1,000+ new voices, our most natural set yet
Developers now have 1,000+ new voices that are more expressive, built for more use cases, and fluent in more accents. They cover English, French, Spanish, German and Portuguese, extend beyond customer service into narration, ads, social media and characters, and speak in 29 accents, from Scottish and Indian English to Quebecois French, Rioplatense Spanish and Swiss German.
This post explains how the new voices improve in naturalness and expressivity, which new use cases and accents are now covered, and how the voices were chosen.
1,000+
New voices
Added in this release
1.4k
In the catalog
Up from ~380
29
Accents
Covered by the new voices
4
Use cases
Customer service, narration, ads, characters
More natural voices for customer service and voice agents
Customer service and voice agents were the catalog's first use case, and this release raises the bar there first. Native listeners compared the new customer service voices with the ones we already recommended, reading the same support lines, and chose which they preferred. New voices came out ahead.
| Voice | Score vs median | Before | After |
|---|---|---|---|
| Tilly→Elsie Standard British · en-gb | −27 1949+118 2094 | Tilly | Elsie |
| Declan→Eoin Irish · en-ie | +5 1922+151 2068 | Declan | Eoin |
| Marcos→Nuno Castilian Spanish · es-es | −101 1945+76 2122 | Marcos | Nuno |
| Ximena→Itzel Mexican Spanish · es-mx | +125 2050+256 2181 | Ximena | Itzel |
| Maude→Rosalie Quebecois French · fr-ca | −12 1955+126 2093 | Maude | Rosalie |
| Augustin→Armand Parisian French · fr-fr | ±0 1993+121 2114 | Augustin | Armand |
| Annika→Nele Standard German · de-de | +25 2031+139 2145 | Annika | Nele |
| Ricardo→Duarte European Portuguese · pt-pt | +73 1868+336 2131 | Ricardo | Duarte |
The scores are ELO ratings from those comparisons: a voice climbs when listeners pick it over another, so they rank preference and nothing else. Bars are 95% confidence intervals, and the table shows the widest gap in each accent. The previous voices stay in the catalog, so integrations that use them keep working.
How the new voices were chosen
In August we described how we pick the voice that wins: generate many candidates against a written target, then let listeners decide. The method has been upgraded since, so it scales to thousands of voices. Automated judges score every clip for audio quality, pacing, expressiveness and accent; native listeners then compare the finalists head to head against our existing voices.
- 16,000+voices designed with the voice design model
- ~3,000shortlisted by automated judges
- Human listeningagainst existing catalog voices
- 1,000+added to the catalog
Within each language, accent, gender and use case, up to five voices are flagship: the ones we recommend you try first, decided by those same listener comparisons. Flagship is a shortcut to a good first choice. The other voices in a category passed the same review, and a different brand, script or audience often calls for one of them.
Every voice in this release was designed from a written prompt with the voice design model, the same model you can use in the studio. If the catalog does not have the voice you need, describe it and design your own.
New use cases and accents in the voice catalog
The new voices are cast for four use cases, with a clear step up in naturalness and expressivity: 257 for customer service and voice agents, 221 for narration, 228 for ads and social media, and 329 character voices. Narration and ads now have flagship voices of their own, so a content team gets the same recommended starting point a voice agent team already had.
A use case changes what a voice has to do, and listeners judge each one on different things. A support voice has to stay clear and patient across short turns on a phone line. A narrator needs steady breath, clauses grouped by sense, and sentence endings that land. An ad read needs forward energy and crisp consonants that cut through a music bed. As we wrote in August, use case moves listener preference more than any other axis, so every new voice is cast for one from the start.
Every language grows, and Spanish grows the most.
29 accents across five languages
Teams building voice agents for a specific market ask for the accent their callers speak, and content teams localizing a campaign ask for the same.
