How to Choose the Right AI Voice, and Which One

Choose the accent your listeners will meet first, and the voice second. For a story in English, OpenAI's voices are a reasonable place to start. Six of the nine in Kukarella are catalogued female. OpenAI's own demo reads its bedtime story with Sage and its dramatic scene with Ash. But OpenAI says its voices are "currently optimized for English". For a lesson in Mexican Spanish or Canadian French, a native voice of that variety serves learners better. And do not trust the label at the top of a list. In our own Voice Library, under "Spanish (Argentina)", 2 of the 67 voices listed on 1 October 2026 were Argentine.

The short answer, in five steps:

1. Language and country. Pick the variety your listeners will hear outside the classroom: Mexican or Castilian Spanish, British or American English.
2. Native first. Pick a voice native to that variety. Use a voice that reads many languages only when one voice must speak several of them.
3. The job. Calm and even for instructions, warmer for stories, two clearly different voices for a dialogue.
4. Gender and age last. Names and labels mislead. Listen.
5. Your own hardest line. Test it, with its numbers and names, before you commit.

When OpenAI cast the voices for ChatGPT in 2023, more than 400 submissions arrived in under a week; five voices were chosen, recorded and launched in a process that took five months. Those are ChatGPT's voices, not the ones in Kukarella. But the method works anywhere: the team wrote down what the voices had to be like before anyone listened. Its brief asked for "An approachable voice that inspires trust". It said nothing about gender, or about "best". You have one lesson and a library of 2,000+ voices. This page walks through the same decision for a Spanish teacher, a safety trainer and a children's story.

We make Kukarella, a text-to-speech tool, so keep that in mind. Where a human voice, your own voice or another tool is the better choice, this article says so.

The same words, three different speakers.

Which OpenAI female voice is best for storytelling and drama?

Short answer: Kukarella offers nine of OpenAI's voices, and our catalogue lists six of them as female: Alloy, Coral, Fable, Nova, Sage and Shimmer. OpenAI's own demo site, openai.fm, pairs its "Bedtime Story" preset with Sage. It pairs its "Dramatic" preset with Ash, a voice we list as male (openai.fm source code, April 2025). But these voices are tuned for English, so choose the voice by ear and write the drama into your words.

For a warm storyteller, start with Alloy, Nova and Shimmer. Fable gives an even, professional read. Sage suits an older, sterner character better than a narrator, and Coral sounds like a brisk lecturer. For a male narrator with weight, or a villain, try Ash, the voice OpenAI's demo uses for drama, or the slower Onyx.

What OpenAI itself says

OpenAI's text-to-speech guide lists 13 voices by name only. Its one piece of advice about a particular voice is this: "For best quality, we recommend using marin or cedar." Kukarella offers nine of the 13, the older set that OpenAI's tts-1 models support. We do not offer marin or cedar, nor ballad or verse. If those two are what you want, you will need to use OpenAI's API directly.

Note one detail: OpenAI's "Bedtime Story" preset asks Sage for "a gentle, curious narrator with a British accent" (openai.fm). The accent is part of the brief.

The nine voices as we hear them

The gender labels below are ours, made by listening, and we got two wrong: until August 2026 our catalogue listed Alloy and Fable as male.

To us, Sage sounds older than the others, and Echo sounds like a teenager rather than a grown man. OpenAI chose Sage for a bedtime story; we hear Sage as strict. That is the first reason to listen for yourself.

Listen for yourself: the same two lines in each of the six voices, picked under English (US), with no style.

