How to Use the ElevenLabs API: Voices, Previews and Examples

Choose a voice for your video’s use case, preview its delivery, then open its ready-filled fal playground or use its MCP prompt. Pass the selected voice ID explicitly when calling Eleven v4 through the API or CLI.

A text-to-speech request can succeed and still give you the wrong voice for your video. The words are correct. The audio is clean. But the delivery feels like a generic voiceover instead of someone explaining an idea.

Start by choosing the speaker. Then choose the model and audition a short passage from your script. In this tutorial, you'll use Eleven v4 on fal through fal's API, MCP connection, and command-line tool, with a specific voice ID carried through every request. You can also browse the full ElevenLabs model collection on fal.

Quick answer: Use elevenlabs/tts/eleven-v4 and pass your chosen ElevenLabs voice ID in voice. This guide includes six social-media voice candidates and five each for education, advertising, and entertainment, with playable fal v4 previews using their ElevenLabs demo scripts. Compare them in the voice tables below.

This guide assumes basic familiarity with JSON and a terminal. Research and generation checks: October 28, 2026.

Which ElevenLabs voice should I use for my video?

Choose an ElevenLabs voice by the job your video needs to do, then pass its exact ID in the fal voice field. For a casual GRWM or lifestyle UGC explainer, audition Siren (eXpIbVcVbLo8ZJQDlDnl). For a warm tutorial, compare Zara (jqcCZkN6Knx8BJ5TBdYR). For upbeat short-form storytelling, try Liam (VCgLBmBjldJmfphyB8sZ). These are starting points for casting, not guarantees of a human-sounding result.

Video brief	Voice to audition first	Why consider it	Try this voice
Make a casual “get ready with me” video or lifestyle UGC explainer	Siren	Editorial listening notes describe a chill, Southern California influencer-style delivery	Playground · MCP
Explain a product or teach a short tutorial in a warm conversational tone	Zara	Supplied description emphasizes warmth, clarity and grounded expression	Playground · MCP
Narrate a story for TikTok, Reels or YouTube Shorts	Liam	Supplied description emphasizes playful energy and story-led delivery	Playground · MCP
Present celebrity news, a how-to or an upbeat creator update	Brittney	Supplied description emphasizes youthful, informative delivery	Playground · MCP
Add an enthusiastic narrator to a fast-paced Short	Tim Hook	Supplied description emphasizes cheerful, high-energy storytelling	Playground · MCP

How do I use Siren for a GRWM or UGC explainer through fal?

Try Siren’s lipstick GRWM script in the playground. Click Edit input to see the ready-filled script and custom voice.

Send the following JSON input to elevenlabs/tts/eleven-v4. The voice field selects Siren; omitting it lets the endpoint choose its default speaker. The settings below match the fal renditions in this article. Replace text with a short passage from your video, listen to the result, and keep the same voice ID when generating the full narration.

json

{
  "text": "Hey guys, welcome back to the channel. Today I'm going to be going to a party but I really wanted to show you this new lipstick from Terra. It's really good and it doesn't smudge.",
  "voice": "eXpIbVcVbLo8ZJQDlDnl",
  "stability": 0.5,
  "similarity_boost": 0.75,
  "output_format": "mp3_44100_128",
  "apply_text_normalization": "auto"
}

How should an automated video workflow choose a voice?

Read the video brief for audience, format and desired delivery: for example, a casual beauty walkthrough versus a structured lesson.
Match that brief to a voice’s suggested role and copy the exact voice ID, rather than choosing by name alone.
Generate a short audition with the actual script. A successful API response confirms compatibility, not naturalness or suitability.
Have the creator compare phrasing, emphasis and pronunciation. Save the chosen ID and settings with the project.
Pass the selected voice explicitly in subsequent API, MCP or CLI requests. If the voice is unavailable, report the issue and select another candidate instead of silently reverting to the default.

Each preview below uses that voice’s own ElevenLabs library demo script. Their text descriptions make the casting rationale available even when a reader or retrieval tool cannot play audio. Only Siren’s specific delivery note reflects the supplied editorial listening feedback; the other roles are audition suggestions based on supplied descriptions.

