Best AI voices for narration, ranked and priced

Narration is the voice job where nobody is waiting on the line. You are turning a written script into finished audio for a video, a course, an advert or a podcast, and that flips what matters. The response speed that decides a live phone agent stops mattering at all. What matters instead is how the voice holds up across three unbroken minutes of listening, what each 1,000 characters of script costs to generate, and whether the licence lets you publish the result. This page ranks the seven engines in our directory that put a usable per-character narration price on the table, best first, for exactly that job.

If you landed here looking for a live phone agent, you want a different page. The business-calls ranking weighs ease of use and running cost the way a call buyer should, and the low-latency roundup covers the speed specialists. Narration flips those priorities, which is why this list looks nothing like either: two of the engines here do not place phone calls at all, and it costs them nothing in this ranking.

The usual honesty note, stated plainly because it matters more on a quality question than anywhere else. We have not run scored blind listening tests yet. The order below comes from our narration-weighted editorial score, a documented preview, and every price is sourced and dated on the linked platform page. When the listening tests land, measured results replace the preview, and if the two contradict each other, the measured result wins.

What a narration price actually is

Every engine here bills the same underlying way: by how much text you feed it. The unit is 1,000 characters of script, roughly 150 to 180 words, and a finished minute of speech comes out of roughly 900 to 1,000 characters (the same characters-per-minute basis we show on the Rime page). So when you see $0.10 per 1,000 characters, read it as about ten cents per finished minute. That conversion is the whole trick to comparing these platforms, because no two of them quote the same headline unit.

The packaging differs, and the packaging is where budgets go wrong. ElevenLabs and Cartesia sell monthly credits (on ElevenLabs’ standard models one character costs one credit). Speechify sells characters directly, with the rate falling as the tier rises. Hume meters its Octave engine per character on top of a subscription. Fish Audio prices its API in UTF-8 bytes, $15 per million, and a byte is one character as long as the script is in English. Murf does not sell characters at all: you buy bundled hours of finished audio per year, so its per-character figure here is our division, not Murf’s price sheet. And Rime publishes one starting-at rate with volume discounts it does not spell out. Same job, six different wrappers. The cost table further down converts all of them to the same two numbers.

How I ranked these

The order is our narration-weighted editorial score, the same published weighting behind the narration ranking table: voice quality carries 40% and voice range 35%, because for narration the voice is the product, while ease of use gets 15% and value for money 10%. Run each platform’s four sub-scores through that weighting and you get the order below: ElevenLabs 9.3, then Cartesia and Fish Audio tied on 7.9, Speechify 7.8, Murf 7.6, Hume 6.5, Rime 6.4. One tie to be honest about: Cartesia holds second because its numbers have survived months of re-captures, while Fish Audio joined the directory on 11 July 2026, so the newcomer sits third until its figures have some history behind them. And the middle is tight: second to fifth place sit three tenths of a point apart, so where each one wins matters more than the gap.

Say it louder here than on most pages: those scores are an editorial preview, not listening tests. They are our provisional opinion, published so the framework is transparent, and how we test explains the weighting in full. No Voxrater listening panel has scored these voices blind yet. The prices are a different matter: every rate below comes from the vendor’s own pricing page, with the capture date on the linked profile.

One disclosure up front, the same one on every roundup we publish. Some of these platforms run affiliate programmes we may earn from. The ranking is not for sale, no vendor saw this page before it went live, and if a platform ever pays to appear it will be labelled sponsored and kept out of the ranked positions, so a paid slot can never pose as an earned one.

1. ElevenLabs: the quality bar, priced like it

ElevenLabs is first because the voice itself is the best you can buy, and for narration that is most of the argument. The library runs past 10,000 voices in 70-plus languages, the cloning is the strongest in the business (instant for speed, professional for a long-term brand voice), and on a blind listen it is the one most people cannot tell from a human. Its narration-weighted score of 9.3 sits 1.4 points clear of second place, the widest gap anywhere on this page.

The rate is the trade. Narration runs on subscription credits at roughly $0.09 to 0.20 per 1,000 characters depending on tier and model, with the multilingual model around $0.10, ten times Speechify’s entry rate for the same script. Model choice matters more than people expect: v3 and Multilingual v2 are the narration engines, with v3 covering the 70-plus languages at the full credit rate, while Flash v2.5 costs about half as much per character but is built for real-time agents, not long reads. Paying the premium and then narrating on the cheap model buys the wrong half of the product.

