We replaced ElevenLabs with a free local TTS. Here is exactly what it cost.

Disclosure: This site intends to carry affiliate links, and every page that has one will say so above the first link. As of publication this page has none — we are not enrolled in any of the programmes for the tools named below, so every outbound link here is a plain untracked link and this article earns nothing whichever option you pick. No vendor paid for, reviewed, or saw it. See the disclosure policy.

Autocomplete for "elevenlabs alternative" completes to free, local, open source, github, and reddit. Nobody is asking whether ElevenLabs is good. They are asking whether they can stop paying for it.

We did stop paying for it — in fact we never started. Our voiceover runs on Kokoro-82M, locally, on a 2015-era MacBook, for $0/month. This article is the honest accounting: what it took to get there, what it actually costs in RAM and seconds, and the one specific situation where we would still hand ElevenLabs $6.

---

Start with the thing that catches most people

ElevenLabs' free tier does not include a commercial licence.

From elevenlabs.io/pricing, checked 2026-07-31:

If you are putting the audio on a monetised YouTube channel, in a client deliverable, or in an ad, the free tier is not a free tier for you. Starter at $6/month is the cheapest tier that makes the output legally yours to monetise. That single fact resolves a surprising number of "is the free plan enough" threads.

Kokoro, by contrast, ships under Apache-2.0: full commercial use, no attribution requirement, no tier games, no credit meter. For a monetised pipeline that licence difference is worth more than the audio quality difference, and it is the actual reason we run it.

---

What the market costs per million characters

Roughly 1,000 characters ≈ 1 minute of speech. These are list prices from vendor pricing pages at the time of our July 2026 survey, converted to a common unit — this is arithmetic on published rates, not a benchmark and not a quality ranking:

Read that carefully before you conclude anything dramatic. Ten cents a minute is not expensive for one hero video. It is expensive for a thousand automated clips. The gap between $0.0007 and $0.10 only matters at volume — which is precisely the case we are in, and possibly not the case you are in.

Note also the middle of that table: hosted Kokoro costs about $0.0007/minute. If you like the free model but do not want to run it on your own hardware, you can rent it for roughly a thousandth of ElevenLabs' rate. That option gets skipped in most "free alternatives" posts and it is often the right answer.

---

Running Kokoro locally: the real numbers

Measured on our machine — 8 GB RAM, 2 cores, macOS 13. Deliberately weak hardware, because if it runs here it runs anywhere:

• Model load: 4.0 seconds.
• Synthesis: 1.62× realtime. A 45-second read takes ~28 seconds of compute, ~32 s including load.
• Peak RSS: 613 MB. Comfortable alongside other processes on an 8 GB box.
• Disk: 310 MB model + 27 MB voices.

Verdict from those numbers: great for batch, unusable for interactive. If you want a voice that answers you in real time, local Kokoro on modest hardware is not it. If you are generating narration for clips in the background, it is free and it is fast enough.

One quality data point we can offer honestly: on an identical script, macOS's built-in say -v Samantha produced a 52.57-second read and Kokoro produced 45.48 seconds — about 7 seconds tighter per minute, with noticeably better pacing. We have not run a formal listening test against ElevenLabs and we are not going to pretend otherwise. Kokoro has scored well for its size on public TTS leaderboards; go listen to samples yourself rather than trusting anyone's adjective, including ours.

Installation gotchas that cost us an afternoon

• kokoro-onnx needs a real espeak-ng. The pip espeakng-loader wheel has a hard-coded data path that breaks on macOS. Install espeak-ng properly (brew install espeak-ng — a prebuilt bottle exists even for Ventura) and pass the library and data paths explicitly rather than trusting the loader.
• Get the weights separately. They are not in the pip package — download the .onnx and the voices file yourself and point the code at them.
• Alternative route: the kokoro pip package's misaki G2P avoids espeak entirely for most English text. If you only need English, try that first and skip the espeak problem completely.
• Budget the RAM. 613 MB peak is fine on 8 GB if you are not also running a browser and a video encode. We have had jobs die on this box for exactly that reason.

---

When we would still pay ElevenLabs

We are an affiliate for ElevenLabs, so treat this section with appropriate suspicion — and then notice that the three cases below are narrow, and that we told you hosted Kokoro is a thousand times cheaper.

1. The read carries the video. For a hero piece where voice quality is the difference between someone watching and leaving, $6 for a month of Starter credits is not a real cost. Our own rule is "local Kokoro for the pipeline, ElevenLabs for hero clips."
2. You need voice cloning. Kokoro gives you a fixed voice set. Cloning your own or a licensed voice is a different product, and it is what ElevenLabs is genuinely best known for.
3. You need commercial rights and you were about to use the free tier anyway. $6 versus a licensing problem is not a close call.

When we would not: high-volume automated narration, anything where 10¢/minute × hundreds of clips becomes a line item, or any project on hardware that can spare 613 MB.

Where to get ElevenLabs: elevenlabs.io — a plain untracked link. We are not in ElevenLabs's affiliate programme, so this page earns nothing either way.

One more honest caveat about that link: if you sign up through it, we get paid a percentage for a year — but only after your subscription has been active 90 days. That is a long time to be motivated to tell you the truth about a $6 product, which is roughly the point. Use the free model if it fits.

---

The decision, in one table

Prices verified against vendor pricing pages on 2026-07-31; per-million-character figures are from our July 2026 survey of published list rates. Performance figures measured on 8 GB / 2-core macOS 13. Everything here changes — re-check before you commit.