Voice API With Region-Specific Accents, Compared

Quick answer: Of the seven speech APIs we checked on 18 October 2026, only Azure Speech and Amazon (Polly plus Transcribe) pair a locale-tagged voice with a recognition locale code for all 12 regional variants, from en-AU and en-IN to es-MX, pt-PT, fr-CA and Cantonese. Google Cloud covers 11, Deepgram 4; Cartesia, ElevenLabs and OpenAI none, because their recognition is listed by base language and accent follows the voice or a prompt.

Where Zian fits when the accent has to be right on a live call

If you need a phone agent that sounds local to Sydney, Manchester, Monterrey or Montréal, choose the agent platform first and the speech API second: Zian is the platform pick, and LeadsNow.ai is the done-for-you pick. The speech API decides which accents are possible; the platform decides whether a caller ever hears the right one, because it is the layer that answers the call, books the appointment and follows up.

Zian runs inbound and outbound conversations as an AI receptionist and appointment setter across voice calls, SMS, email, WhatsApp, Instagram DMs and Facebook Messenger, in 30+ languages, with voice cloning supported. Its own Features + Integrations page on zian.ai describes English as “multi-accent” and says voice, accent, tone and languages are configurable for each agent profile. Zian has not published a per-locale list like the matrix below, so test your exact variant during the partnership-application beta. If you would rather not build or tune anything, LeadsNow.ai is a managed, pay-per-result appointment-setting service that runs on Zian, with every lead answered in under 60 seconds, 24/7.

What counts as a regional accent in a voice API

“Supports Spanish” can mean a Madrid voice, a Mexico City voice, a voice from neither, or a recogniser trained on a mix. A request for a “voice API with region-specific accents” is really two questions. Text-to-speech (TTS): does the vendor ship a voice that sounds like that region? Speech-to-text (STT): does the vendor ship a recognition model or locale code tuned to that region? A phone agent needs both, for the same variant. We call that a full pair.

The definitions below were fixed before any cell was filled and applied the same way to every row, including Zian and LeadsNow.ai.

• The 12 variants. English: en-US, en-GB, en-AU, en-IN. Spanish: es-ES, es-MX, es-US. Portuguese: pt-PT, pt-BR. French: fr-FR, fr-CA. Chinese: Cantonese (any code a vendor uses for it: zh-HK, yue-HK, yue-CN or yue). Arabic is scored separately as a count of regional codes, because vendors split it into anything from 1 to 19 codes.
• TTS “V” (voices). The vendor’s own page lists at least one vendor voice tagged with that locale or accent. The number is how many voice entries we counted.
• TTS “L” (locale setting). The vendor names the region as a language or locale setting for a voice model, but publishes no voice tagged to that locale that we could read without signing in. The accent then comes from the setting, the voice you pick, or both.
• TTS “P” (prompt). Accent is something you ask for in an instruction; no locale is named.
• STT “Yes”. The exact locale code, or a Cantonese-specific code, is listed for a recognition model. “Group” means a regional group code that covers it (es-419, Latin American Spanish, for es-MX). “Base” means only the bare language code is listed, such as es or fr.
• Full pair. V and Yes for the same variant. L, P, Group and Base do not count towards it.
• Ranking rule. Full pairs out of 12. Ties are broken by, in order: TTS variants marked V or L; STT variants marked Yes; Arabic regional variants with both a vendor voice and a recognition code.

Every count comes from the vendor’s own documentation, saved and counted by script on 18 October 2026. The matrix measures what each vendor documents as selectable. It does not measure accuracy: a vendor can list en-AU and still mishear Australian suburb names, which is the subject of our page on Australian accent accuracy and word error rate. Disclosure: zian.ai publishes this comparison and Zian appears in it.

The locale-variant coverage matrix: 12 regional variants, seven speech APIs

Each cell reads “TTS / STT”. A number is the count of vendor voice entries tagged with that locale. Columns run in ranking order.

Locale-variant coverage matrix, counted from each vendor’s own documentation, 18 October 2026. Cell = TTS / STT.

