AI voiceover in 2026: ElevenLabs, OpenAI, Google, and when to pay vs use open source
Same script, four engines. The picks aren't where you'd expect — and one of them costs nothing.
I ran the same 90-second documentary script through ElevenLabs Multilingual v2, OpenAI's gpt-4o-mini-tts (Sage), Google Studio voices, and a local F5-TTS clone of my own voice. The surprise: on the documentary read, Sage beat ElevenLabs. On the same script delivered as a video-game character monologue, it wasn't close — ElevenLabs walked it.
That gap is the whole point of this post. There is no "best" TTS in 2026. There are best fits.
Quick verdict per provider
ElevenLabs (Multilingual v2 / Turbo v2.5) — still the king of expressive speech. Whispers, laughs, sighs, character work, ad reads with attitude. Multilingual v2 is the quality model; Turbo v2.5 is half the cost and 75ms-class latency for real-time use. ElevenLabs themselves now recommend the Flash v2.5 model over Turbo for almost all cases.
OpenAI (gpt-4o-mini-tts, voices Sage / Coral / Shimmer) — the dark horse for narration. The gpt-4o-mini-tts model is a step up in naturalness from the older tts-1-hd, and crucially it accepts a style prompt ("read this like a calm documentary narrator"). Sage is calm and grounded; Coral is the corporate-explainer voice; Shimmer is intimate. They sound less "performed" than ElevenLabs — which is exactly what you want for long-form narration where ElevenLabs voices can drift into an unwelcome theatrical mode.
Google Cloud TTS (Studio voices) — the IVR and enterprise pick. Polished, predictable, robotically dependable. Studio voices are the premium tier above WaveNet/Neural2/Chirp 3 HD. Excellent for accessibility, dashboards, phone trees. Don't reach for it when you need warmth.
Open source (F5-TTS, Coqui XTTS v2) — closer than the paid camp wants you to believe. F5-TTS is the strongest open model on naturalness and zero-shot cloning right now, edging out XTTS v2 on quality. XTTS v2 still wins on streaming and ~200ms time-to-first-chunk for live applications. Both clone a usable voice from ~10 seconds of reference audio.
Best use case per voice
| Job | Pick | Why |
|---|---|---|
| Documentary / long-form narration | OpenAI Sage | Restrained delivery, style-prompt steerable, no theatre |
| Punchy ad reads, energetic VO | ElevenLabs (custom voice) | Emotion + emphasis are still its moat |
| Character / animation / games | ElevenLabs Multilingual v2 | Range and non-verbals nothing else touches |
| Corporate explainer / e-learning | OpenAI Coral or Google Studio | Crystal articulation, no quirks |
| IVR, accessibility, dashboards | Google Studio | SSML-mature, predictable, enterprise SLAs |
| Personal-channel weekly podcast | F5-TTS local clone | Free, your voice, runs on a 4090 |
| Real-time agent / live dub | XTTS v2 (or ElevenLabs Flash v2.5 if budget allows) | Streaming with low TTFB |
The pricing math, normalized
A 5-minute video script is roughly 4,500 characters. Run that 50 times a month (say, a weekly newsletter, a YouTube channel, plus drafts):
- ElevenLabs Multilingual v2 — $0.10 per 1k chars. 50 × 4,500 = 225k chars = ~$22.50/mo. Or absorb it in the Creator tier (~$11/mo at the regular rate, 121k credits included).
- ElevenLabs Flash/Turbo v2.5 — $0.05 per 1k chars. ~$11.25/mo.
- OpenAI tts-1-hd — $0.030 per 1k chars (per OpenAI's pricing page). ~$6.75/mo.
- OpenAI gpt-4o-mini-tts — token-priced; the community-confirmed estimate is ~$0.015 per minute of audio output. 250 minutes/mo = ~$3.75/mo.
- Google Studio voices — $160 per million chars = $36/mo. (Standard is $4/M with 4M free, but those voices sound like 2019.)
- F5-TTS local — $0 in API spend, ~$0.40/hour in electricity if you bought the GPU.
OpenAI's narration model is now the cheapest paid option by a wide margin, and it's not even close to the worst-sounding. That alone should make you re-test it if you wrote it off in 2024.
When open source actually beats paid
Three real scenarios:
- 01You're cloning your own voice and you publish weekly. F5-TTS removes a recurring bill and removes the platform from your supply chain. Your voice never leaves your machine.
- 02You need bulk — audiobooks, language learning, dataset narration. Even at Turbo v2.5 prices, a 12-hour audiobook costs real money. Local inference is free after the first one.
- 03You need a voice that doesn't exist. Both XTTS v2 and F5-TTS clone from short reference audio. Hire one voice actor for a 60-second sample (with a written license) and you have an unlimited-use brand voice — no per-character meter.
Open source loses on convenience, on noisy-source robustness, and on real-time agent use cases where ElevenLabs Flash's sub-100ms TTFB is the product.
A mildly contrarian take
Most "ElevenLabs vs everyone" posts treat ElevenLabs as the ceiling. It isn't, for narration. The expressive cleverness that makes ElevenLabs win on character work is exactly what makes it worse for documentary — it keeps reaching for emotion the script didn't ask for. Sage in gpt-4o-mini-tts, given a one-line style prompt, sits flatter and trusts the writing more. If your VO sounds like it's acting when it should be narrating, swap engines before you swap scripts.
One legal note on cloning — read this once
ElevenLabs' Prohibited Use Policy forbids cloning anyone's voice without consent or legal right, and a growing list of US states — California (Civ. Code 3344), New York (Civ. Rights 50-51), Tennessee's ELVIS Act, and others — make written consent a statutory requirement. The same rule applies to the open-source models, even though no platform is policing you. "I'll just clone the YouTuber's voice for the parody" is not a defense; the ELVIS Act in particular treats unauthorized vocal likenesses as a property-rights violation regardless of intent. Get the release in writing, store it with the project file, move on.
The takeaway: before your next project, run the same paragraph through Sage, Multilingual v2, and an F5-TTS clone of yourself. Pick by ear, not by brand. Most creators are paying for the wrong tier — usually too much for narration, sometimes too little for character work.