Why the default voice can miss the mark

The fal v4 endpoint currently defaults to Rachel when voice is omitted. That is a technical default, not a casting decision for your project. A product explainer, a classroom lesson, and a character story can use the same model while needing very different speakers. fal v4 API reference

Keep three choices separate:

Model: the speech engine, such as Eleven v4.
Voice: the speaker identity, selected using a voice ID.
Delivery: the script, punctuation, and generation settings that shape the performance.

Switching models alone doesn't select a better narrator. Likewise, increasing a setting doesn't guarantee a more natural read. Cast a suitable voice first, then tune the delivery against a real passage from your video.

Find a voice and copy its ID

Open the ElevenLabs Voice Library. Choose a category, narrow by language and accent, and preview candidates. In a voice's three-dot menu, choose Copy voice ID. Save the ID alongside its full display name; similar names can refer to different speakers.

A library voice ID is not an API key. It identifies a speaker. It also doesn't mean you're creating a clone: here, you're selecting an existing voice.

ElevenLabs documents searching the library by voice ID and notes that direct API access to library voices isn't available on its free tier. Availability through another provider must be checked separately. The 21 IDs in this guide were each accepted in a live fal v4 generation on the research date. Voice Library guide

Voices to audition, organized by use case

How this shortlist was made: Six social-media voices were selected with editorial input, alongside five candidates each for education, advertising and entertainment. Each of the 21 listings and default audio demos was opened in the ElevenLabs library. The spoken words were recovered with speech-to-text, then sent to fal’s Eleven v4 endpoint with the corresponding voice ID, stability 0.5, and similarity boost 0.75.

About the previews: Each player is a fal Eleven v4 generation using that voice’s automatically transcribed ElevenLabs demo script. Transcripts may contain errors. Promotional claims in a demo are sample copy, not independently verified claims by this article.

Suggested uses come from the live library descriptions and supplied editorial feedback, not a listening-ranked evaluation of the whole library. Listen to the previews and test your own script before choosing a voice.

Choose Try in playground to open a saved fal example with the voice and demo script already filled in. Click Edit input to change the words, then Run. Choose Try via MCP for a ready-to-copy prompt for your connected assistant.

Social media

Voice	Use case	Try this voice
Brittney	Youthful, upbeat celebrity news, how-to videos and creator conversation	Try in playground · Try via MCP
Liam	Short-form storytelling with a balance of authority and playful energy	Try in playground · Try via MCP
Tim Hook	High-energy, cheerful YouTube Shorts storytelling	Try in playground · Try via MCP
Jamal	Casual creator commentary and social-media voiceovers	Try in playground · Try via MCP
Zara	Warm, grounded tutorials, brand storytelling and conversational ads	Try in playground · Try via MCP
Siren	Orange County / Los Angeles influencer-style “get ready with me” videos, lifestyle UGC and casual explainers	Try in playground · Try via MCP

Education

Voice	Use case	Try this voice
Daniel	Audio guides and accessible explainers	Try in playground · Try via MCP
Quentin	Structured lessons and documentary explanations	Try in playground · Try via MCP
Dr Chris	Lectures and authoritative course narration	Try in playground · Try via MCP
Megan	Professional training and onboarding	Try in playground · Try via MCP
Casey	Product walkthroughs and concise lessons	Try in playground · Try via MCP

Advertising

Voice	Use case	Try this voice
Kristen	Lifestyle and product promos	Try in playground · Try via MCP
Jesse	Friendly benefit-led brand reads	Try in playground · Try via MCP
Markmont	Grounded sports and product campaigns	Try in playground · Try via MCP
Asher	Commercial product introductions	Try in playground · Try via MCP
Emma	Clear launches and brand explainers	Try in playground · Try via MCP

Entertainment

Voice	Use case	Try this voice
Mike	POV and storytime narration	Try in playground · Try via MCP
Hannah	Bright presenter-style storytelling	Try in playground · Try via MCP
Odom — Upbeat	Game-show and playful presenter delivery	Try in playground · Try via MCP
Celeste	Period stories and vintage characters	Try in playground · Try via MCP
Tyler Cruz	Music-show intros and energetic hosting	Try in playground · Try via MCP

If you want a compact first round, start with Zara for conversational explainers and compare Jamal’s library demo, or Siren for lifestyle UGC, Megan and Casey for training, Jesse and Kristen for ads, and Hannah and Mike for storytelling. Those pairings give you different delivery profiles to compare; they are not quality scores.

