Deepgram 官方 JavaScript/TypeScript SDK 发布

The official JavaScript/TypeScript SDK for Deepgram's automated speech recognition, text-to-speech, and language understanding APIs. Power your applications with world-class speech and Language AI models.

Documentation

Comprehensive API documentation and guides are available at developers.deepgram.com.

• Flux TTS Controls - SDK-specific batch and WebSocket guidance for speed, pauses, and pronunciation

Migrating From Earlier Versions

• v2 to v3
• v3 to v4
• v4 to v5 (current)

Installation

Install the Deepgram JavaScript SDK using npm:

npm install @deepgram/sdk

Reference

• REST API Reference - Fern-generated methods and parameters for HTTP endpoints

Streaming WebSocket API

Every streaming client exposes both connect() and createConnection(). They are aliases that return a start-closed socket: register handlers, call the socket's connect(), and then await waitForOpen() before sending data.

All sockets expose on("open" | "message" | "close" | "error", callback), off(event, callback), connect(), waitForOpen(), close(), readyState, and a lower-level socket property. Registering on() again for the same event replaces the prior callback. Prefer the typed send methods below instead of calling socket.send() directly.

Pass shouldReconnect: (event) => boolean when a connection needs a custom close policy. Native Flux Listen V2 automatically treats the server no-status (1005) close after sendCloseStream() as terminal; other native socket closes retain the default retry behavior. Custom transports disable wrapper retries by default; with reconnect: true, a 1005 close after CloseStream or TTS Close is terminal unless shouldReconnect overrides it.

The public connection argument types and Deepgram-specific wrapper behavior are defined in src/CustomClient.ts. Use the linked socket classes for exact message and event types.

For Voice Agent function calls, do not send FunctionCallResponse for an ID listed in a FunctionCallCancelled event. Set defer_until_eot: true on an agent.think.functions entry when its action cannot be undone: the agent waits until the user's turn is confirmed, and discards the deferred call if that turn resumes.

Usage

Quick Start

The Deepgram SDK provides clients for all major use cases:

Real-time Speech Recognition (Listen v1)

Connect to our WebSocket and transcribe live streaming audio:

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();

const connection = await client.listen.v1.connect({
  model: "nova-3",
  language: "en",
  punctuate: "true",
  interim_results: "true",
});

connection.on("open", () => console.log("Connection opened"));

connection.on("message", (data) => {
  if (data.type === "Results") {
    console.log(data);
  }
});

connection.connect();
await connection.waitForOpen();

// Send audio data
connection.sendMedia(audioData);

Pass an abortSignal to stop a connection attempt or active session and disable automatic reconnection. If you await waitForOpen(), make that wait abort-aware as shown in Canceling a WebSocket Connection (AbortSignal).

Every streaming connection is also an async iterable, so it can be consumed with for await instead of a callback:

const connection = await client.listen.v1.connect({ model: "nova-3", interim_results: "true" });

connection.connect();
await connection.waitForOpen();

for await (const message of connection) {
  if (message.type === "Results" && message.is_final) {
    console.log(message.channel.alternatives[0].transcript);
  }
}

Iteration ends when the connection closes, throws if it errors, and break closes the connection. Messages that arrive while your loop body is still running are buffered rather than dropped, and on("message", ...) keeps working alongside iteration if you want both. Each connection supports one active iterator. Its queue is capped at 1,000 messages or 16 MiB; an overflow throws and closes the connection rather than retaining unbounded audio in memory. This works on listen.v1, listen.v2, agent.v1, speak.v1 and speak.v2; on the sockets that carry audio, binary frames are delivered as a Blob.

File Transcription

Transcribe pre-recorded audio files (API Reference):

import { createReadStream } from "fs";
import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();

const response = await client.listen.v1.media.transcribeFile(
  createReadStream("audio.wav"),
  { model: "nova-3" }
);
console.log(response.results.channels[0].alternatives[0].transcript);

Text-to-Speech

Generate natural-sounding speech from text (API Reference):

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();

const response = await client.speak.v1.audio.generate({
  text: "Hello, this is a sample text to speech conversion.",
  model: "aura-2-thalia-en",
  encoding: "linear16",
  container: "wav",
});

// Save the audio file
const stream = response.stream();

Text Analysis

Analyze text for sentiment, topics, and intents (API Reference):

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();

const response = await client.read.v1.text.analyze({
  text: "Hello, world!",
  language: "en",
});

Voice Agent (Conversational AI)

Build interactive voice agents:

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();

const connection = await client.agent.v1.connect();

connection.on("open", () => console.log("Connection opened"));

connection.on("message", (data) => {
  if (data.type === "ConversationText") {
    console.log(data);
  }
});

connection.connect();
await connection.waitForOpen();

connection.sendSettings({
  type: "Settings",
  audio: {
    input: { encoding: "linear16", sample_rate: 24000 },
    output: { encoding: "linear16", sample_rate: 16000, container: "wav" },
  },
  agent: {
    language: "en",
    listen: {
      provider: { type: "deepgram", version: "v1", model: "nova-3" },
    },
    think: {
      provider: { type: "open_ai", model: "gpt-4o-mini" },
      prompt: "You are a friendly AI assistant.",
    },
    speak: {
      provider: { type: "deepgram", model: "aura-2-thalia-en" },
    },
  },
});

Authentication

The Deepgram SDK supports two authentication methods:

API Key Authentication

Use your Deepgram API key for server-side applications:

import { DeepgramClient } from "@deepgram/sdk";

// Explicit API key
const client = new DeepgramClient({ apiKey: "YOUR_API_KEY" });

// Or via environment variable DEEPGRAM_API_KEY
const client = new DeepgramClient();

Access Token Authentication

Use access tokens for temporary or scoped access (recommended for client-side applications):

import { DeepgramClient } from "@deepgram/sdk";

// Explicit access token
const client = new DeepgramClient({ accessToken: "YOUR_ACCESS_TOKEN" });

// Or via environment variable DEEPGRAM_ACCESS_TOKEN
const client = new DeepgramClient();

// Generate access tokens using your API key
const authClient = new DeepgramClient({ apiKey: "YOUR_API_KEY" });
const tokenResponse = await authClient.auth.v1.tokens.grant();
const tokenClient = new DeepgramClient({ accessToken: tokenResponse.access_token });

Environment Variables

The SDK automatically discovers credentials from these environment variables:

• DEEPGRAM_ACCESS_TOKEN - Your access token (takes precedence)
• DEEPGRAM_API_KEY - Your Deepgram API key

Precedence: Explicit parameters > Environment variables

Getting an API Key

To access the Deepgram API you will need a free Deepgram API Key.

Browser Usage

The SDK works in modern browsers with some considerations:

WebSocket Features (Full Support)

• Live Transcription: Direct connection to wss://api.deepgram.com
• Voice Agent: Direct connection to wss://agent.deepgram.com
• Live Text-to-Speech: Direct connection to wss://api.deepgram.com

REST API Features (Proxy Required)

Due to CORS header restrictions in the Deepgram API, you must use a proxy server when making REST API calls from browsers. Pass "proxy" as your API key and point baseUrl to your proxy:

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient({
  apiKey: "proxy",
  baseUrl: "http://localhost:8080",
});

Your proxy must set the Authorization: token DEEPGRAM_API_KEY header and forward requests to Deepgram's API. See our example Deepgram Node Proxy.

Setup Options

<!-- CDN (UMD) -->
<script src="https://cdn.jsdelivr.net/npm/@deepgram/sdk"></script>
<script>
  const { DeepgramClient } = deepgram;
</script>

<!-- CDN (ESM) -->
<script type="module">
  import { DeepgramClient } from "https://cdn.jsdelivr.net/npm/@deepgram/sdk/+esm";
</script>

Exception Handling

When the API returns a non-success status code (4xx or 5xx), a DeepgramError is thrown:

import { DeepgramClient, DeepgramError } from "@deepgram/sdk";

const client = new DeepgramClient();

try {
  await client.listen.v1.media.transcribeFile(audioData, { model: "nova-3" });
} catch (err) {
  if (err instanceof DeepgramError) {
    console.log(err.statusCode);
    console.log(err.message);
    console.log(err.body);
  }
}

Request And Response Types

The SDK exports all request and response types as TypeScript interfaces:

// Direct import (recommended)
import { ListenV1Response, SpeakV1Response } from "@deepgram/sdk";

// Or via namespace
import { Deepgram } from "@deepgram/sdk";
type Response = Deepgram.ListenV1Response;

Advanced Features

Request Configuration

Configure timeouts, retries, and other request options:

const response = await client.listen.v1.media.transcribeFile(audioData, {
  model: "nova-3",
  timeoutInSeconds: 60,
  maxRetries: 3,
});

Canceling a WebSocket Connection (AbortSignal)

All real-time connect() methods — listen.v1.connect(), listen.v2.connect(), agent.v1.connect(), speak.v1.connect(), and speak.v2.connect() — accept an abortSignal. Use it when cancellation must be controlled outside the connection owner, such as in apps that start and stop sessions rapidly.