Alloy:

Coral:

Fable:

Nova:

Sage:

Shimmer:

For a male narrator, Ash:

And the slower Onyx:

The "buts" before you choose

• Tuned for English. OpenAI lists 50+ languages for its speech model, "despite voices being optimized for English" (OpenAI). In October 2025, someone building a Japanese-learning app asked OpenAI's forum for a soft female voice. One forum member warned of "a significant American foreigner accent". That is one user's view, not OpenAI's.
• No accent setting. In our Voice Library the nine OpenAI voices appear under every one of the 152 languages and accents. Do not expect them to carry the accent of the entry you pick them under.
• Fewer controls. Pauses work on them. Speed, pitch and volume do not, so the pace you hear is the pace you get.
• A name is not a description. The voice called "Captivating Storyteller" is a MiniMax voice, catalogued male. Its name will not tell you whether it suits your story.
• A professional narrator still does better. In a 2024 study of 197 adults, a professional recording beat modern synthetic voices on recall, on applying what was learned, and on ratings (Siegle and Craig).

The Style menu on these voices offers presets such as Dramatic. A style is a request. Some voices follow it, many barely change. Write the feeling into the words, then listen.

Does a female voice tell a story better?

Short answer: the evidence is weak and mixed, so decide the role and the listener first, and gender after. But listeners do hear gender in voices, even clearly synthetic ones.

In a study presented at CHI 2000, people stereotyped even clearly synthetic voices by gender (Lee, Nass and Brave). Preferences also follow habit. In German surveys by the digital association Bitkom, two-thirds of voice-assistant users preferred a female voice in 2022, and six in ten in 2023. Bitkom's expert put it down to habit: the assistants speak in a female voice by default.

The best-known story about voice gender is shaky. The Stanford researcher Clifford Nass told NPR in 2010 that BMW "actually had to have a product recall because German drivers would not take directions from a woman" (NPR). We found no BMW record of a recall. Every telling leads back to him. And when Bitkom asked German navigation users in 2011, 46% preferred a female voice and 9% a male one. The useful part of Nass's account is his method. His team chose the voice's role first, a co-pilot, and only then described it: "Male, not at all dominant, somewhat friendly and highly competent" (PBS).

So a female narrator is a fine choice for a story. Then comes the question that matters more: who will listen, and in which accent? If your story is for learners of English in London, or of Spanish in Buenos Aires, the next sections decide the voice.

Who will listen, and in which accent?

Short answer: choose the variety of the language before the voice: the one your listeners will hear outside the lesson. Exam boards mix accents on purpose, and research shows that unfamiliar accents cost listeners points. But that variety is often not the one you speak.

Listeners hear a person in every voice, and they notice whose. When Apple added new Siri voices in 2021, Jason Allen, a public relations manager in the San Francisco Bay Area, told Consumer Reports: "You play it back and you hear someone who sounds like you, or a friend of yours, or someone in your family." (Consumer Reports, April 2021)

Exam boards build this in. IELTS says of its listening test: "Different accents, including British, Australian, New Zealand and North American, are used." (IELTS). TOEFL says you may hear the same four. Every SIELE Spanish test must combine "un mínimo de tres variedades diferentes del español", at least three varieties of Spanish.

Research explains why. In a study of 21,726 TOEFL test-takers, both the strength of a speaker's accent and the listener's familiarity with it changed listening scores, "even with quite light accents" (Ockey and French, 2016). In a 2003 experiment, students who heard a lesson in a standard accent learned more than students who heard it in a foreign accent (Mayer, Sobko and Mautone, Experiment 1; effect size d = 0.90, as listed by Dinçer, 2022). In a British Council survey published in July 2024, 80% of 1,112 English teachers in 115 countries and regions agreed that AI "should be developed to support the learning of different varieties of English".

The accent your learners need is often not your own: you may speak Castilian Spanish while your students are moving to Mexico. As our 2024 article on this site put it, a voice cannot be good or bad, only suitable or not.

Your listeners decide which voice is right, not the library.

What are the most important things to look for in AI text-to-speech?

Short answer: check these eight things, in this order. Most checklists start with quality and naturalness, but a natural voice in the wrong accent still teaches the wrong accent.