The tiers are the real budget question. Starter at $6 a month covers 30,000 credits, about half an hour of finished audio, and the commercial licence arrives with it; Creator at $22 covers 121,000; Pro at $99 covers 600,000. The free tier’s 10,000 monthly credits are a listening post only, with no commercial licence and no cloning. Pick ElevenLabs when read quality is the deciding factor and volume is modest; at bulk volume the rate is the reason our ElevenLabs alternatives page exists. Full tier detail in the ElevenLabs review.

2. Cartesia: the fast engine hiding a narration bargain

Cartesia is second, and it is here almost by accident. Its Sonic engine is built for live phone agents (the sub-100ms speed pitch we ranked first on the low-latency page), but the same engine reads scripts, and worked back from its credit metering the cost comes out at roughly $0.035 per 1,000 characters. That is about a third of what the premium narration platforms charge, from an engine whose voice quality previews at 9 out of 10 on our sheet. A 10-minute script for about 32 cents is a genuine bargain.

Be clear about what that figure is, though. Cartesia no longer publishes a flat per-character rate; it meters Sonic in credits, so our $0.035 is a derived effective figure, our arithmetic on its published credit pricing rather than a line on its price list. It has held through repeated re-captures, most recently 11 July 2026, but treat it as a sourced estimate, not a quoted rate.

The trades sit in the library and the paperwork. The catalogue is around 100 voices across 42 languages, and Cartesia no longer states a single public voice total, so you are choosing from a far smaller set than ElevenLabs’ thousands; cloning from a three-second clip softens that for a branded read. Entry is cheap, with Pro at $5 a month including 100,000 monthly credits and the commercial licence. But the compliance posture is thin: SOC 2 and HIPAA are advertised for Enterprise and we have not verified either from a primary certificate, which matters if your scripts are medical or financial content. The Cartesia review shows the credit maths.

3. Fish Audio: the open-weight bargain, new to the table

Fish Audio is third, and it is the newest name here: we profiled it on 11 July 2026, so treat this placement as a first reading rather than a settled one. On the narration-weighted score it ties Cartesia at 7.9, and what earns that is the spread of strengths. The community library is put at 2,000,000-plus voices on the homepage, cloning works from a claimed 15-second clip, and the flagship s2.1-pro model is credited with 83 languages in the docs, though the homepage still says 30-plus and the release post’s training set spans about 50, a spread we flag rather than resolve.

The rate is the second-cheapest on this page. The API charges $15 per million UTF-8 bytes on every current model, and a byte is one character in English, so the derived figure comes out at about $0.015 per 1,000 characters, roughly 14 cents for a 10-minute script. Two catches sit right next to that number. Pay-as-you-go API access needs a paid subscription first, and the subscription cards disagree with their own pricing FAQ on how many credits a minute costs (600 to 625 in the FAQ, about 1,250 implied by the Plus and Pro cards), so get the conversion in writing before sizing a big tier on minutes.

The licence and paperwork lines are plain. Free is 8,000 credits a month, about 7 minutes, personal use only, with commercial use crossed out on the card; the commercial licence starts at Plus, $15 a month. Compliance is GDPR and nothing else documented, since the Enterprise card lists SOC 2 with no certificate anywhere we could find. The real differentiator is the exit route: the S2 model weights are open-sourced on GitHub and Hugging Face, so if the hosted service ever moves against you, the voices do not vanish with it. The Fish Audio review shows the byte arithmetic.

4. Speechify: the value pick with leaderboard evidence

Speechify is fourth, and on price it is not a contest. The entry rate is $0.01 per 1,000 characters, and it falls with volume: Starter at $10 a month includes a full million characters, Pro’s overage drops to $8 per million, Scale’s to $6 per million. The next cheapest engine here, Cartesia’s derived $0.035, is three and a half times the price. The catalogue holds up too, at 1,500-plus voices across 30-plus languages.

