ElevenLabs V4 vs FishAudio Drama 3: Emotional AI Voices Compared
Compare ElevenLabs V4 vs FishAudio Drama 3 with 14 audio samples, identical-script tests, emotional voice prompts, and October 2026 pricing.
ElevenLabs V4 and FishAudio Drama 3 are, for us, the two standout AI voice releases of the past month. ElevenLabs launched V4 on September 28; FishAudio followed with Drama 3 on October 13. Both are chasing something more interesting than a smooth reading voice: a performance you can actually direct.
That’s an exciting prospect if you make dubbed videos, character dialogue or narrated stories. A character might sound cheerful while trying not to cry. An angry accusation might collapse into an embarrassed laugh. Those little turns are where a scene comes alive—and where a voice that simply reads the words can leave you cold.
V4 arrives with ElevenLabs’ focus on expressive delivery and keeping a speaker recognizable through a scene. Drama 3 brings an open-weight model with a strong emphasis on emotional direction. We wanted to give both room to show what they can do, then take away the acting instructions and see how they handle the same everyday scripts.
Below, you’ll find the exact prompts, 14 playable takes, the settings we used, and a pricing comparison checked on October 16, 2026. Bring headphones: the interesting moments here are often a quiet phrase or a change of pace.
Inside this guide
- The models at a glance
- How we designed the tests
- Emotional voice showcases
- Same-script baseline comparisons
- What to listen for
- Pricing and promotional offers
- Where this leaves the comparison
ElevenLabs V4 vs FishAudio Drama 3 at a glance
| ElevenLabs V4 | FishAudio Drama 3 | |
|---|---|---|
| Release | September 28, 2026 | October 13, 2026; Preview |
| Performance controls | Natural-language instructions and inline audio tags | Natural-language instructions, inline tags and scoped directions |
| Published architecture | New architecture; no parameter breakdown in the launch post | 35B-parameter mixture of experts, with about 3B active per token |
| Advertised languages | More than 90 | 83 |
| Our showcase voice | Alicia | Freya, female British English |
| Our baseline voice | Alicia | Verity, male |
| What we generated | Two emotional scenes and three shared scripts | Two emotional scenes and three shared scripts |
The language figures are vendor claims; our clips are in English. Architecture details come from the V4 introduction and the Drama 3 technical report.
ElevenLabs V4: can the voice follow the scene?
What interests us about V4 is its promise to read the intention behind a line. The words “wake the team” could be an order, a plea or a moment of quiet wonder. Our first scene builds toward exactly that choice.
ElevenLabs describes improved handling of tone, pacing and inline directions, alongside stronger speaker consistency. We tested standard V4 with one voice, Alicia. V4 Turbo is a separate model aimed at faster conversational responses; its latency claims aren’t results from this comparison. ElevenLabs’ V4 introduction.
FishAudio Drama 3: room for mixed emotions
Drama 3 caught our attention because its launch examples lean into feelings that pull in different directions. A forced smile over grief is a much more interesting challenge than simply asking for “sad.” That became the starting point for our workshop scene.
Under the hood, Drama 3 uses a mixture-of-experts backbone: roughly 3B parameters are active at each token out of 35B total. Its audio decoder builds the acoustic detail of the speech. Those are useful architectural differences to understand, but the listening samples are where the creative promise has to earn its keep. Drama 3 technical report.
We used the official FishAudio text-to-speech app for the clips below. The research also explores editing an existing performance; this article tests fresh generations, not those editing features.
How we gave both models a fair chance
First, let each model stretch
We studied the vendors’ example prompts, then wrote our own scenes with different situations and wording. Each has several emotional beats and a quieter landing. The idea was to give both models a rewarding acting challenge, rather than recycle a launch demo.
There are two showcase scripts per model and two takes of each. We kept both takes so you can hear variation. This explores what each model can do with tailored direction; different scripts aren’t a head-to-head contest.
Then, give them exactly the same words
For the baseline round, we wrote three fresh scripts without borrowing from either vendor’s examples: a reunion, a detailed announcement and a quiet museum opening. Both models received identical text, punctuation and capitalization, with no acting tags or pronunciation hints.
“Neutral” doesn’t mean lifeless. The reunion still has a story to tell; we’re asking whether the voice finds it without us spelling out the emotion.
Each baseline presents take one from each model. FishAudio’s app automatically generated a second take, which remains in our research record; using the first take consistently keeps the comparison from favoring a hand-picked result.
The voices and settings behind the clips
| Setting | ElevenLabs V4 | FishAudio Drama 3 |
|---|---|---|
| Generation setup | API-generated audio | Official text-to-speech app |
| Showcase voice | Alicia | Freya – Female British English, by Fish Official |
| Baseline voice | Alicia | Verity, male preset |
| Controls | Stability 0.5; similarity 0.75 | Volume 0; speed 1x; Auto Tag off |
| Takes presented | Two per showcase; one per baseline | Two per showcase; one per baseline |
We chose Freya for the showcases because Fish lists her as a female British English voice. All three baselines use Verity. Keep that difference in mind: preferring a speaker’s tone isn’t necessarily preferring the underlying model.