1. The language and the country your listeners will hear.
2. A native voice of that country, and enough of them.
3. The job: one narrator, a conversation, or an announcement.
4. Age and gender, chosen by ear, not by the name or the label.
5. Numbers, names and acronyms, read the way your listeners must hear them.
6. The controls on the voice you chose, not "on the platform": pauses, speed, pronunciation.
7. A test on your hardest line, at full length, in the room where it will play.
8. Whether you can keep the voice for the whole course, and get audio you can drop into it. Kukarella downloads MP3 or WAV.

How to choose an AI voice for a lesson: one case, step by step

Short answer: name the country, open the Voice Library on it, pick native voices, make the speakers sound different, then test your hardest line. But the list under a country is not all from that country, so step 3 matters most.

A lesson is where our own library is easiest to get wrong. Say you teach Spanish, your students are going to Mexico, and you want a short market dialogue they can hear several times.

Step 1: Name the variety your learners will meet

Your students will hear Mexican Spanish, so the entry you need is "Spanish (Mexico)", not "Spanish (Spain)". Kukarella lists English in 14 accents and Spanish in 23, but the number of voices differs a great deal from one country to the next.

Step 2: Open the Voice Library on that country

In Text to Voice, click "Add voice" in the left sidebar. In DialoguesAI, create a speaker and click "Select voice". In the window that opens, type the language in the search box on the left, then choose the country: "Spanish (Mexico)".

The search box finds languages, not voice names. So write down the name and country of a voice you like, or click its heart. "Favorite" under "Filters" shows your hearts again.

The Voice Library: languages and countries on the left, voices with a play button on the right.

Our own video walks through the same window, step by step:

Step 3: Pick a native voice, not one that reads every language

On 1 October 2026 the list under "Spanish (Mexico)" showed 148 voices, and 26 of them were Mexican native voices. A native voice speaks one language in the accent of the country you picked. The rest read many languages:

• The nine OpenAI voices. They appear under all 152 entries. Do not expect them to sound Mexican.
• Microsoft voices whose names end in " M." They take the accent of the entry you pick them under. I heard one of them, Ada M., pronounce tomato, potato and garage differently under English (UK) and English (US).
• MiniMax voices. Under Spanish (Mexico) they play a recording filed under Spain.
• 48 more voices with a "Multilingual voice" label. They appear under most varieties of the big languages and carry no accent of the entry you pick.

There is no filter for native voices. OpenAI's voices and some others carry a "Multilingual voice" label, Microsoft's end in " M.", and MiniMax voices carry no mark. Their names read like descriptions, such as "Confident Woman". So listen. In my own listening, a voice that reads many languages often sounds right for about fifteen seconds. Then a stress or a vowel gives it away, and a learner may copy it.

The gap is widest where a variety has few native voices. Under "Spanish (Argentina)" the list showed 67 voices, and two of them, Elena and Tomas, were Argentine.

Under Spanish (Argentina), 2 of the 67 voices listed were Argentine on 1 October 2026.

Native voices on the same day:

Two voices are enough for a dialogue between a woman and a man, not for a course that needs six different Argentine speakers. For that, see when a human voice is the better choice.

Hear the difference: one Spanish sentence, read by a native voice of each country.

From Spain:

From Mexico:

From Argentina:

And one English line, first in Arlo, a native British voice:

Then in Alloy, an OpenAI voice, picked under English (UK):

When a multilingual voice is the right choice:

• a lesson that switches between two languages sentence by sentence, in one voice;
• one presenter across several language versions of a course, where sounding like the same person matters more than the local accent;
• a script with a few foreign names or words: start native, put those words in your test, and switch only if they come out wrong.

When people write to me that a voice sounds bad, the answer is often a different voice in the same library, usually a native voice of their country. But in two of those cases I suggested a multilingual one, because for what they were making it sounded more natural. The voice was rarely broken. It was the wrong one for the job.

Step 4: Cast two voices that sound clearly different

Give each speaker a native voice of the same variety that differs in gender, age or pace, unless the lesson is about comparing accents. DialoguesAI's assistant can write the market scene and give each speaker a voice in seconds. But it picks "from the same fixed set of 22 voices" (Kukarella home page FAQ). It matches each character's gender and age, never your script's language or accent. So the accent is your one job after it. Click the speaker's card, click "Voice", choose a native voice under "Spanish (Mexico)", then click "Save & close". Every line of that speaker follows.