Variant	Azure Speech	Amazon Polly + Transcribe	Google Cloud TTS + STT	Deepgram	Cartesia	ElevenLabs	OpenAI
en-US	V 118 / Yes	V 15 / Yes	V 99 / Yes	V 57 / Yes	V 3 / Base	L / Base	P / Base
en-GB	V 20 / Yes	V 4 / Yes	V 63 / Yes	V 10 / Yes	V 2 / Base	V 1 / Base	P / Base
en-AU	V 19 / Yes	V 3 / Yes	V 49 / Yes	V 3 / Yes	L / Base	V 1 / Base	P / Base
en-IN	V 20 / Yes	V 3 / Yes	V 49 / Yes	V 3 / Yes	L / Base	L / Base	P / Base
es-ES	V 24 / Yes	V 6 / Yes	V 49 / Yes	V 6 / Base	L / Base	L / Base	P / Base
es-MX	V 23 / Yes	V 2 / Yes	L / Yes	V 6 / Group	L / Base	L / Base	P / Base
es-US	V 2 / Yes	V 4 / Yes	V 48 / Yes	— / Base	L / Base	— / Base	P / Base
pt-PT	V 5 / Yes	V 2 / Yes	V 4 / Yes	— / Yes	L / Base	L / Base	P / Base
pt-BR	V 28 / Yes	V 4 / Yes	V 43 / Yes	— / Yes	L / Base	L / Base	P / Base
fr-FR	V 23 / Yes	V 6 / Yes	V 42 / Yes	V 2 / Base	L / Base	L / Base	P / Base
fr-CA	V 6 / Yes	V 3 / Yes	V 45 / Yes	— / Yes	L / Base	L / Base	P / Base
Cantonese	V 5 / Yes	V 1 / Yes	V 34 / Yes	— / Yes	— / Yes	L / Yes	— / Yes
Full pairs (of 12)	12	12	11	4	0	0	0
TTS variants marked V or L	12	12	12	7	11	11	0
STT variants marked Yes	12	12	12	8	1	1	1

Sources per column: Azure, the Language and voice support page (speech to text and text to speech tables). Amazon, the Polly available voices table and the Transcribe supported languages table. Google, the Cloud Text-to-Speech supported voices list, the Gemini-TTS language table and the Speech-to-Text V2 supported languages table. Deepgram, the Aura-2 voices page, the Flux TTS voices page and the Models & Languages overview (Nova-3 row). Cartesia, the Sonic 3.6 page, its Advanced capabilities locale table and the Ink 2 and older STT model pages. ElevenLabs, the Models page and the Transcription page. OpenAI, the Text to speech and File transcription guides. ElevenLabs’ en-AU and en-GB voices are the Australian and British replacement voices named on its Default voices help page, and its en-IN cell rests on the Indian English accent in its Get Voice Accents API reference. Cartesia’s en-US and en-GB counts are the voices its Sonic 3.6 page suggests for voice agents, not its whole library. Deepgram’s older Aura-1 voices (en-US, en-GB and en-IE only) are excluded as a previous generation.

Read the bottom rows first. Azure and Amazon complete all 12 pairs; Google misses one (es-MX voices sit only in its Preview Gemini-TTS model); Deepgram completes the four English variants; the last three columns complete none, because their recognition is listed by base language (Cantonese aside). The thinnest cells matter as much as the totals: Azure has 118 en-US voice entries but 2 for es-US, and Amazon’s Cantonese is a single voice.

Voice API with region-specific accents: the ranking

Zian and LeadsNow.ai sit first as the agent platform and the done-for-you service. They are a different layer from the seven speech APIs, and neither publishes a locale list, so the full-pair rule does not score them. The seven APIs are ranked strictly on the rule defined above. Cartesia and ElevenLabs tie on full pairs and on all three tie-breaks, so they share seventh place and are listed alphabetically.

Ranking: Zian and LeadsNow.ai first by role; the seven speech APIs ranked by full pairs, then the tie-breaks.