Standard USen-us
Standard Britishen-gb
Australianen-au
Canadianen-ca
Scottishen-gb-x-scottish
Irishen-ie
Indian Englishen-in
New Zealanden-nz- Southern Americanen-us-x-southern
Californianen-us-x-western
Castilian Spanishes-es
Mexican Spanishes-mx
Rioplatense Spanishes-ar
Chilean Spanishes-cl
Colombian Spanishes-co
Dominican Spanishes-do
Puerto Rican Spanishes-pr
Parisian Frenchfr-fr
Quebecois Frenchfr-ca- Southern Frenchfr-fr-x-southern
Moroccan Frenchfr-ma
Senegalese Frenchfr-sn
Standard Germande-de
Austrian Germande-at
Swiss Germande-ch
Bavariande-de-x-bavarian
Brazilian Portuguesept-br
European Portuguesept-pt- Brazilian Nordestinopt-br-x-nordestino
Listen to the new voices by language
Pick an accent to hear two of its voices side by side: a female and a male voice, or two different use cases. Each card plays the voice's own preview line, so no two cards say the same sentence. Lines in French, Spanish, German and Portuguese carry an English gloss underneath. The ID on each card is the voice_id you pass to the API, and Try this voice opens that voice in the studio catalog.
English voices: Standard US, Southern US, British, Scottish, Irish, Indian, Australian and New Zealand
Nadia
Low and resonant with real chest weight under every line, holding even dynamics and a settled fall at each sentence end.
“Judging by the worn thresholds, thousands of children passed through these doors.”
Warren
Low, chest-deep resonance with controlled breath and sense-grouped phrasing that holds structure across long paragraphs.
“Glaciers retreat slowly, leaving behind ridges of gravel that mark each year of loss.”
French voices: Parisian, Quebecois and Southern French
Ombeline
Low and chest-resonant, with even dynamics and a definite fall at the close of every sentence.
“Chaque chapitre s'ouvre sur une question simple, puis déroule patiemment ses conséquences.”
EN · Each chapter opens on a simple question, then patiently works through its consequences.
Baptiste
Lively and expressive, with pitch that lifts and dips across a phrase and a playful edge to every line.
“Votre abonnement reste actif, la coupure vient d'une simple erreur de facturation.”
EN · Your subscription is still active; the cut-off came from a simple billing error.
Spanish voices: from Castilian to Rioplatense and Caribbean Spanish
Itziar
Low and resonant, with chest weight and steady breath, it groups clauses by sense and lands the end of every sentence.
“Cuando el río bajaba de nivel, los vecinos cruzaban por las piedras sin mojarse los pies.”
EN · When the river ran low, the neighbours crossed on the stones without getting their feet wet.
Gonzalo
Low and chest-resonant, steady in dynamics, with clauses grouped by sense and a real fall at the end of each sentence.
“El viajero anotó en su cuaderno el nombre de cada arroyo, cada puente y cada campanario.”
EN · The traveller wrote down the name of every stream, every bridge and every bell tower in his notebook.
German voices: Standard, Swiss, Austrian and Bavarian
Mareike
Low and chest-resonant with even dynamics and a definite fall at the close of each sentence, authority coming from steadiness rather than volume.
“Was geschieht eigentlich mit einer Sprache, wenn ihre letzten Sprecher verstummen?”
EN · What actually happens to a language when its last speakers fall silent?
Nele
Rounded, open tone with gentle lift at the ends of phrases and an easy, flowing rhythm.
“Möglicherweise liegt es an Ihrem Browser, versuchen Sie es bitte in einem anderen Fenster.”
EN · It may be your browser; please try again in another window.
Portuguese voices: European, Brazilian and Nordestino
Leonor
Low, chesty resonance with settled breath control, real pauses at paragraph breaks and a clean fall at every full stop.
“Neste capítulo, vamos observar como as plantas transformam a luz em alimento.”
EN · In this chapter, we'll look at how plants turn light into food.
Duarte
Rounded and resonant, with open vowels and a gentle lift at the end of phrases.
“Prefere o reembolso por transferência, ou desconto o valor na próxima fatura?”
EN · Would you prefer the refund by bank transfer, or shall I take the amount off your next invoice?
Character voices for games, animation and stories
The catalog also gets character voices in all five languages: pixies, dragons, ogres, pirates and sorcerers. Each one sounds like a human actor playing the role, with the gravel, the pauses and the timing of a performance. Our next model iteration will add fictional and cartoon-like character voices. Below, each language has two of its best-rated character voices, each a different character type.
OrmagarStandard Britishen-gb
Vast chest weight and a dark, gravelled low register, with slow deliberate phrasing and long controlled pauses.