Across a unit, use more than one voice of the variety. In a 2008 study, listeners trained on several speakers of an accent then understood a new speaker of it (Bradlow and Bent; native listeners and a foreign accent, not learners). Duolingo, writing in 2021 about how it voiced nine characters in five languages, put it this way: "hearing a variety of voices in your lessons is so important: it helps you develop flexible listening skills for real-life language situations!"

Step 5: Test your hardest line, not the sample

Press play beside a voice to hear its sample. The sample reads a stock paragraph, not your text. A padlock stops you using a voice, not hearing it. But the nine OpenAI voices have no sample in 33 of the 152 entries, among them Mandarin (China) and the general "Arabic" entry.

So build one test paragraph and run it through each voice on your shortlist:

1. Your hardest sentence, long, with clauses. For the market: "¿Cuánto cuesta el kilo de jitomate, si me llevo dos y usted me regala el cilantro?"
2. A number or a year, and a phone number. In September 2026 I typed "It happened in 28 28" into a range of our voices. Some read a year, some two pairs, one a single four-digit number. The reading varied voice by voice, even within one provider.
3. An acronym. Every voice I tried read acronyms letter by letter, but some ran the letters together until they sounded like a word.
4. A name, or a word from another language.
5. A question and an exclamation.
6. A word with two readings. Duolingo's example: "read" in "I read the book" sounds different in a present-tense lesson and a past-tense one.
7. An instruction: "Turn to page twelve and listen again."
8. For a dialogue, one exchange between your two speakers. Can you tell them apart with your eyes closed?

Listen past the first fifteen seconds and judge the whole paragraph. Researchers at Google found that rating sentences one by one and rating whole paragraphs gave different results for the same voices (Clark and colleagues, 2019).

When you test again, change a word: the same text with the same voice may play back the audio already made. Each test uses 1 credit per character on any catalogue voice, HD voices included. A guest can try 10,000 characters and a free account 30,000. My own rule: choose a voice that already reads your text the way you expect. It beats fixing a voice with settings.

If a line keeps coming out wrong, change the voice before you fight the settings.

The tool below turns five answers about your learners into the Voice Library entry to open, how many native voices it holds, a first-minute test and what to listen for. It is a starting point, not the final choice: you still listen.

How do I choose the best AI voice for training videos?

Short answer: choose one calm, clear voice in your staff's own variety, test every number in it, and keep it for every module. But voices can change or leave a library, so write down which one you chose.

Say you train warehouse staff in Quebec. There is no Quebec entry: choose "French (Canada)", which had 56 native voices on 1 October 2026. Write your brief in one sentence: who is speaking, to whom, and why. For example: "A calm narrator explaining a safety procedure to new staff."

A voice can lose your staff's trust before it costs them any learning: in a 2020 study, voice quality did not significantly change learning, but it did change trust (Chiou, Schroeder and Craig). And match the energy to the listener's state, not to your mood. In a 2005 driving-simulator study, 40 people were first made happy or upset. Then they drove while the car's voice spoke in an energetic or a subdued tone. When the voice matched the driver's mood, "drivers had fewer accidents" (Nass and colleagues). A module after an incident may want calm, not cheer.

The wrong tone for the moment.

A lively voice will not rescue a muddled script: in a 2001 study, 135 people listening to training lectures remembered most when the lecture was both expressive and well organised (Towler and Dipboye). In a 2009 follow-up with 132 participants, the expressive lecture scored best on problem solving only when it also carried interesting asides. For recall, trainees focused on mastering the material did best with an expressive lecture without them (Towler).

Problem-solving scores by trainer expressiveness, in the design of Towler (2009).

