AI Voiceover Compared: Kokoro, Qwen3, MiMo and ElevenLabs

Picking a text-to-speech engine is a cost versus quality tradeoff that plays out differently depending on what you are making and how much volume you need. Four options matter most right now: Kokoro, Qwen3 TTS, MiMo voice design, and ElevenLabs. They sit at very different price points and have genuinely different strengths.

Listen First

Here is the same 342-character passage read by each engine available in InkSlop, rendered with no editing. It mixes plain narration, a number, a question and a change of mood, so the differences are easy to hear.

On the twelfth morning he found footprints in the wet sand below the tower. They came up from the waterline, circled the building twice, and stopped at the door. He told himself it was a seal. Then he noticed they were barefoot. Who walks out of the sea at three in the morning? He locked the door that night for the first time in nine years.

Engine and voice	Sample
Kokoro, af_heart
Kokoro, am_michael
Qwen3, preset voice Ryan
Qwen3, Vivian with the instruction "tense and hushed, like telling a ghost story late at night"
MiMo voice design: "an old Scottish fisherman in his 70s with a low, weathered, slightly hoarse voice", speaking slowly "like telling a story by a fire"

All five are AI-generated. There is no ElevenLabs sample because it is not built into InkSlop (see below).

Kokoro: Most Choices, Cheapest to Run

Kokoro is an open-source model from Hexgrad, released under the Apache 2.0 license. At 82 million parameters it is tiny by modern AI standards, which is exactly what makes it interesting: it runs on CPU without a GPU and costs under $1 per million input characters when served via API.

InkSlop runs Kokoro on its own servers, which means the cost is compute time rather than API fees. That makes it the cheapest voice in InkSlop.

Voices available in InkSlop: all 28 English voices, 20 American and 8 British, with British voices read with British pronunciation. For other languages, set lang on the narration (for example lang="es") and InkSlop picks a native voice: Kokoro's own for Spanish, French, Italian, Brazilian Portuguese and Hindi. Every English Kokoro voice has a sample of each.

What Kokoro is good at: high-volume narration where you need a reliable, natural voice and you are not trying to match a specific person or capture subtle emotional range. Reddit story narrations, explainer videos, listicles, and educational content all work well. The voice is clean and the word-level timing data it produces is accurate enough for subtitle sync.

Where it falls short: emotional depth. Kokoro produces natural-sounding speech but it does not handle dramatic shifts in tone the way a larger model does. A monologue that needs to sound genuinely sad or angry will come across flat.

Cost in InkSlop: 0.5 credits per 1,000 characters, so a 60-second script of about 800 characters adds 0.4 credits.

Qwen3 TTS: Best Quality for the Price

Qwen3 TTS is Alibaba's open-weight TTS model, served through DeepInfra. It supports 10 languages, produces speech quality that beats ElevenLabs on speaker similarity benchmarks across several languages, and offers voice cloning from just 3 seconds of reference audio.

It does not run locally: InkSlop calls the DeepInfra API for every Qwen3 generation, which is why it costs more than Kokoro.

Voices available in InkSlop: nine preset voices (Vivian, Serena, Dylan, Eric, Ryan, Aiden, and others) plus any voice you clone from a reference clip.

What Qwen3 is good at: higher-fidelity output for English and other supported languages, native voices for Chinese, Japanese and Korean, and voice cloning for channel consistency. If you have a specific narrator voice you want to replicate, Qwen3 is the tool. It also accepts an instruct direction, but when we measured it, the effect was smaller than the difference between two takes; for directing delivery, MiMo's instructions work far better. Before cloning anyone else's voice, read the voice cloning legality guide first.

Where it falls short: ElevenLabs v3 still edges it for English prosody and naturalness in real-world listening tests, even if benchmark numbers favour Qwen3. It also covers fewer languages than ElevenLabs (10 vs 29).

Cost in InkSlop: preset voices cost 1 credit per 1,000 characters. Cloned voices cost 2 credits per 1,000 characters. Creating a new voice clone costs 5 credits flat.

MiMo Voice Design: Describe the Voice You Want

MiMo voice design, from Xiaomi and served through DeepInfra, works differently from the others. There is no list of voices to pick from. You describe the speaker in plain English (age, sex, accent, timbre, pace, attitude) and the model creates a voice to match. The fisherman sample above was made from a two-sentence description.

What it is good at: characters. Stories with several speakers, where each needs to sound like a different person, and narration that needs a specific delivery, such as deadpan comic timing or a hushed, tense read. Describing how the speaker talks ("pausing for a beat before each punchline") shapes the performance directly. The Reddit story case study shows four designed voices in one video.

Where it falls short: each description produces a new voice, so it is less predictable than a preset. Keep the exact description and reuse it word for word when a character needs to sound the same across videos.

Cost in InkSlop: 2 credits per 1,000 characters, the same as a cloned Qwen3 voice.

ElevenLabs: Best Quality, Priced for It

ElevenLabs is the benchmark everyone compares against. English prosody is better than any open alternative available in 2026, emotional range is wider, and the voice library is the largest. The API supports 29 languages and the voice cloning is fast and accurate.

The price reflects all of that. API costs run from $0.06 to $0.30 per 1,000 characters depending on plan tier and model. The Creator plan ($22/month) gives you 100,000 characters; the Pro plan ($99/month) gives 500,000. If you are producing high volumes of narrated content, ElevenLabs costs add up quickly.

That is why InkSlop does not offer ElevenLabs as a built-in option. At the volumes content creators actually work at, the per-video cost would push InkSlop's pricing beyond what the platform can absorb. The quality is real, but it is priced for enterprise scale.

If you need ElevenLabs quality specifically, you can generate audio externally and upload it into any InkSlop project as a custom audio track. The [tts engine="upload_audio"] path in the DSL accepts user-provided audio files, so you can slot ElevenLabs output into a fully scripted InkSlop video without losing the rest of the production pipeline.

Side by Side

	Kokoro	Qwen3 TTS	MiMo voice design	ElevenLabs
Model type	Open source, run on InkSlop servers	Open weight, API	API	Proprietary API
How you pick a voice	Preset list	Preset list, or clone from 3s of audio	Describe it in words	Library, or clone
Languages	9	10	English and Chinese	29
English quality	Good	Very good	Very good, most expressive	Best
InkSlop credits per 1,000 characters	0.5	1 preset, 2 cloned	2	Not built in
Available in InkSlop	Yes	Yes	Yes	Bring your own audio

Which One to Use

For most short-form content, Kokoro is the right default. It is fast, cheap, and accurate enough that viewers will not notice a quality difference compared to more expensive options.

For a channel with a consistent narrator voice, Qwen3 voice cloning is worth the extra credits. You clone once, use it across every video, and the consistency reads as professional.

For stories with several characters, or a narrator with a very particular delivery, MiMo voice design is the one to use. It costs the same as a cloned voice and needs no reference recording.

For long-form documentary narration or content where audio quality is the core product rather than background infrastructure, ElevenLabs is worth paying for separately and uploading.

For the full cost-per-video breakdown using each engine, see how much 100 AI videos actually cost. Try the voices in the creator before committing. The differences are real but they are also context-dependent, and the best way to evaluate is to render a test segment with the content you actually plan to make.