@apicity/elevenlabs

v0.7.3

Published

3 days ago

ElevenLabs provider for sound effect generation, text-to-speech, and audio APIs.

0High
0Medium
0Low

justintanner

ai elevenlabs sound-effects sfx text-to-speech tts audio apicity

@apicity/elevenlabs

ElevenLabs provider for sound effect generation, text-to-speech, and audio APIs.

Runtime dependencies:

zod@^4.4.3 — request schemas attached to every POST endpoint as .schema

Installation

npm install @apicity/elevenlabs
# or
pnpm add @apicity/elevenlabs

Quick Start

import { createElevenLabs } from "@apicity/elevenlabs";

const elevenlabs = createElevenLabs({ apiKey: process.env.ELEVENLABS_API_KEY! });

Real-world example: generate a sound effect, then run it through Scribe v2

ElevenLabs' two flagship audio surfaces fit together cleanly: text-to- sound-effects spits out raw MP3 bytes, and Scribe v2 hands back a typed transcript with word-level timestamps plus tagged audio events. The round-trip below — generate a UI click, then transcribe a separate clip with tag_audio_events: true — mirrors what tests/integration/elevenlabs-sound-generation.test.ts and tests/integration/elevenlabs-speech-to-text.test.ts replay against tests/recordings/elevenlabs_2379486140/, so every payload, response field, and byte count below comes straight from the recorded HARs.

import { readFileSync, writeFileSync } from "node:fs";
import { createElevenLabs } from "@apicity/elevenlabs";
import type { ElevenLabsTranscript } from "@apicity/elevenlabs";

const elevenlabs = createElevenLabs({ apiKey: process.env.ELEVENLABS_API_KEY! });

// 1. Generate a 0.5s UI click. soundGeneration returns the raw MP3 as
//    an ArrayBuffer — there's no JSON wrapper, the response body is
//    audio/mpeg straight off the wire. duration_seconds (0.5–30) caps
//    the clip length and prompt_influence (0–1) trades prompt-fidelity
//    for creative variation. The factory also accepts `output_format`
//    on the same request object and silently moves it to the URL query.
const audio = await elevenlabs.v1.soundGeneration({
  text: "soft ui click",
  duration_seconds: 0.5,
  prompt_influence: 0.3,
});

writeFileSync("./click.mp3", new Uint8Array(audio));
console.log(`Generated ${audio.byteLength} bytes of audio/mpeg`);
// → "Generated 11764 bytes of audio/mpeg"
//   ElevenLabs charged 10 characters for this call (visible in the
//   `character-cost` response header on the original request).

// 2. Transcribe a separate audio clip with Scribe v2. The request goes
//    up as multipart/form-data — pass a Blob and the rest as ergonomic
//    fields; the factory packs the form, sets xi-api-key, and parses
//    the JSON response. tag_audio_events: true tells Scribe to surface
//    non-speech events ([phone beeping], [laughter], [applause]) inline
//    with words instead of dropping them.
const phoneBeep = readFileSync("./phone-beeping.mp3"); // 2,528 bytes
const file = new Blob([phoneBeep], { type: "audio/mp3" });

const result = (await elevenlabs.v1.speechToText({
  file,
  model_id: "scribe_v2",
  language_code: "eng",
  tag_audio_events: true,
})) as ElevenLabsTranscript;

// 3. The transcript is rich. `text` is the human-readable form;
//    `words` is the per-token breakdown with absolute timestamps and a
//    `type` discriminator ("word" | "spacing" | "audio_event") plus a
//    `logprob` confidence. `transcription_id` is durable — you can
//    retrieve the same transcript later through the history API.
console.log(
  `${result.language_code} · ${(result.language_probability * 100).toFixed(0)}% confident`,
);
// → "eng · 100% confident"
console.log(
  `${result.audio_duration_secs}s · transcription_id=${result.transcription_id}`,
);
// → "0.5s · transcription_id=CeeidI2QJ8kkN1mcq8HX"
console.log(result.text);
// → "[phone beeping]"

// 4. Walk the words array, splitting audio events from spoken words
//    via the `type` discriminator. On a clip with no speech every
//    entry is an audio_event; on real speech you'll see "word" and
//    "spacing" entries interleaved with bracketed events.
for (const w of result.words) {
  const tag =
    w.type === "audio_event"
      ? "event"
      : w.type === "word"
        ? "word "
        : "space";
  console.log(
    `  ${tag}  [${w.start.toFixed(2)}–${w.end.toFixed(2)}s]  ${w.text}` +
      (w.logprob !== undefined ? ` (logprob ${w.logprob.toFixed(3)})` : ""),
  );
}
// → "  event  [0.00–0.44s]  [phone beeping] (logprob -0.335)"

Notes

soundGeneration returns binary, not JSON — the provider already reads it as arrayBuffer() and hands you an ArrayBuffer. Pass output_format: "mp3_44100_128" (or any other ElevenLabs codec string) on the request object and the factory will strip it from the body and move it to the ?output_format= URL query.
speechToText accepts either a file Blob or a cloud_storage_url (S3/GCS/HTTP). For long-form audio set webhook: true — the call returns a small ElevenLabsWebhookAcknowledgement instead of the transcript, and the finished result is delivered to your registered webhook. Type-narrow the union with "text" in result before reading transcript fields.
Set diarize: true and num_speakers to label words by speaker; the per-word speaker_id field gets populated in that mode. Combine with use_multi_channel: true for stereo audio and the response switches to ElevenLabsMultichannelTranscript (one transcript per channel under transcripts[]).
Errors throw ElevenLabsError with status, code, and the parsed body attached. ElevenLabs returns either FastAPI's { detail: [{msg, ...}] } shape or { detail: { status, message } }; the client normalises both into error.message.

API Reference

303 endpoints across 28 groups. Each method mirrors an upstream URL path.

audioIsolation

POST https://api.elevenlabs.io/v1/audio-isolation

Cost tier: expensive

Published

Vulnerabilities

Links

Maintainers

Keywords

Readme

@apicity/elevenlabs

Installation

Quick Start

Real-world example: generate a sound effect, then run it through Scribe v2

API Reference

audioIsolation

audioNative

convai