Safety training is full of numbers, and voices read them differently: listen to every number once. Then keep the voice. Write down its name and its language entry, and click its heart. A voice can leave our library when its maker retires it. It is rare. If it happens, send us a few recordings you made with it, and we can clone it back for everyone (home page FAQ). Voices also drift. Amazon warns of "slight variations" after model updates across "a season of podcasts" (Amazon Polly). ElevenLabs says v4 may sound "substantially different" from v3 for the same voice (ElevenLabs). Make a module in one sitting, and listen again before you add a lesson months later.

A training video is not a show: match the tone to the content.

Last, tell your staff the voice is AI. OpenAI's usage policies require apps that use its voices to tell listeners "that the TTS voice they are hearing is AI-generated" (OpenAI). That binds the app, not you, but it costs nothing.

Which voice for a children's story?

Short answer: start with one warm narrator in the children's own accent, not many character voices. But child voices are few, and a name will not make a voice young.

Enthusiasm helps children listen. In three studies, fourth and fifth graders recalled more of a text read aloud with enthusiasm, partly because it held their attention (Moè and colleagues, 2021). Those readers were people. When researchers made a synthetic voice sound more enthusiastic, scores did not rise. In a study of 244 students, enthusiastic versions of Alexa's voice did not improve what they remembered or applied. Only one of three versions lifted their mood a little (Liew and colleagues, 2023).

So choose the narrator's warmth with the voice, not afterwards. Our Voice Library has no age filter and few child voices. When our assistant needs a boy under ten, it casts a voice called "Assertive Queen" or "Kind-Hearted Girl", because none of the 22 voices it casts from is catalogued as a boy.

A style cannot make a voice younger. Choose a voice that already sounds the right age.

If the children speak Spanish, French or Portuguese, go back to step 3: their own accent first. With so few child voices, a story with several child characters may still need actors.

Which voice AI avoids the "robotic voice" effect?

Short answer: no platform avoids it on every line. When a voice sounds robotic in a lesson today, the cause I see most often is the wrong voice for the language. But a good modern voice in the right variety can teach as well as an ordinary human recording.

What 'robotic' usually means: flat intonation, odd pauses, no breath.

The old rule against machine voices came from old machines. In 2003, students learned more from a human voice than from a machine voice (Mayer, Sobko and Mautone, Experiment 2; effect size d = 0.79, as listed by Dinçer, 2022). In a 2017 online study, a modern synthetic voice matched a recorded human voice on most measures and scored higher on applying what was learned (Craig and Schroeder). A professional voice actor still did better in 2024 (Siegle and Craig).

In a 2020 study, 29 Brazilian learners of English understood a synthetic voice as well as a human one, but rated it less natural on average: 3.12 against 4.58 out of 6 (Bione and Cardoso).

What gives a voice away? In a study run in 2018 and 2019 with 95 adults, listeners named flat or "too perfect" intonation, odd pauses, no breathing, and emotion that does not fit the words (Kühne, Fischer and Zhou, 2020). For a teacher, add a cause they did not test: a voice reading a language or an accent it was not built for. So:

1. Choose a native voice of the variety. It is the fix I see work most often.
2. Write for the ear: short sentences, abbreviations written out, pauses where a listener needs time.
3. Test names and numbers (step 5).
4. Choose a voice whose delivery fits rather than pushing one with settings.
5. Judge a full paragraph, not one sentence.

If natural sound is all you need, other engines lead. In the Artificial Analysis Speech Arena, where listeners pick the more natural of two clips of the same text, ElevenLabs' Eleven v4 ranked first on 1 October 2026. Kukarella does not offer ElevenLabs voices. But the arena does not say which languages its clips use, so it ranks natural sound, not fit for your learners.

Great words, flat delivery, lost audience.

How do I choose the right AI voice style for my content?

Short answer: choose a voice whose own delivery already fits the job: calm and even for instructions, warmer for a story or a welcome, two clearly different voices for a role-play. But the controls differ from voice to voice: check what yours accepts in its panel in Text to Voice before you count on it.

Many voices also have a Style menu. A style is a request. Some voices follow it, many barely change. Write the feeling into the words, then listen. A style cannot change a voice's age.

