AI/LLM

If you've been living under a rock, the open-source voice scene has been on a tear. XTTS-V2, F5-TTS, OpenVoice — all free, all runnable on an M4 Mac. So is ElevenLabs still worth the bill? Spoiler: yes, for one specific thing.

The Setup

I A/B'd these on the same 12-second voice sample for a Rebuild Relief explainer narration. Same script, same reference audio, blind test with the team.

{`# XTTS-V2 — the open contender
pip install TTS

tts --model_name tts_models/multilingual/multi-dataset/xtts_v2 \\
    --text "Hi, this is Aiden from Rebuild Relief." \\
    --speaker_wav ref.wav \\
    --language_idx en \\
    --out_path out.wav`}

The Money Pattern

For app-side TTS where the voice doesn't need to be a specific person, XTTS-V2 is free and good enough. For voice cloning of a real person where listeners know the voice, ElevenLabs still wins — the prosody and breath pacing aren't close yet.

{`import { ElevenLabsClient } from "elevenlabs";

const eleven = new ElevenLabsClient({ apiKey: process.env.ELEVEN_KEY });

const audio = await eleven.generate({
  voice: "aiden-clone-v2",
  model_id: "eleven_turbo_v2_5",
  text: "Your claim has been approved.",
  voice_settings: {
    stability: 0.45,
    similarity_boost: 0.85,
    style: 0.2,
  },
});

// pipe straight into FFmpeg for the final mix
await Bun.write("voiceover.mp3", audio);`}

The Catch

ElevenLabs is still expensive at scale and the rate limits will bite you the moment you hit production traffic. XTTS-V2 is free but the inference latency on CPU is brutal — you need a GPU or an Apple Silicon Mac to get realtime.

F5-TTS is the dark horse — flow-matching architecture, faster than XTTS, but quality is a coin flip on short references under 5 seconds.

The Verdict

Hybrid stack wins. ElevenLabs for the hero asset where the voice has to be perfect. XTTS-V2 or F5-TTS for everything else — push notifications, in-app TTS, batch narration over CSVs of insurance copy. Your spend drops 90% and nobody hears the difference where it matters.