Every player uses a 96 kbps MP3 listening copy. We compressed the exports for the page without trimming, splicing, normalizing or enhancing them. Lossy compression can soften fine detail, so these copies are best used to compare delivery rather than tiny differences in audio fidelity. We also aren’t ranking generation speed: the two setups didn’t give us equivalent timing measurements.
The emotional scenes: give the voices something to act
ElevenLabs V4 showcase 1: fatigue, discovery and wonder
Six hours staring at an empty patch of sky, and then something changes. This scene gives V4 a journey from exhaustion to discovery, with a final instruction that should feel almost too precious to say loudly. Listen to “Wait,” and then to “Gently.” Can the voice make both turns without overplaying them?
The script we sent:
[quietly, exhausted but attentive] The telescope has been pointed at the same patch of darkness for six hours. Nothing. Wait. [suddenly alert, holding back excitement] There, beside the blue star. One pulse. Two. [laughs softly in disbelief] We checked that sensor three times. It is not the sensor. [warm, breathless with wonder] Wake the team. Gently. After all these years, the universe has finally knocked.ElevenLabs V4, take 1 · 20.4 seconds
ElevenLabs V4, take 2 · 20.2 seconds
ElevenLabs V4 showcase 2: reassurance under pressure
Here’s a different kind of pressure: the curtain is moving, the performer is nervous, and someone has to keep the show together. The speaker addresses three audiences in a few lines—the performer, the crew, and finally herself. We’re looking for those shifts in intention, especially the move from command to proud relief.
The script we sent:
[calm and reassuring, speaking to a nervous performer] Look at me. You know the first line. You only need the first line. [whispering urgently] The curtain is moving early. Keep walking; I will handle it. [firm, controlled, addressing the crew] Hold the lights. Give her the stage. [softly, with proud relief] There she is. I knew she would find her voice.ElevenLabs V4, take 1 · 15.1 seconds
ElevenLabs V4, take 2 · 15.4 seconds
FishAudio Drama 3 showcase 1: warmth with grief underneath
A broken radio is an excuse to talk to someone who isn’t there. We wrote this scene to explore the gap between what a character says and what she can bear to admit: affectionate teasing, a laugh that catches, then one whispered confession. Both takes use Freya, FishAudio’s female British English voice. Listen for whether the warmth survives the grief.
The script we sent:
[trying to sound cheerful while holding back tears] I finally fixed your ridiculous little radio. It only gets one station, and the volume knob still falls off. [a fragile laugh that catches in the throat] So, exactly how you left it. <whisper>I thought fixing it would make this room feel less empty.</whisper> [takes a breath, steadier but still tender] Anyway. Your song is on. I will leave the light on for a while.FishAudio Drama 3 · Freya, take 1 · 22.4 seconds
FishAudio Drama 3 · Freya, take 2 · 20.3 seconds
FishAudio Drama 3 showcase 2: anger, discovery and embarrassment
This one should be fun. Our speaker is halfway through a beautifully controlled complaint when she discovers the missing envelope under a coffee cup. The challenge is comic timing: let the anger build, give the discovery a beat, then allow relief and embarrassment to catch up. Freya handles both takes here too.
The script we sent:
[restrained anger, precise and quiet] You mailed the only copy to the wrong address. [trying to remain composed] No, I am not shouting. I am counting. One envelope. Six months of work. [a short, incredulous laugh] And you used the decorative stamp. <whisper>Wait. Is that it under your coffee?</whisper> [relief breaking into laughter, then embarrassed composure] Right. Excellent filing system. I owe the postal service an apology.FishAudio Drama 3 · Freya, take 1 · 25.8 seconds
FishAudio Drama 3 · Freya, take 2 · 23.6 seconds
Same script, different model: the baseline round
Reunion: emotion from context
After those heavily directed scenes, let’s take the tags away. This reunion ends with an offer of tea, not a grand speech. Does the voice find the anticipation and relief while letting that small gesture stay small?
The same script for both models:
I kept your seat by the window. For three months, everyone told me to move the chair. Then your message arrived this morning, just four words: I am coming home. I have rehearsed this moment a hundred times. Now you are here, and all I can think to ask is whether you would like some tea.ElevenLabs V4, take 1 · 19.0 seconds
FishAudio Drama 3 · Verity, take 1 · 18.4 seconds
Precision: numbers and contextual pronunciation
Now for a less glamorous test that matters enormously in real work. A lovely voice is no help if it reads the wrong price. Follow both amounts, the time and the access code, then listen to how each model handles “read,” “record” and “live” in context.