Other providers document a similar limit. Microsoft's own documentation says that if a style is missing or invalid, the style instruction is ignored and the service "uses the default neutral speech". So check that a voice lists the style before you count on it, and listen to the result.

An OpenAI voice's panel in Text to Voice: a Style list, no speed or pitch sliders.

A human speaker can say one line in many tones. With an AI voice, choose the tone when you choose the voice: listen to several voices on the same line before you change any setting.

A person can play all ten tones; with an AI voice, choose the voice that has the tone.

A delay and a good-news announcement at the same gate are not read alike.

The tone has to fit the news.

For health, money and safety, trust matters more than friendliness. In a 2016 study of written website copy, trustworthiness explained far more of how much people wanted a service (Nielsen Norman Group). That was text, not voices, but the lesson carries.

When a voice sounds wrong, start with the voice, not the settings:

Need advice on choosing AI dubbing software?

Short answer: choose by the country your viewers live in, not only by the language, and prefer a native voice of that country for each version. But a voice that reads many languages may not read your viewers' variety: test it in each language you need.

Before you compare tools, write down the variety each audience will meet: Portuguese (Brazil) or (Portugal), Spanish (Mexico) or (Spain), French (Canada) or (France). Then check each tool for native voices in it. Most providers build the accent into the voice. Microsoft lists 14 English locales and 22 Spanish ones. Amazon shows its US voice Joanna reading French "in American-accented French". Accent behaviour can also change in an update. ElevenLabs says that in v4 a voice cloned from a Korean speaker now speaks fluent English rather than Korean-accented English. It advises you to "test this behavior directly with Eleven v4 before migrating" (ElevenLabs).

In Kukarella the rule is simple. One language per recording: a native voice for each version. One speaker who switches languages within a recording: a voice that reads many languages, tested on the switch. Either way, have a native speaker read the translated script first.

When a human voice, your own voice or another tool is the better choice

Short answer: use an AI voice for clear, repeatable material that changes often. But record yourself, hire an actor or record native speakers in the four cases below.

When it has to be you. A welcome to a new class, feedback on one person's work, a message from the person who leads the team. Even a clone of your voice will not do: our own home page FAQ warns that learners who know your voice hear the difference. If the point is that it comes from you, record it. If the point is the language, use voices that are not you.

When the relationship is the point, your own voice carries it.

The reverse can also be true. One teacher told us in a survey that they correct mistakes in a computer voice precisely because it is not theirs, so the correction feels less personal.

When the performance is the product. A drama for an audience, or a story with several child characters. A professional recording still beat synthetic voices in 2024, and you can direct an actor line by line. If you need only the most natural-sounding narration, also listen to ElevenLabs: its Eleven v4 ranked first in Artificial Analysis's blind listening tests on 1 October 2026.

With an actor you direct each line; with an AI voice you choose the voice.

When your variety is thin, or you need real-life speech. Six different Argentine speakers, Egyptian Arabic beyond two voices, or a street scene with noise and people talking over each other. Record native speakers. Use AI voices for what they do best: clear, repeatable, quick to change.

When every number must be exact and you cannot listen to all of it. Dosages, phone numbers, emergency steps. How a voice reads numbers varies voice by voice, even within one provider. Listen to every one, or record those lines.

Check yourself. Three quick questions on what you have read:

Questions people ask

Can an OpenAI voice speak with a British accent?

Do not count on it. The nine OpenAI voices in Kukarella have no accent setting. For British English, pick a native English (UK) voice, or a Microsoft " M." voice under English (UK).

What is the best OpenAI voice?

OpenAI recommends marin or cedar "for best quality"; Kukarella does not offer them. Among our nine, the best is the one that reads your two hardest lines the way you want.

Why did my Spanish sound American?

A voice that reads many languages was reading it. Change the speaker to a native Spanish voice and pick the country you want. Our home page FAQ says the same.

Can the assistant pick the accent for me?

