrl + "?model=" + MODEL;
        logger.info("Connecting to server: " + url);

        client = new WebSocketClient(new URI(url)) {
            @Override
            public void onOpen(ServerHandshake handshake) {
                logger.info("Connected to server.");
                sendSessionUpdate();
            }

            @Override
            public void onMessage(String message) {
                try {
                    JSONObject data = new JSONObject(message);
                    String eventType = data.optString("type");

                    logger.info("Received event: " + data.toString(2));

                    // When the finish event is received, stop the sending thread and close the connection.
                    if ("session.finished".equals(eventType)) {
                        logger.info("Final transcript: " + data.optString("transcript"));
                        logger.info("Closing WebSocket connection after session finished...");

                        isRunning.set(false); // Stop the audio sending thread.
                        if (this.isOpen()) {
                            this.close(1000, "ASR finished");
                        }
                    }
                } catch (Exception e) {
                    logger.severe("Failed to parse message: " + message);
                }
            }

            @Override
            public void onClose(int code, String reason, boolean remote) {
                logger.info("Connection closed: " + code + " - " + reason);
            }

            @Override
            public void onError(Exception ex) {
                logger.severe("Error: " + ex.getMessage());
            }
        };

        // Add request headers.
        client.addHeader("Authorization", "Bearer " + API_KEY);
        client.addHeader("OpenAI-Beta", "realtime=v1");

        client.connectBlocking(); // Block until the connection is established.

        // Replace with the path of the audio file to recognize.
        String localAudioPath = "your_audio_file.pcm";
        Thread audioThread = new Thread(() -> {
            try {
                sendAudio(localAudioPath);
            } catch (Exception e) {
                logger.severe("Audio sending thread error: " + e.getMessage());
            }
        });
        audioThread.start();
    }

    /** Session update event (enable/disable VAD). */
    private static void sendSessionUpdate() {
        JSONObject eventNoVad = new JSONObject()
                .put("event_id", "event_123")
                .put("type", "session.update")
                .put("session", new JSONObject()
                        .put("modalities", new String[]{"text"})
                        .put("input_audio_format", "pcm")
                        .put("sample_rate", 16000)
                        .put("input_audio_transcription", new JSONObject()
                                .put("language", "zh"))
                        .put("turn_detection", JSONObject.NULL) // Manual mode.
                );

        JSONObject eventVad = new JSONObject()
                .put("event_id", "event_123")
                .put("type", "session.update")
                .put("session", new JSONObject()
                        .put("modalities", new String[]{"text"})
                        .put("input_audio_format", "pcm")
                        .put("sample_rate", 16000)
                        .put("input_audio_transcription", new JSONObject()
                                .put("language", "zh"))
                        .put("turn_detection", new JSONObject()
                                .put("type", "server_vad")
                                .put("threshold", 0.0)
                                .put("silence_duration_ms", 400))
                );

        if (enableServerVad) {
            logger.info("Sending event (VAD):\n" + eventVad.toString(2));
            client.send(eventVad.toString());
        } else {
            logger.info("Sending event (Manual):\n" + eventNoVad.toString(2));
            client.send(eventNoVad.toString());
        }
    }

    /** Send the audio file stream. */
    private static void sendAudio(String localAudioPath) throws Exception {
        Thread.sleep(3000); // Wait for the session to be ready.
        byte[] allBytes = Files.readAllBytes(Paths.get(localAudioPath));
        logger.info("Start reading the file.");

        int offset = 0;
        while (isRunning.get() && offset < allBytes.length) {
            int chunkSize = Math.min(3200, allBytes.length - offset);
            byte[] chunk = new byte[chunkSize];
            System.arraycopy(allBytes, offset, chunk, 0, chunkSize);
            offset += chunkSize;

            if (client != null && client.isOpen()) {
                String encoded = Base64.getEncoder().encodeToString(chunk);
                JSONObject eventd = new JSONObject()
                        .put("event_id", "event_" + System.currentTimeMillis())
                        .put("type", "input_audio_buffer.append")
                        .put("audio", encoded);

                client.send(eventd.toString());
                logger.info("Sending audio event: " + eventd.getString("event_id"));
            } else {
                break; // Avoid sending after disconnection.
            }

            Thread.sleep(100); // Simulate real-time sending.
        }

        logger.info("Finished reading the file.");

        if (client != null && client.isOpen()) {
            // A commit is required in non-VAD mode.
            if (!enableServerVad) {
                JSONObject commitEvent = new JSONObject()
                        .put("event_id", "event_789")
                        .put("type", "input_audio_buffer.commit");
                client.send(commitEvent.toString());
                logger.info("Sent commit event for manual mode.");
            }

            JSONObject finishEvent = new JSONObject()
                    .put("event_id", "event_987")
                    .put("type", "session.finish");
            client.send(finishEvent.toString());
            logger.info("Sent finish event.");
        }
    }

    /** Initialize the logger. */
    private static void initLogger() {
        logger.setLevel(Level.ALL);
        Logger rootLogger = Logger.getLogger("");
        for (Handler h : rootLogger.getHandlers()) {
            rootLogger.removeHandler(h);
        }

        Handler consoleHandler = new ConsoleHandler();
        consoleHandler.setLevel(Level.ALL);
        consoleHandler.setFormatter(new SimpleFormatter());
        logger.addHandler(consoleHandler);
    }
}