The same script for both models:
The live broadcast begins at 7:05 p.m. on October 21. Please read the revised total: $1,204.50, not $1,240.15. I read the first draft yesterday. Today I will read the final version, record the result, and check the record before we go live. The access code is A, seven, B, zero, nine.ElevenLabs V4, take 1 · 27.7 seconds
FishAudio Drama 3 · Verity, take 1 · 26.6 seconds
Restraint: a quiet narrative arc
The museum is ready to open after a near disaster. Nobody cheers; the door simply gets unlocked. Listen for an easy pace and a sense of relief that never needs to announce itself. It is a short scene, but it asks for something easy to overlook: knowing when a calm reading is enough.
The same script for both models:
The museum opens in ten minutes. The lights are working, the labels are straight, and the painting is finally back on the wall. Yesterday, I was certain we would have to cancel. This morning, the room looks almost ordinary. That is what surprises me. After all that work, success can look like a quiet room with the door unlocked.ElevenLabs V4, take 1 · 21.4 seconds
FishAudio Drama 3 · Verity, take 1 · 20.0 seconds
What to listen for when you compare
The most revealing moment may not be the biggest emotional flourish. It might be the word just after a laugh, or the last sentence of a take. Does the voice recover naturally? Does it finish the thought? Would you use the whole clip, or reach for the regenerate button?
We checked that all 14 embedded files download and decode correctly. The acting and pronunciation review is still pending, so we haven’t turned those technical checks into a quality score. These are the questions we’ll use for the listening verdict:
| Listen for | A useful checkpoint |
|---|---|
| The words | Does every sentence survive, including the ending? |
| Emotional turns | Does the delivery change at the phrase that calls for it? |
| Restraint | Can the voice leave an ordinary line alone? |
| Pronunciation | Are the numbers, codes and context-dependent words correct? |
| Character | Does it still sound like the same speaker after a strong emotion? |
| Practical usefulness | Could you use the take intact, or would it need repairs? |
Pricing comparison: checked October 16, 2026
The price gap deserves attention, especially while Drama 3 is on sale. But the launch price and the price you budget for a continuing project are two different numbers. Here’s the snapshot in US dollars, with introductory offers separated from regular prices. Taxes and checkout terms can affect the final bill.
Creator subscriptions
| Plan | Current introductory offer | Regular monthly price | Included monthly allowance |
|---|---|---|---|
| ElevenLabs Starter | $1 for the first month; offer advertised through October 18, 2026 | $6 | 30,000 credits |
| ElevenLabs Creator | $11 for the first month | $22 | 121,000 credits |
| FishAudio Plus | $7.50 for the first month under the launch offer | $15 | 250,000 credits |
| FishAudio Pro | $50 for the first month under the launch offer | $100 | 2,000,000 credits |
Sources: ElevenLabs creator pricing, FishAudio plans and Drama 3 launch offer. FishAudio’s first-month amounts apply its advertised 50% discount to the monthly list prices. Credits are provider-specific, so a bigger allowance alone doesn’t tell you how much comparable V4 or Drama 3 audio you’ll get.
API generation: compare the same unit
| Model | Price per 1 million input characters on October 16 | Regular published rate |
|---|---|---|
| ElevenLabs V4 | $80 | $80 |
| FishAudio Drama 3 | $10 during the launch promotion | $50 |
ElevenLabs publishes V4 at $0.08 per 1,000 characters, converted here to $80 per million. FishAudio publishes Drama 3 at $10 per million during the offer and $50 at list price. These are generation rates, separate from the creator subscriptions above. ElevenLabs API pricing · FishAudio’s dated offer.
Reading this after November 12, 2026? FishAudio’s announced API promotion runs October 13–November 12. Unless extended or replaced, budget for $50 per million characters afterward, five times the promotional rate. Its creator offer has a different clock: eligible new Plus, Pro and Max subscriptions receive 50% off their first month on monthly billing or first year on annual billing. That is not a permanent half-price renewal.
For a concrete example, 100,000 input characters would cost about $8 with V4, $1 with Drama 3 during the offer, or $5 with Drama 3 at list price, using those published API rates before taxes and any account-specific terms. Reruns add usage. If a lower-priced voice needs several attempts to get a line right, the saving can shrink quickly.
That’s why we’d track cost per usable take alongside the headline rate. Our AI dubbing cost guide looks at the other costs that show up once speech generation becomes a finished video.
Drama 3 also offers downloadable weights under a community license. Self-hosting is another route, but compute costs and license conditions still belong in the budget.
Where this leaves the comparison
What makes these releases worth exploring is the possibility of giving a voice more precise creative direction: not just “sound happy,” but a discovery that starts quietly, or a joke that only works because someone tries to stay serious.
For your own audition, start with the scene closest to your work. Try the workshop for emotional subtext, the envelope for comic timing, and the announcement for speech that has to get every detail right. Then listen to the quieter baseline. A voice that can deliver a dramatic moment also needs to know when to step back.
Drama 3 has the lower published API rate in this snapshot, even after its scheduled promotion ends. Whether its performance—or V4’s—is the better fit is the listening question we still need to answer. We’d rather leave that choice open than mistake a compelling feature list or a low price for a finished review.
If you’re developing your own cast, our guide to creating character voices includes voice-brief ideas to try. The analog horror voice guide explores a very different mood.