No. It casts from the same fixed set of 22 voices whatever the script's language. Give each speaker a native voice afterwards.

How do I choose an AI voice provider?

By your listeners, not by the size of the library: native voices in their varieties, your numbers and names read correctly, the controls you need, and a voice that will stay. Then generate the same paragraph of your own text on two or three providers. Play the files to people without saying which is which, and judge the whole paragraph.

What speed should narration be?

There is no single right number. In a 2020 Mozilla study, listeners heard one English article read by many voices and found a "just right speed" at about 163 to 177 words per minute. In 6.9 million edX viewing sessions from 2012, median watching time usually rose, up to two times, with the instructor's speaking rate. The authors read this as enthusiasm, which they did not measure (Guo, Kim and Rubin, 2014). For beginners in a foreign language, start slower and listen.

What are the best settings for an AI voice?

Usually none. Choose the voice that reads your text right as it is, add pauses where learners need time, and use speed or "Say as" only where the voice offers them.

Does Google's text-to-speech have emotional styles?

Two of Google's classic voices, en-US-Neural2-F and en-US-Neural2-J, take named styles, with five values in preview: apologetic, calm, empathetic, firm and lively (Google Cloud). Google's newer Gemini-TTS voices take no named styles. You describe the tone in a written prompt instead (Google Cloud).

Is Kukarella free to try?

Yes, with limits. A guest can try 10,000 characters and a free account 30,000. Every catalogue voice uses 1 credit per character. A padlock stops you using a voice, not hearing it.

Which languages does Kukarella support?

2,000+ voices in 152 languages and accents, with English in 14 accents and Spanish in 23. But some countries have dozens of native voices and some have two: see the table above.

How we checked these numbers

Voice counts were made on 1 October 2026 with the rule the Voice Library uses to list voices under an entry. You can repeat them: open the Voice Library in Text to Voice, pick the entry and count. "Native" means a voice that speaks one language in the accent of its entry. The genders and notes of the OpenAI voices are our team's, from listening, not OpenAI's. Missing samples were checked the same day: every sample of Alloy and spot checks of Nova and Sage. Outside sources were read on 1 October 2026, several in abstract only. A study older than 2020 carries its year, so you can judge how old its voices were.

A correction: an earlier version of this page quoted unnamed "users", "creators" and "a teacher" whose words we could not trace, described a selection method we could not document, and promised that a style setting would fix a flat voice. We removed all three.

About the author. Nazim Ragimov has been building Kukarella since 2019, first for himself and his team, then for the users whose feedback shaped almost everything in it. Most of them teach, and he wants Kukarella to delight them and their students: a lesson that makes the class say "Now we get it." He started these articles himself, with help from his wife, Anna; today he sometimes uses AI for research and first drafts.

Sources

All pages read on 1 October 2026 unless stated.