When the signal aborts, the SDK stops the pending or active transport and disables automatic reconnection. A registered close callback can run as part of cancellation. AbortSignal does not clear callbacks registered with connection.on(), and waitForOpen() does not observe the signal directly. If a connection can be canceled while opening, make the wait abort-aware:

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient();
const controller = new AbortController();

function waitForOpenOrAbort(
  connection: { waitForOpen(): Promise<unknown> },
  signal: AbortSignal,
): Promise<void> {
  return new Promise((resolve, reject) => {
    const cleanup = () => signal.removeEventListener("abort", onAbort);
    const onAbort = () => {
      cleanup();
      reject(signal.reason ?? new Error("Connection aborted"));
    };

    if (signal.aborted) {
      onAbort();
      return;
    }

    signal.addEventListener("abort", onAbort, { once: true });
    connection.waitForOpen().then(
      () => {
        cleanup();
        resolve();
      },
      (error) => {
        cleanup();
        reject(error);
      },
    );
  });
}

const connection = await client.listen.v1.connect({
  model: "nova-3",
  language: "en",
  abortSignal: controller.signal,
});

connection.on("open", () => console.log("Connection opened"));
connection.on("message", (data) => console.log(data));

connection.connect();
await waitForOpenOrAbort(connection, controller.signal);

// Cancel the session later. If this happens before open, the helper above rejects.
controller.abort();

An aborted signal is terminal for that connection. To start a new session, create a fresh AbortController and connection. AbortSignal does not provide removeAllListeners(); manage the lifecycle of callbacks registered on the old connection separately.

Access Raw Response Data

const { data, rawResponse } = await client.listen.v1.media
  .transcribeFile(audioData, { model: "nova-3" })
  .withRawResponse();

console.log(rawResponse.headers["X-My-Header"]);

Custom Fetch Client

Use a custom fetch implementation for unsupported environments:

import { DeepgramClient } from "@deepgram/sdk";

const client = new DeepgramClient({
  apiKey: "YOUR_API_KEY",
  fetcher: yourCustomFetchImplementation,
});

WebSocket Proxy / Custom HTTP Agent

In Node.js you can route streaming WebSocket connections (listen, speak, agent) through any compatible custom http.Agent implementation. For example, install a Node 18-compatible version of https-proxy-agent:

npm install https-proxy-agent@8

Then create an HttpsProxyAgent to route traffic through a corporate HTTP/HTTPS egress proxy. Set agent on the client to apply it to every connection, or pass it per connection to override the client-level default:

import { DeepgramClient } from "@deepgram/sdk";
import { HttpsProxyAgent } from "https-proxy-agent";

const agent = new HttpsProxyAgent(process.env.HTTPS_PROXY!);

// Applies to every streaming connection from this client.
const client = new DeepgramClient({ apiKey: "YOUR_API_KEY", agent });

// ...or per connection (overrides the client-level agent).
const socket = await client.listen.v1.createConnection({ model: "nova-3", agent });

The agent option is Node-only; it is ignored in browser and web-worker runtimes, which use the native WebSocket and cannot accept a custom agent.

Logging

import { DeepgramClient, logging } from "@deepgram/sdk";

const client = new DeepgramClient({
  apiKey: "YOUR_API_KEY",
  logging: {
    level: logging.LogLevel.Debug,
    logger: new logging.ConsoleLogger(),
    silent: false,
  },
});

Runtime Compatibility

The SDK works in the following runtimes:

• Node.js 18+
• Vercel
• Cloudflare Workers
• Deno v1.25+
• Bun 1.0+
• React Native

Contributing

We welcome contributions to improve this SDK! However, please note that this library is primarily generated from our API specifications.

Development Setup

1. Install dependencies:

   pnpm install

2. Build:

   make build

3. Run tests:

   make test

Contribution Guidelines

See our CONTRIBUTING guide.

Backwards Compatibility

Older SDK versions will receive Priority 1 (P1) bug support only. Security issues, both in our code and dependencies, are promptly addressed. Significant bugs without clear workarounds are also given priority attention.

Getting Help

We love to hear from you so if you have questions, comments or find a bug in the project, let us know!

• Open an issue on GitHub
• Join the Deepgram Discord Community
• Join the Deepgram GitHub Discussions

Community Code of Conduct

Please see our community code of conduct before contributing to this project.

License

This project is licensed under the MIT License - see the LICENSE file for details.