Node.js

Install the required dependency:

npm install ws

/**
 * Qwen-ASR Realtime WebSocket Client (Node.js version)
 * Features:
 * - Supports VAD and Manual modes.
 * - Sends session.update to start a session.
 * - Continuously sends input_audio_buffer.append audio chunks.
 * - Sends input_audio_buffer.commit in Manual mode.
 * - Sends a session.finish event.
 * - Closes the connection after receiving a session.finished event.
 */

import WebSocket from 'ws';
import fs from 'fs';

// ===== Configuration =====
// API keys for the Singapore and Beijing regions are different. To get an API key, see https://www.alibabacloud.com/help/zh/model-studio/get-api-key.
// If you have not configured an environment variable, replace the following line with your Model Studio API key: const API_KEY = "sk-xxx"
const API_KEY = process.env.DASHSCOPE_API_KEY || 'sk-xxx';
const MODEL = 'qwen3-asr-flash-realtime';
const enableServerVad = true; // true for VAD mode, false for Manual mode
const localAudioPath = 'your_audio_file.pcm'; // Path to the PCM16, 16 kHz audio file

// The following is the base URL for the Singapore region. If you use a model in the Beijing region, replace the base URL with wss://dashscope.aliyuncs.com/api-ws/v1/realtime.
const baseUrl = 'wss://dashscope-intl.aliyuncs.com/api-ws/v1/realtime';
const url = `${baseUrl}?model=${MODEL}`;

console.log(`Connecting to server: ${url}`);

// ===== Status Control =====
let isRunning = true;

// ===== Establish Connection =====
const ws = new WebSocket(url, {
    headers: {
        'Authorization': `Bearer ${API_KEY}`,
        'OpenAI-Beta': 'realtime=v1'
    }
});

// ===== Event Binding =====
ws.on('open', () => {
    console.log('[WebSocket] Connected to server.');
    sendSessionUpdate();
    // Start the audio sending thread.
    sendAudio(localAudioPath);
});

ws.on('message', (message) => {
    try {
        const data = JSON.parse(message);
        console.log('[Received Event]:', JSON.stringify(data, null, 2));

        // Received finish event.
        if (data.type === 'session.finished') {
            console.log(`[Final Transcript] ${data.transcript}`);
            console.log('[Action] Closing WebSocket connection after session finished...');

            if (ws.readyState === WebSocket.OPEN) {
                ws.close(1000, 'ASR finished');
            }
        }
    } catch (e) {
        console.error('[Error] Failed to parse message:', message);
    }
});

ws.on('close', (code, reason) => {
    console.log(`[WebSocket] Connection closed: ${code} - ${reason}`);
});

ws.on('error', (err) => {
    console.error('[WebSocket Error]', err);
});

// ===== Session Update =====
function sendSessionUpdate() {
    const eventNoVad = {
        event_id: 'event_123',
        type: 'session.update',
        session: {
            modalities: ['text'],
            input_audio_format: 'pcm',
            sample_rate: 16000,
            input_audio_transcription: {
                language: 'zh'
            },
            turn_detection: null
        }
    };

    const eventVad = {
        event_id: 'event_123',
        type: 'session.update',
        session: {
            modalities: ['text'],
            input_audio_format: 'pcm',
            sample_rate: 16000,
            input_audio_transcription: {
                language: 'zh'
            },
            turn_detection: {
                type: 'server_vad',
                threshold: 0.0,
                silence_duration_ms: 400
            }
        }
    };

    if (enableServerVad) {
        console.log('[Send Event] VAD Mode:\n', JSON.stringify(eventVad, null, 2));
        ws.send(JSON.stringify(eventVad));
    } else {
        console.log('[Send Event] Manual Mode:\n', JSON.stringify(eventNoVad, null, 2));
        ws.send(JSON.stringify(eventNoVad));
    }
}

// ===== Send Audio File Stream =====
function sendAudio(audioPath) {
    setTimeout(() => {
        console.log(`[File Read Start] ${audioPath}`);
        const buffer = fs.readFileSync(audioPath);

        let offset = 0;
        const chunkSize = 3200; // Approx. 0.1 second of PCM16 audio

        function sendChunk() {
            if (!isRunning) return;
            if (offset >= buffer.length) {
                isRunning = false; // Stop sending audio.
                console.log('[File Read End]');
                if (ws.readyState === WebSocket.OPEN) {
                    if (!enableServerVad) {
                        const commitEvent = {
                            event_id: 'event_789',
                            type: 'input_audio_buffer.commit'
                        };
                        ws.send(JSON.stringify(commitEvent));
                        console.log('[Send Commit Event]');
                    }

                    const finishEvent = {
                        event_id: 'event_987',
                        type: 'session.finish'
                    };
                    ws.send(JSON.stringify(finishEvent));
                    console.log('[Send Finish Event]');
                }

                return;
            }

            if (ws.readyState !== WebSocket.OPEN) {
                console.log('[Stop] WebSocket is not open.');
                return;
            }

            const chunk = buffer.slice(offset, offset + chunkSize);
            offset += chunkSize;

            const encoded = chunk.toString('base64');
            const appendEvent = {
                event_id: `event_${Date.now()}`,
                type: 'input_audio_buffer.append',
                audio: encoded
            };

            ws.send(JSON.stringify(appendEvent));
            console.log(`[Send Audio Event] ${appendEvent.event_id}`);

            setTimeout(sendChunk, 100); // Simulate real-time sending.
        }

        sendChunk();
    }, 3000); // Wait for session configuration to complete.
}