Studies

• Mayer, Sobko, Mautone. Social cues in multimedia learning: Role of speaker's voice. Journal of Educational Psychology 95(2), 2003, 419–425. DOI
• Dinçer. The voice effect in multimedia instruction revisited: Does it still exist? Journal of Pedagogical Research 6(3), 2022, 17–26 (source of the effect sizes for Mayer et al. 2003). PDF
• Craig, Schroeder. Reconsidering the voice effect when learning from a virtual human. Computers & Education 114, 2017, 193–205. Abstract
• Chiou, Schroeder, Craig. How we trust, perceive, and learn from virtual humans: The influence of voice quality. Computers & Education 146, 2020. Abstract
• Siegle, Craig. The voice quality of pedagogical agent impacts learning and agent perceptions. Journal of Computer Assisted Learning 40(5), 2024, 2278–2291. DOI
• Ockey, French. From one to multiple accents on a test of L2 listening comprehension. Applied Linguistics 37(5), 2016, 693–715. Abstract
• Bradlow, Bent. Perceptual adaptation to non-native speech. Cognition 106(2), 2008, 707–729. PubMed
• Bione, Cardoso. Synthetic voices in the foreign language context. Language Learning & Technology 24(1), 2020, 169–186. Article
• Kühne, Fischer, Zhou. The Human Takes It All: Humanlike synthesized voices are perceived as less eerie and more likable. Frontiers in Neurorobotics 14, 16 December 2020. Frontiers
• Clark, Silen, Kenter, Leith. Evaluating long-form text-to-speech: Comparing the ratings of sentences and paragraphs. ISCA Speech Synthesis Workshop, 2019. ISCA Archive
• Cambre, Colnago, Maddock, Tsai, Kaye. Choice of Voices, CHI 2020, as reported by Mozilla, 7 May 2020. Mozilla
• Guo, Kim, Rubin. How video production affects student engagement: An empirical study of MOOC videos. ACM Learning@Scale, 2014, 41–50. PDF
• Towler, Dipboye. Effects of trainer expressiveness, organization, and trainee goal orientation on training outcomes. Journal of Applied Psychology 86(4), 2001, 664–673. DOI
• Towler. Effects of trainer expressiveness, seductive details, and trainee goal orientation on training outcomes. Human Resource Development Quarterly 20(1), 2009, 65–84. ERIC
• Moè, Frenzel, Au, Taxer. Displayed enthusiasm attracts attention and improves recall. British Journal of Educational Psychology 91(3), 2021, 911–927. Europe PMC
• Liew, Tan, Pang, Khan, Kew. I am Alexa, your virtual tutor! Education and Information Technologies 28, 2023, 1455–1489. Europe PMC
• Nass and colleagues. Improving automotive safety by pairing driver emotion and car voice emotion. CHI 2005 Extended Abstracts. DOI
• Lee, Nass, Brave. Can computer-generated speech have gender? CHI 2000 Extended Abstracts. DOI
• Edmett, Ichaporia, Crompton, Crichton. Artificial intelligence and English language teaching, 2nd edition. British Council, July 2024, pp. 32 and 38. PDF
• Moran. The Impact of Tone of Voice on Users' Brand Perception. Nielsen Norman Group, 7 August 2016. NN/g

Exams

• IELTS, Academic Listening format. ielts.org
• ETS, TOEFL iBT Listening. ets.org
• Instituto Cervantes, Qué es el SIELE. cervantes.es; Cómo son los exámenes DELE. cervantes.es

Stories and surveys

• OpenAI, How the voices for ChatGPT were chosen, 19 May 2024, updated 22 May 2024. OpenAI; archived copy
• OpenAI, openai.fm demo, src/lib/library.ts, 15 April 2025. GitHub
• OpenAI Developer Community, How to get a soft and cute female voice with OpenAI TTS?, posts of 25 and 29 October 2025. Forum
• Consumer Reports, Kaveh Waddell, Hey Siri, Is That You? Apple's New Voices Resonate With Some Black iPhone Users, updated 23 April 2021. Consumer Reports
• Duolingo Blog, Character voices, 23 August 2021. Duolingo
• NPR, Talk of the Nation, Studying Computers To Learn About Ourselves, 3 September 2010. NPR
• PBS, Do You Speak American?, Clifford Nass, "Machine Voices", about 2005. PBS
• Bitkom, navigation voices survey, 19 September 2011 (press release reprint). verbaende.com
• Bitkom, voice assistants and female voices, 2 September 2022. Bitkom; 1 August 2023. Bitkom

Providers

• OpenAI, Text to speech guide. OpenAI
• Google Cloud, SSML reference, last updated 30 September 2026. Google; Gemini-TTS, last updated 30 September 2026. Google
• Microsoft, Language and voice support for the Speech service, 9 September 2026. Microsoft Learn
• Amazon Polly, Using the lang tag. AWS; Generative voices. AWS
• ElevenLabs, Eleven v4 overview. ElevenLabs
• Artificial Analysis, Speech Arena leaderboard. Artificial Analysis

Kukarella

• Home page and FAQ · Voice Library in Text to Voice · DialoguesAI · Pricing