The entertainment choices are deliberately varied. A vintage character and a game-show host shouldn't be judged by the same expectation as a quiet educational narrator. An expressive voice can be right for a performance and wrong for an otherwise conversational explainer.

Make an Eleven v4 request through the fal API

You'll need a fal account with generation access, a fal API key, and Node.js with npm. Keep the key on a server or in a local environment variable; don't put it in browser code. Install the official client, saving the resolved version exactly and retaining the lockfile:

bash

npm install --save-exact @fal-ai/client
export FAL_KEY="YOUR_FAL_API_KEY"

Save this as voiceover.mjs. It uses a Tyler creator voice ID that completed a live v4 generation during this research.

javascript

import { fal } from "@fal-ai/client";
import { writeFile } from "node:fs/promises";

if (!process.env.FAL_KEY) throw new Error("Set FAL_KEY first.");

try {
  const result = await fal.subscribe("elevenlabs/tts/eleven-v4", {
    input: {
      text: "Here's the idea. Choose a voice that fits your story, test a short scene, and listen before you render the whole video.",
      voice: "rPMkKgdwgIwqv4fXgR6N",
      stability: 0.5,
      similarity_boost: 0.75,
      output_format: "mp3_44100_128",
      apply_text_normalization: "auto"
    }
  });

  const url = result.data?.audio?.url;
  if (!url) throw new Error("The response did not contain an audio URL.");
  const response = await fetch(url);
  if (!response.ok) throw new Error(`Audio download failed: ${response.status}`);
  await writeFile("voiceover.mp3", Buffer.from(await response.arrayBuffer()));
  console.log("Saved voiceover.mp3. Request:", result.requestId);
} catch (error) {
  console.error(error);
  process.exitCode = 1;
}

Run it with:

bash

node voiceover.mjs

The client submits the job, waits for completion, and exposes the generated file at result.data.audio.url. The example then saves that file. Replace only voice to audition another candidate with the same script. The endpoint and request structure follow the official fal client example.

For this standard queued request, omit response_format: "bytes". The verified schema returns an audio-file object with a URL. A streaming implementation is a separate integration; don't assume its transport options belong in this request body.

If you copied a payload from a /stream request, check its options against the queued endpoint before reusing it. A field appearing in a streaming payload does not establish that it belongs in a queued request.

Use Eleven v4 through fal MCP

MCP lets an assistant discover and run the model with your connected fal account. Connect a compatible client to https://mcp.fal.ai/mcp-relay using Streamable HTTP and browser OAuth, or follow the setup guide for your client. The OAuth connection doesn't require pasting a fal API key into chat. fal MCP setup

Use a specific instruction so the assistant preserves the model and voice:

Use fal's elevenlabs/tts/eleven-v4 endpoint. Generate the text “Here's the idea. Choose a voice that fits your story, test a short scene, and listen before you render the whole video.” Use voice ID rPMkKgdwgIwqv4fXgR6N, stability 0.5, similarity_boost 0.75, output_format mp3_44100_128, and apply_text_normalization auto. Check the live schema, submit one generation, wait for that request to finish, and return the audio and request ID.

For clients exposing the fal submit_job tool, the tool arguments are:

json

{
  "endpoint_id": "elevenlabs/tts/eleven-v4",
  "input": {
    "text": "Here's the idea. Choose a voice that fits your story, test a short scene, and listen before you render the whole video.",
    "voice": "rPMkKgdwgIwqv4fXgR6N",
    "stability": 0.5,
    "similarity_boost": 0.75,
    "output_format": "mp3_44100_128",
    "apply_text_normalization": "auto"
  }
}

This is a tool argument object, not the HTTP request body. The assistant should check the returned request ID until it completes, rather than submit another generation to check progress. Tool names can have a client-specific namespace.