“Three kingdoms burned before your grandmother drew breath, and I remember each of them.”
MabScottishen-gb-x-scottish
Sparkling and airy, with a forward smiling placement and quick, bouncing melody that darts between tiny pauses.
“Follow the glow, quick now, before the thistles close over the path for the night.”
How to find the right voice in the catalog
In the studio, open the voice catalog and filter by language, accent, gender, age and use case. The filters combine, so "Irish, female, customer service" is one view. To link a single voice, add its ID to the catalog URL: https://studio.gradium.ai/voices/library?voice=<voice_id>. That is what the Try this voice buttons above do.
From code, pass the same ID as voice_id in a TTS request. The docs cover the request format and streaming.
What's next: regional accents within a language
Most of the 29 accents in this release are national or broad regional standards: the English of Scotland, the French of Quebec, the Spanish of Buenos Aires. The next step goes finer, to city and regional accents inside a country, where a caller hears the difference within a few words.
That work happens in the voice design model. Each accent it learns opens a new set of voices to design, and native listeners still decide which ones ship. Accent granularity is where we are putting the effort, because a team serving one region needs a voice from that region, and a national standard does not cover it.
Further reading:
- New voices: how we pick the one that wins, the selection method this release builds on.
- Launching Voice Design: prompt the voice your agent needs, on the model that designed every voice above.
- Giving voice agents a voice that adapts to context, on choosing a catalog voice per conversation.
Need an accent, a language or a brand voice we do not cover yet? Tell us what you are building.
Related posts
Giving voice agents a voice that adapts to context
Most voice agents speak with one preselected voice, whoever is on the line. Once a voice can be created from a text description, the voice becomes another property your code computes from context. This article builds the full pipeline: classify the message, map the signals to a Voice Design prompt, resolve it to an approved voice on the hot path, and design the missing ones in the background.
Launching Voice Design: prompt the voice your agent needs
Developers building voice agents ask us for voices matched to the use case in front of them: a Québécoise receptionist for a Montréal dealership, a Paulista support agent for a São Paulo fintech, a narrator in his sixties with the authority of a lecture hall. Briefs outnumber any catalog. Voice Design starts from a prompt: write one or two sentences, get complete new voices back in seconds, keep the one you want.
New voices: how we pick the one that wins
Picking a voice works like casting: one candidate in a thousand has a hit factor you cannot explain. Here is the methodology behind every flagship voice in the Gradium catalog, from generating hundreds of candidates per category to ranking them head-to-head against the incumbents on an ELO benchmark, why the same prompt does not travel across locales, and how to write the target yourself.
Frequently Asked Questions
1,422. The catalog grew from 387 voices with the addition of 1,000+ new voices across English, French, Spanish, German and Portuguese. By language, English has 539 voices, Spanish 310, French 221, German 201 and Portuguese 151.
Four: 257 voices for customer service and voice agents, 221 for narration, 228 for ads and social media, and 329 character voices. Narration and ads now have flagship voices of their own, alongside the customer service voices the catalog is known for.
29 accents. English: Standard US, Standard British, Scottish, Irish, Indian English, Australian, Southern American, Californian, Canadian and New Zealand. French: Parisian, Quebecois, Southern French, Moroccan and Senegalese. Spanish: Castilian, Mexican, Rioplatense, Chilean, Colombian, Dominican and Puerto Rican. German: Standard German, Austrian, Swiss and Bavarian. Portuguese: Brazilian, European and Brazilian Nordestino.
Within each language, accent, gender and use case, up to five voices are flagship: the ones Gradium recommends trying first, chosen head to head by native listeners. Flagship is a starting point rather than a quality line. The other voices in a category passed the same review and are often the right choice for a different brand, script or audience.
Every voice was designed from a written prompt with Gradium's voice design model. Automated judges score every clip for audio quality, pacing, expressiveness and accent; native listeners then compare the finalists head to head against existing catalog voices. Out of more than 16,000 designed voices, about 3,000 reached native listeners and 1,000+ joined the catalog. The methodology is proprietary, and humans check every voice before it ships.
Open the voice catalog in the Gradium studio and filter by language, accent, gender, age and use case, or open a single voice with https://studio.gradium.ai/voices/library?voice=<voice_id>. From code, pass the same ID as voice_id in a TTS request.