# AI voice vs hiring a voice actor — the honest cost comparison > What a minute of AI narration really costs, how fast revisions land, and the jobs where hiring a human voice actor is still the right decision. Published 2026-09-05 · eroq.ai — canonical: https://eroq.ai/blog/ai-voice-vs-hiring-a-voice-actor This comparison usually gets written by someone with a side to sell. Here is the version with the arithmetic shown and the losses admitted. Synthetic voice is priced per character, a human is priced per project, and those two units behave so differently that the interesting question is not "which is cheaper" but "which failure mode can you afford on this job". ## What a minute of narration actually costs Speech models bill per character, not per second, so the first thing to do is convert. A comfortable narration pace is around 150 words a minute, and an English word plus its space runs about six characters. So **one spoken minute is roughly 900 characters**. [Voice One](/models/eroq-voice-one) charges 3 credits per started block of 100 characters. Voice Turbo charges 2. - **Voice One** — 900 ÷ 100 = 9 blocks × 3 = **27 credits a minute**. About 27 cents at the entry pack rate, about 22 cents at the Creator plan rate. - **Voice Turbo** — 9 blocks × 2 = **18 credits a minute**. About 18 cents, or 15 on Creator. Now scale it to real deliverables: - A 90-second explainer script is around 1,350 characters — 14 blocks — **42 credits on Voice One**, roughly 34 cents. - A 30-minute narration is around 27,000 characters — **810 credits on Voice One**, about $6.62 at the Creator rate. On Turbo, 540 credits, about $4.41. - The 50 free credits a new account starts with cover about 1,600 characters on Voice One — a minute and three quarters of speech. Enough to hear whether the voice is right, not enough to narrate anything. One practical constraint: a single speech request takes up to 5,000 characters, which is about five and a half minutes. Long-form work is chunked by section anyway, which is also how you keep re-records small. A human voice actor prices differently — a session fee, often plus usage, and the number depends on the market, the medium and the length of the run. There is no honest single figure to quote, and anyone quoting one is generalizing. What is safe to say is the shape: it is a project price with a floor, so it does not shrink much for a 30-second job, and it does not scale linearly to a 30-minute one. ## Turnaround, and what it does to a schedule A render comes back in seconds and lands in your library. A booking involves availability, a session, and delivery, which for most projects means a day or several. The schedule effect is bigger than it sounds. When audio takes a day, the script freezes before the edit is locked, and the edit bends around the read. When audio takes seconds, you can cut the picture first and record the narration against the finished timing — which is the correct order and almost nobody gets to work that way. ## Revisions are where the two diverge This is the honest center of the comparison. Change one line in a 30-minute narration. With synthetic speech you re-render that line only: a 120-character sentence is 2 blocks, **6 credits on Voice One**, about five cents, and it is back before you have finished reading the note. With a session recording you are scheduling a pickup, and small pickups carry disproportionate overhead for everyone involved. This is also what makes generated voice good for work that is *inherently* iterative — versioned ad copy, per-market variants, a product name that changes twice before launch. Ten variants of a 200-character line is 10 × 6 = 60 credits. That is a rounding error, and it means you can actually test the copy. The flip side, stated plainly: your only direction channel is the text. On eroq you steer the read with punctuation, with the speed and expressiveness controls, and by rewriting the line — commas slow it, a full stop lands it, a question mark lifts the end. You cannot say "warmer, and lean on the second word". A director in a room with a performer can, and that is a real capability, not a nostalgia. ## Languages and casting The roster voices — Aria, warm and close; Orion, low and unhurried — each read in 13 languages, which turns localization from a casting exercise into a translation exercise. The same script in six markets is six renders and six translations, not six bookings. There is a full method for that in the [multi-language voice-over guide](/blog/multi-language-voice-over-guide). You can also clone a voice from your own audio, which is the right tool for a founder who narrates their own product videos and does not want to re-record every update. The consent rule is absolute and worth repeating: clone a voice you own or have written permission to use, never a real person's voice without it. The [voice cloning guide](/blog/ai-voice-cloning-guide) covers what makes a usable sample. Two limits to plan around: there is no lip-sync on eroq, so generated speech is narration and voice-over rather than a character speaking on camera in sync, and unusual proper nouns sometimes need to be respelled phonetically to come out right. ## Where a human voice actor is still the right call Hire the person when: - **The voice is the brand.** A campaign voice that will run for years, that an audience will recognize and associate, is a casting decision with a person attached. Treat it like casting an actor, because it is. - **The read needs direction.** If the script is comedic, ironic, or emotionally specific in a way that depends on timing, a performer who can take a note in real time will beat any number of prompt rewrites. - **There are union, consent or rights considerations.** Broadcast, games and many agency contexts have contractual frameworks around performance. Those exist for good reasons and are not optional inputs. - **The performer's identity is part of the product.** A known narrator on an audiobook, a founder who genuinely should be the one speaking, a documentary subject. - **Dialogue has to sit in a scene.** On-camera dialogue, matched ADR, two characters overlapping — that is performance work with picture, not narration. ## How to decide on a given job Two questions settle most briefs. First: will this line change? Versioned, localized or provisional copy strongly favors synthesis, because the revision cost is near zero. Second: is the voice doing creative work or delivering information? Information — tutorials, product walkthroughs, audiobooks of your own writing, internal video, social narration — is exactly where synthesis is strong. Creative performance is where it is not. The most common good answer is both. Synthesize the temp track so the edit can be cut to real timing, lock the script, and if the piece deserves a performance, book one for the final with the timing already proven. You will pay for one session instead of three. Read the two speech engines side by side in [Voice One vs Voice Turbo](/blog/voice-one-vs-voice-turbo), check the plan rates on the [pricing page](/pricing), then write a line and hear it at [the voice studio](/studio/voice). Long-form workflows live in [audiobooks and podcasts](/use-cases/audiobooks-podcasts). ## FAQ ### How much does one minute of AI voice-over cost? About 27 credits on Voice One or 18 on Voice Turbo, since a spoken minute is roughly 900 characters and the models charge 3 and 2 credits per 100 characters. In dollars that is roughly 27 and 18 cents at the entry pack rate, and less on a plan. ### Can AI voice replace a voice actor entirely? For information-carrying narration — tutorials, product videos, internal content, localization — it usually can. For a performance that carries a brand, for dialogue in a scene, or anywhere direction and timing are the point, a human performer is still better, and the sensible workflow uses synthesis for temp tracks and drafts either way. ### Is it legal to clone a voice? Clone your own voice, or one you have written permission to use. Cloning a real identifiable person's voice without consent is not allowed on eroq and carries obvious legal exposure elsewhere. The model roster and your own clones are the safe ground. ### How many languages can the voices read? Each roster voice reads in 13 languages, so the same cast can carry a localized set. Translate the script, render each version, and keep the delivery consistent across markets.