Use Eleven v4 from the command line

fal's model-generation CLI is genmedia. The separate fal CLI handles deploying and managing applications. Install genmedia using its official instructions, then configure your key with genmedia setup or the FAL_KEY environment variable. genmedia guide

bash

genmedia schema elevenlabs/tts/eleven-v4

genmedia run elevenlabs/tts/eleven-v4 \
  --text "Here's the idea. Choose a voice that fits your story, test a short scene, and listen before you render the whole video." \
  --voice "rPMkKgdwgIwqv4fXgR6N" \
  --stability 0.5 \
  --similarity_boost 0.75 \
  --output_format mp3_44100_128 \
  --apply_text_normalization auto \
  --download

The CLI accepts model inputs as flags and waits for completion by default. --download saves returned media. The MCP sections below include the exact request body for each shortlisted voice.

If you're calling ElevenLabs directly

The direct ElevenLabs API uses a different request shape. The voice ID goes in the URL, the model is selected with model_id, and delivery settings are nested under voice_settings. Use an ElevenLabs key rather than a fal key. ElevenLabs' current introductory example uses eleven_v4. ElevenLabs introduction, Create speech reference

bash

export ELEVENLABS_API_KEY="YOUR_ELEVENLABS_API_KEY"

curl --fail-with-body --silent --show-error \
  "https://api.elevenlabs.io/v1/text-to-speech/rPMkKgdwgIwqv4fXgR6N?output_format=mp3_44100_128" \
  -H "xi-api-key: $ELEVENLABS_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "text": "Choose a voice that fits your story, then test a short scene.",
    "model_id": "eleven_v4",
    "voice_settings": {"stability": 0.5, "similarity_boost": 0.75}
  }' \
  --output direct-voiceover.mp3

On success, this endpoint returns audio bytes. On an HTTP error, inspect the error response rather than treating the output file as valid audio. This direct-provider example was checked against documentation; it wasn't executed with an ElevenLabs key during this research.

Make your explainer sound less like a generic voiceover

Start with a script that someone could comfortably say aloud. “In our fast-paced world, let's explore the power of patience” gives a narrator little concrete intent. “I kept rushing the first step. That was why I had to do it twice” gives the speaker a situation, a turn, and a reason to emphasize particular words.

For auditions, use the actual opening of your video plus one difficult sentence. Include a technical term, a number, or a change in emotion. A voice that handles a greeting well may struggle with a dense explanation.

Listen for five things:

Phrasing: Do pauses separate ideas, or interrupt them?
Emphasis: Does the important word land without every phrase sounding equally dramatic?
Pronunciation: Are names, abbreviations, and numbers correct?
Consistency: Does the speaker keep the same identity and energy across sentences?
Fit: Does the delivery suit this audience and scene?

Use the shared settings as a baseline, not a universal preset. Lower stability allows more variation; higher stability favors consistency. Similarity boost controls adherence to the reference voice, and increasing it can reduce naturalness. Change one setting at a time and compare the resulting takes. fal parameter definitions

For a first explainer, audition a few untagged reads before adding emotional direction. Keep the version that best supports the message. Then generate the longer narration in coherent scenes and check the joins in your editor.

Troubleshooting

The result still uses the default speaker. Inspect the submitted JSON. For fal, the field is voice, not voice_id. Check that your application isn't dropping it when it builds the request.

A voice ID fails. Confirm it in the library and check current availability through your chosen provider. A name, preview, or previous success doesn't guarantee permanent availability. Don't silently replace it with a default and deliver the wrong speaker.

The script is too long. The fal schema inspected for this guide caps text at 5,000 characters. Respect the endpoint's current limit even if the underlying model has a different limit elsewhere. Split at scene or paragraph boundaries.

The preview doesn't match the library demo. The library sample may use another script, model, or configuration. Listen to the previews in this guide, then audition your own text. Generation is variable; one take isn't a full quality evaluation.

You need a permanent article asset. Retain the chosen generated files in your publishing storage before publication. Generated-media URLs can expire; a preview link is not an archival guarantee.

For an explainer, the useful workflow is short: pick a suitable speaker, generate one representative scene, and listen before rendering the entire script. Once the casting works, carry that exact voice ID through the API, MCP, or CLI workflow you prefer.

Try these voices via MCP

Connect your assistant using the fal MCP setup guide. Copy the prompt and input for your chosen voice into that assistant. Generation uses your fal account credits.