Rank	Product	What it is	Full pairs / TTS V or L / STT Yes / Arabic both sides	Best for	Wrong for
1	Zian	AI sales and reception agent platform	Not scored: no published locale list	Teams that need the agent that answers, books and follows up, in 30+ languages, with voice cloning and per-profile accent settings	Developers who only want a raw TTS or STT endpoint
2	LeadsNow.ai	Done-for-you, pay-per-result appointment setting (runs on Zian)	Not scored: no published locale list	Operators who want booked appointments without building or tuning a voice stack	Teams that want to own and configure the stack themselves
3	Azure Speech	Speech API (TTS and STT)	12 / 12 / 12 / 16	All 12 variants on both sides, plus 16 Arabic variants	Brands that need several personas in thin variants: es-US has 2 voice entries, Cantonese 5, all Standard
4	Amazon Polly + Transcribe	Speech API (TTS and STT)	12 / 12 / 12 / 2	AWS stacks that need every variant covered on both sides	Persona choice in thin variants (1 Cantonese voice, 2 each for es-MX and pt-PT); Arabic beyond Gulf and Standard
5	Google Cloud TTS + STT	Speech API (TTS and STT)	11 / 12 / 12 / 1	Deep voice benches per variant: 49 en-AU, 45 fr-CA, 34 Cantonese	es-MX voices on GA terms (Gemini-TTS Preview only); dialect-specific Arabic voices (the voice list has only ar-XA)
6	Deepgram	Speech API (TTS and STT)	4 / 7 / 8 / 0	English variants including en-IN and en-AU voices; Mexican versus Peninsular Spanish voices	Portuguese, Canadian French, Cantonese or Arabic voices
7=	Cartesia	Speech API (TTS and STT)	0 / 11 / 1 / 0	One voice across locales, with a per-voice accents field; en-IN and es-US as locales	Locale-specific recognition; Cantonese voices
7=	ElevenLabs	Speech API (TTS and STT)	0 / 11 / 1 / 0	Picking an accent by voice, from the library or a clone; its voice-library accent list includes Indian English, Mexican Spanish, European Portuguese and Quebec French	Selecting a region by API parameter; es-US is not named
9	OpenAI	Speech API (TTS and STT)	0 / 0 / 1 / 0	Prompt-steered accents in English-first applications	Any requirement for a guaranteed regional accent

The ranking is a statement about documented regional coverage, not overall quality. On a page about language counts or turn detection, the order changes; our comparison of voice agent APIs for multilingual support counts languages across four layers of the stack and reaches a different shortlist for that job.

A listed accent is not always the accent you get

The most useful sentence on each vendor’s page is the one describing what happens when the voice and the locale disagree. Four vendors describe, in their own words, how the accent you hear can differ from the one you asked for.

• Cartesia. Sonic 3.6 accepts locale codes such as en-GB, es-MX and fr-CA, and each voice carries an accents field. Cartesia’s Advanced capabilities page is explicit: “If the voice does not support that regional accent, it uses its default English accent.” Its Multilingual Voices guide gives the example of a voice whose only English accent is General American, so “a request for en-GB falls back to her American English rather than failing”. Check the voice’s accents list, not just the locale table.
• ElevenLabs. The language_code parameter accepts ISO 639-1 codes, which carry no region. ElevenLabs’ help centre says “The accent that is used for your generation comes from the voice you’re using”, and warns that a voice not native to the language “might retain its native accent, or drift between different accents.” Its Multilingual v2 model description names regions (English for the USA, UK, Australia and Canada; French for France and Canada; Portuguese for Brazil and Portugal; Spanish for Spain and Mexico; Arabic for Saudi Arabia and the UAE), so an “Australian” claim there describes the model’s training coverage. The accent itself is chosen by picking a library voice with the accent label. Another help page says all pre-made and generated voices are English, and that community-shared Professional Voice Clones generally carry a language tag.
• Azure. Multilingual voices “can speak in the auto-detected language of the input text in the default locale”, and Microsoft adds that you can “use the <lang xml:lang> element to set the preferred speaking accent of each language, such as a British accent (en-GB) for English.” Without that element, a multilingual voice speaks in its default locale. Our matrix counts voices by their primary locale, which Microsoft says is the first segment of the voice name.
• OpenAI. OpenAI lists accent among the things you can prompt gpt-4o-mini-tts to control, alongside tone and speed, and says its voices “are currently optimized for English.” Note too that the same OpenAI page says “11 built-in voices” in one paragraph and “13 built‑in voices” in another.