Cheap usually means an audible compromise, and here the evidence mostly says otherwise. The dated result we verified is Simba 3.0 placing 7th of 76 on the Artificial Analysis text-to-speech leaderboard in May 2026, already the cheapest model in that top ten. The current site banner claims the newer Simba 3.2 tops that board; that is the vendor’s claim about a third-party leaderboard and we have not re-verified it, so weigh the sourced seventh place, not the boast.

Two honest limits sit in the same breath as the price. The compliance paperwork is thin: SOC 2 Type II is now named on the Enterprise tier but there is no certificate you can open, and HIPAA is not mentioned anywhere, so regulated content should look elsewhere for now. And the platform is young in its 2026 shape: it moved onto speechify.ai, restructured its pricing between June and July, and shipped a voice-agents product this year, so expect the ground to keep moving under it. One genuine nicety the others do not offer: the free tier’s 50,000 monthly characters carry commercial use, so you can publish a real test read without paying anything. The Speechify review has the tier detail.

5. Murf: the easiest studio, sold in bundled hours

Murf is fifth on the strength of its workflow rather than its rate. It is the one tool here built entirely for people who do not want to touch code: paste the script, pick from 200-plus voices across 35 languages, tune pacing and emphasis in a timeline editor, export. For a marketing team cutting explainer videos or a course creator narrating modules, that studio is the product, and it is the easiest first hour in this group.

The pricing works in bundled hours of finished audio, not characters, and you should read it that way before comparing. Creator is $19 a month billed annually, which is $228 a year for 24 hours of generation; Business is $66 a month for 96 hours. Divide it out and finished audio costs about $0.16 a minute on Creator and $0.14 on Business, which is why our per-1,000-character figure of $0.16 is a derived equivalent, not a Murf price. Bundled hours also mean you pre-buy: an unused allowance is money already spent, and a heavy month hits the ceiling.

Two gates to know about in the same breath. Cloning your own voice is locked to the Enterprise plan, so a custom brand voice on a budget rules Murf out where ElevenLabs sells instant cloning on its $6 tier. And the free allowance is a one-off 10 minutes with no export and no commercial rights, so you can hear it but never use it. There is no phone-agent product at all, which costs Murf nothing here and everything on our calls pages. The Murf review covers the hours arithmetic.

6. Hume: the expressive specialist, dearest at entry

Hume is sixth, and it is the most distinctive voice in the list. Its Octave engine generates the words and the voice together, so the delivery carries real emotional register rather than one flat tone per sentence, and instead of a fixed library you clone a voice from about thirty seconds of audio or describe the one you want in plain words and have it built. For a branded voice with a specific temperament, or any read where the delivery is the point, nothing else here works quite this way.

The price is why it sits sixth on a narration list. Octave meters at $0.15 per 1,000 characters on the entry tiers, the top of this group’s range and fifteen times Speechify’s entry rate, and it only falls to $0.05 per 1,000 on the Business plan at $500 a month. The language list is modest too: Octave 2 covers about 11 languages, against 30-plus at Speechify and 70-plus at ElevenLabs.

Voice range is where the preview score gives most ground, a 5 out of 10 with no fixed library to browse, and under a weighting that prizes range that drags the weighted score to 6.5. One paperwork note in the same breath: Hume states HIPAA compliance with a BAA on request, more than most of this group can say, but SOC 2 and GDPR appear only as a pricing-page footer with no openable certificate. Commercial use starts at the Creator tier, $14 a month at its regular price. The Hume review explains how the Octave engine and the EVI agent product split.

7. Rime: the phone engine you can press into narration

Rime is last, and the honest framing is that narration is not its fight. Rime’s pitch is pronunciation accuracy on high-stakes phone calls, the names, medications, account numbers and addresses that quietly lose a customer when fumbled, and its models are tuned for that live use. But it publishes a real per-character rate, its 300-plus voice library reads scripts perfectly well, and a team already running Rime for calls may reasonably ask whether it can cover the voiceover work too. It can, with limits worth naming.

The rate is $0.05 per 1,000 characters, and that is a starting-at figure. Rime redesigned its pricing page between June and July 2026 and removed the old per-model rates (Mist was $0.03), so the single published number now implies volume discounts the page does not spell out, and parts of the page still held placeholder text at our 11 July capture. If you are negotiating, get the current rate card in writing. At the published rate a 10-minute script runs about 45 cents: cheaper than ElevenLabs and Hume, dearer than Cartesia and Speechify.

