Turn a script into a natural voiceover

Still the only one built around a whole script, not a request.

Use it when

You have a real script — an explainer, an audiobook chapter, a course module — and you need pronunciation fixed, two voices swapping mid-paragraph, and a WAV you can hand to an editor without writing chunking code.

Cost

Free 10k credits/mo, no commercial use. Paid $6-$990/mo (30k-6M credits).Free $0 (10k credits/mo), Starter $6 (30k), Creator $22 (121k, first month half price at $11), Pro $99 (600k), Scale $299 (1.8M), Business $990 (6M), Enterprise custom. Unused credits roll over but only up to two months' worth, and only while the subscription stays active. Instant voice cloning unlocks at Starter; professional voice cloning at Creator. Pay-as-you-go top-up credits last 12 months and kick in when the monthly quota runs dry.

Good at

Studio is the actual differentiator: import EPUB, PDF, DOCX, TXT, HTML or a URL, split into up to 500 chapters, assign different voices inside a single paragraph, apply pronunciation dictionaries with phoneme or alias tags, export MP3, WAV or video. Nothing else here treats a script as a document.

The catch

It is far and away the most expensive per character on this list — the mid plans work out around ten to twenty times what Speechify or Inworld charge per million characters. And the free tier is a trap dressed as generosity: ElevenLabs states plainly that it includes no commercial licence and cannot be used for any commercial purpose, and that free-plan output published anywhere must carry 'elevenlabs.io' attribution in the title. You need a paid plan before a single second of this reaches a client.

Wrong for

Anyone generating narration at volume, or anyone who just needs an API call. Go to Speechify or Inworld and keep the difference.

Top of the blind arena. Bring your own script pipeline.

Use it when

You are generating voiceover from code, naturalness is the thing you will be judged on, and you are happy to write your own paragraph splitting and stitching.

Cost

Free 20k credits/mo, no commercial use. Paid $5-$299/mo (100k-8M credits).Free $0 (20k credits, ~27 TTS minutes), Pro $5 (100k credits, ~133 min), Startup $49 (1.25M credits, ~1,667 min), Scale $299 (8M credits, ~10,667 min), Enterprise custom. Commercial licence starts at Pro. Instant voice cloning from Pro (2 clones); professional cloning from Startup. Artificial Analysis lists Sonic 3.6 at $49 per 1M characters.

Good at

Raw naturalness. It ranks first on both Artificial Analysis speech arenas at Elo 1288, with the gap to third wider than the gap from third to twelfth. 61 locales, 500+ preset voices, Hinglish code-switching, and accent adherence that survives voice cloning. Sub-90ms replies and ~132 characters/sec generation, so a long script renders fast.

The catch

It is engineered for live voice agents, not narration. The launch material sells sub-90ms latency and a 99.9% uptime SLA — there is no Studio, no chapters, no document import, no pronunciation dictionary workflow. You will spend day three writing the splitter, the retry logic and the concatenation that ElevenLabs gives you in a text box. The free tier's 20k credits is roughly 27 minutes and carries no commercial rights.

Wrong for

Anyone who wants to paste a script into a browser and download an MP3. That is ElevenLabs or Speechify.

Near arena-leading quality at a tenth of ElevenLabs' price.

Use it when

You are shipping narration at volume and the budget matters more than the last five percent of naturalness — or you want a free tier you can legally publish from.

Cost

Free 50k chars/mo, commercial use allowed. Paid $10-$499/mo, $6-$10 per extra 1M chars.Free $0 (50k TTS characters/mo, hard cap, no overages, commercial use permitted). Starter $10/mo then $10 per additional 1M characters. Pro $99/mo then $8 per 1M. Scale $499/mo then $6 per 1M. Enterprise volume pricing. Artificial Analysis lists Simba 3.2 at $10 per 1M characters at Elo 1243 — second on the leaderboard. 1,500+ voices, 30+ languages.

Good at

Price-to-quality, by a distance. Simba 3.2 sits 45 Elo behind the arena leader and costs roughly a fifth per character. The free tier is the only one here that explicitly permits commercial use, and it is a hard cap — no overage, no surprise invoice at the end of the month.

The catch

Voice cloning is not on the free tier. The Free card lists commercial use and 1,500+ voices; 'Voice cloning, streaming, SSML' appears only from Starter. So the free 50k characters gets you stock voices only — if the brief was 'clone the founder's voice', you are paying from day one. The free tier's hard cap also means generation simply stops mid-project rather than degrading.

Wrong for

Long documents needing chapter structure and pronunciation dictionaries. Use ElevenLabs Studio for that.

Free two-speaker dialogue in 100+ languages, with a data catch.

Use it when

You need a two-hander — interview, dialogue, back-and-forth explainer — or you are prototyping and want to spend nothing on non-confidential text.

Cost

Free tier free of charge. Paid $1/1M text-in, $20/1M audio-out tokens.gemini-3.1-flash-tts-preview: paid tier $1.00 per 1M input text tokens, $20.00 per 1M output audio tokens; batch halves both to $0.50 and $10.00. Free tier is listed as free of charge for both input and output. Free-tier rate limits are no longer enumerated in the docs — they are only visible in AI Studio under your own account. Also available: Gemini 2.5 Flash Preview TTS and 2.5 Pro Preview TTS.

Good at

Native multi-speaker generation — up to 2 speakers in one call, which no other tool here does without stitching separate renders. Over 100 languages with automatic detection, a 32k-token session context, and genuinely free access for prototyping.

The catch

Two problems land on day three. Google's own pricing page says free-tier content is 'used to improve our products' while paid-tier content is not — so a client's unreleased script cannot go through the free path. And the speech-generation docs warn that quality and consistency 'may begin to drift with generated outputs that are longer than a few minutes', which is exactly the length of the voiceover you are trying to make. Add: 30 preset voices, no voice cloning, and every TTS model still carries a 'preview' label.

Wrong for

Anyone who needs a specific or cloned voice, or a 20-minute continuous read. Go to ElevenLabs or Speechify.

MIT weights, free cloning, no per-character meter running.

Use it when

Volume is high enough that per-character pricing hurts, the script cannot leave your infrastructure, or you want unlimited voice cloning without a subscription tier deciding when.

Cost

Free. MIT-licensed weights; you pay only for the GPU.No pricing. Weights are MIT — the genuinely permissive licence in this field, and the reason it beats Fish Audio S2 Pro and Breeze TTS 2 here despite both scoring higher on the arena. 500M parameters. Runs on CUDA, CPU or Apple Silicon (mps); the Nano variant runs 3x faster than realtime on 8 CPU cores. 23+ languages plus dedicated single-language packs for six.

Good at

Zero-shot voice cloning from roughly a 10-second reference clip, with no tier gate and no per-clone limit. Exaggeration and cfg_weight parameters give you direct control over how dramatic the read is, and Turbo/Nano support paralinguistic tags like [laugh] and [cough].

The catch

Every file it generates carries Resemble's Perth neural watermark — imperceptible, but designed to survive MP3 compression and editing, and not something you can switch off. On top of that, output quality depends on you tuning cfg_weight and exaggeration per reference clip, and a reference whose language does not match the language tag will bleed its accent into the read. That is an afternoon of fiddling per voice, every voice.

Wrong for

Anyone without a GPU and an afternoon. If you just need a clean read by Friday, pay Speechify $10.