The practical rule: a regional accent is a property of a voice, so verify it per voice in a recorded test call.

Spanish, Portuguese, French and Cantonese: where the lists split

English variants are the easy part; every vendor here except OpenAI names en-GB on the voice side. The splits appear in the second-largest market variant of each language.

• Mexican versus Spain Spanish. Azure lists 23 es-MX and 24 es-ES voice entries; Deepgram’s Aura-2 lists 6 voices marked Mexican and 6 marked Peninsular, plus 2 Latin American, 2 Colombian and 1 Argentine. Amazon Polly has 2 es-MX voices and 6 es-ES. Amazon Transcribe lists es-MX, but its footnote says streaming for that language is not available in five AWS Regions, including Tokyo and Cape Town; it is the only one of our 12 variants with that footnote. Google’s main voice list has es-ES and es-US but no es-MX; es-MX appears only in the Gemini-TTS language table, at Preview. Deepgram’s recogniser lists es and es-419 (Latin America) but not es-ES or es-MX, so its Spanish pairs are incomplete under our rule.
• US Spanish. Azure lists 2 es-US voice entries and Amazon 4, against Google’s 48. Cartesia accepts es-US as a locale; ElevenLabs and Deepgram do not name it.
• European Portuguese. The thinnest major variant in the set. Azure lists 5 pt-PT voice entries against 28 for pt-BR; Google 4 against 43; Amazon 2 against 4. Deepgram recognises pt-PT and pt-BR but has no Portuguese voice.
• Canadian French. Google is the outlier here, with 45 fr-CA voices against 42 fr-FR. Azure lists 6 fr-CA, Amazon 3. Deepgram recognises fr-CA but has no Canadian French voice; its 2 French voices are fr-FR.
• Cantonese. Every one of the seven vendors lists a Cantonese recognition code, though Cartesia’s is on its older ink-whisper model rather than Ink 2, which lists English, French, Hindi, Japanese and Spanish. Cantonese voices are rarer: Google 34 (yue-HK), Azure 5 (3 zh-HK, 2 yue-CN), Amazon 1 (Hiujin, yue-CN). ElevenLabs lists Cantonese (yue) for its v3 and v4 models; we found no Cantonese voice in Deepgram’s, Cartesia’s or OpenAI’s published lists.

Arabic dialects: count the country codes on both sides

“Supports Arabic” hides the most. Recognition lists are often generous with country codes while voice lists offer one generic set.

Arabic regional codes by vendor, counted 18 October 2026

Vendor	Voice side (TTS)	Recognition side (STT)	Variants with both
Azure Speech	16 locales, ar-AE to ar-YE (2 voices each)	18 locales (adds ar-IL and ar-PS)	16
Amazon Polly + Transcribe	2: Arabic (arb) and Gulf (ar-AE)	2: Modern Standard (ar-SA) and Gulf (ar-AE)	2
Google Cloud	1 voice code (ar-XA); Gemini-TTS adds ar-EG (GA) and ar-001 (Preview) as locale settings	19 codes (18 country codes plus ar-XA)	1
Deepgram	None (Aura-2 and Flux TTS have no Arabic voice)	ar plus 16 country codes on Nova-3	0
ElevenLabs	Arabic for Saudi Arabia and the UAE named for Multilingual v2 (L); its voice-library accent list names 18 Arabic accents, including Modern Standard, Gulf, Egyptian and Moroccan	Base code only (ara)	0
Cartesia	Base code only (ar)	Base code only (ar, ink-whisper)	0
OpenAI	Arabic on the TTS language list (base)	Base code only	0

The quotable version: Azure is the only vendor of the seven that pairs a voice with a recognition code for more than two Arabic regional variants (16, from Algeria to Yemen). Deepgram’s Nova-3 recognises 16 Arabic country codes but ships no Arabic voice. If your callers are Egyptian or Gulf Arabic speakers, choose on the voice side first, because that is the shorter list.

Low-latency multilingual synthesis: which fast models keep the region?