The narration limits are plain. Seven languages, the shortest list here by a distance, and a voice line-up organised around call styles (professional, formal, casual, energetic) rather than performance reads, which is why range and quality preview at 6 and 7 for a weighted 6.4. On the other side of the ledger, the free start is generous at 3,000 minutes, and the Enterprise tier states a HIPAA BAA, a real differentiator if your scripts are healthcare content. Full workings in the Rime review.

What a 10-minute script actually costs

Here is the arithmetic laid bare, because a per-1,000-character rate only means something once it touches a real script. A finished minute of narration comes out of roughly 900 to 1,000 characters, so call a 10-minute script 9,000 characters. Multiply each engine’s stored rate by nine and you get the generation cost of that script. Cheapest first; note this is a cost order, not the ranking above.

Four caveats, each flagged in its section above: Cartesia’s rate is derived from credit metering, Fish Audio’s is derived from its per-byte API rate (which needs a paid plan before you can use it), Murf’s is derived from bundled hours you pre-buy, and Rime’s is a starting-at figure with unpublished volume discounts. And the generation cost is rarely the real bill anyway. You pay for the subscription tier that carries your volume and your commercial licence, not just the characters. To model your own volumes rather than our 10-minute example, the cost calculator runs any script length against every engine and model here.

Free tiers, and whether you may publish from them

Every engine here has a free way to hear itself, but free to hear and licensed to publish are different things, and the difference bites after the video ships. The licence column is the one to read twice.

Speechify is the odd one out in a good way: a small channel could genuinely run on its free tier alone. Everyone else holds the licence back for a paid plan, though the cheapest tickets are small at $5 to $6 a month. Treat Murf’s free minutes as a listening booth, nothing more, and confirm Rime’s free-allowance terms in writing before anything commercial ships on it.

Who this list is not for

Live phone work. If the voice is answering calls rather than reading scripts, the weighting flips entirely: speed and running cost dominate, the business-calls ranking is the page you want, and the low-latency roundup covers the engines built for the speed race. Quote-only enterprise suites are also absent on principle: if a platform will not publish a price, we will not slot it into a numbered position against ones that do. And as on every roundup, our own site never appears in its own list. The only names here are engines you could actually publish with.

Before you commit, run the same script through three of them

Do not pick from this page alone, ours included. The point of the free tiers is that a narration decision can be tested in an afternoon, with the script you will actually publish. The narration comparison table puts the sourced numbers side by side while you listen.

• Use your real script, not demo copy. Brand names, figures and the words your audience would notice mispronounced are exactly where engines separate. Demo copy is chosen to flatter.
• Listen at minute three, not sentence one. Every engine sounds good for ten seconds. Pacing drift and flat emphasis only show up in the long read, and the long read is what you are buying.
• Compare the same class of voice. A flagship ElevenLabs voice against a stock voice elsewhere tells you about the voices, not the platforms. Pick each engine’s best fit for your brief, then compare.
• Price your real monthly volume. Run your actual character count through the calculator and check which tier it lands on, because the tier, not the headline rate, is usually the bill.

Bottom line

There is a clear quality winner and a clear price winner, and they are not the same engine, which is why this page does not end on one name.

• The best read and the biggest library: ElevenLabs, around $0.10 per 1,000 characters.
• The cheapest serious quality, with a commercially usable free tier: Speechify, $0.01 per 1,000.
• One cheap engine for both calls and narration: Cartesia, roughly $0.035 per 1,000, derived.
• The open-weight newcomer with the second-cheapest rate: Fish Audio, about $0.015 per 1,000, derived.
• The no-code studio for teams who live in a browser: Murf, bundled hours, easiest workflow here.
• The delivery specialist when the emotion is the brief: Hume, $0.15 per 1,000 at entry.
• The phone engine that can moonlight on scripts: Rime, $0.05 per 1,000, starting-at.

If you are torn, let the script break the tie. Run it through ElevenLabs and Speechify first, and pay the quality premium only if your own ears vote for it. Then check the narration rankings for the current scores, and put your monthly volume through the calculator before you commit, because the cheapest rate on the wrong tier is not the cheapest bill.