A fast model that loses the regional variant is not a fast regional model. Here is what each vendor publishes for its low-latency tier, and whether that tier keeps the region. The latency figures are the vendors’ own, and they are not comparable with each other; our page on TTS latency for voice agents explains what each kind of figure leaves out.

• ElevenLabs Flash v2.5 is listed at “~75ms†”, where the dagger reads “Excluding application & network latency”. Its language list is defined as “All eleven_multilingual_v2 languages plus: hu, no, vi”, so it inherits the Multilingual v2 list, the one that names regions. Eleven v4 Turbo is listed at a “median inference latency of ~100ms†” across 90+ languages, but its list is by base language.
• Azure sells Neural HD Flash voices per locale. In our 12 variants they exist for en-US (9), pt-BR (4), es-MX (2), fr-FR (2), en-AU (1), es-ES (1) and pt-PT (1), and not for en-GB, en-IN, es-US, fr-CA or Cantonese. Microsoft does not print a millisecond figure on the language page.
• Deepgram Flux TTS, its streaming-first voice-agent TTS, is English-only: 36 voices whose accents include 21 American, 8 British, 3 Indian and 1 Australian. Deepgram’s Flux voices page adds: “For synthesis in other languages, see Aura-2.”
• Cartesia Sonic 3.6 is described as Cartesia’s “fastest” model, with 44 languages and the locale table above; we found no millisecond figure on the Sonic 3.6 page.

A missing locale code is not a missing accent

The recognition half of the matrix needs one caution, and it cuts against our own scoring rule. Some vendors deliberately ship one model per language that handles every accent, instead of a model per locale. Deepgram lists its Flux English model as “English (all accents): en”. OpenAI’s gpt-transcribe takes ISO 639-1 codes, selected ISO 639-3 codes such as yue, and regional zh locale codes (zh-cn, zh-tw, zh-hk); the formats it lists name no other regional codes. A “Base” cell therefore means “not separately selectable”, not “cannot understand Australians”.

What a locale code gives you is control: locale-specific vocabulary, number and date conventions, and in Azure’s case per-locale custom speech options, including phrase lists for many locales. What it does not give you is a measured error rate. For that, record 50 to 100 real calls from your callers, transcribe them on each shortlisted model, and score names, suburbs and digit strings separately from overall word error rate. Our comparison of speech-to-text for multiple languages covers the separate problem of callers who switch language mid-sentence.

How to choose a regional voice API in five checks

1. Write down your variants, not your languages. “Spanish” is not a requirement; “Mexican Spanish callers in Texas” is (es-MX voice, es-US recognition).
2. Find the voice-side list first. It is almost always shorter than the recognition list, as the Arabic and Portuguese rows show.
3. Count voices per variant, not just presence. One or two voices means one or two personas for your brand, with no fallback if a voice is retired.
4. Confirm the fallback behaviour. Ask what happens when the locale and the voice disagree. Cartesia and ElevenLabs document a silent fallback, and Azure’s multilingual voices fall back to their default locale.
5. Test with your own callers. A catalogue is a promise; recordings are evidence.

How we built this comparison

We chose the 12 variants as the regional pairs buyers most often ask about for English, Spanish, Portuguese, French and Chinese, plus Arabic as a separate count. We read each vendor’s own language and voice lists on 18 October 2026, saved each page and counted cells by script rather than by eye. Where a hub page named a sub-product, we opened it: Deepgram’s Flux TTS voices page added three Indian-accented and one Australian-accented voice, and Google’s Gemini-TTS page added es-MX. We could not count ElevenLabs’ or Cartesia’s accent-tagged library voices, because both libraries need an account for filtered browsing. Those cells are marked L, not V, unless the vendor names a voice on an open page: ElevenLabs’ Default voices page names an Australian and a British replacement voice, and Cartesia’s Sonic 3.6 page suggests three en-US and two en-GB voices, so those cells count only the named voices. ElevenLabs’ public Get Voice Accents endpoint lists the library’s accent labels, including Indian English, which is why its en-IN cell is L. Vendor lists change often; check the linked page before you build.

Frequently asked questions

Where every figure on this page comes from

Figure	Who published it	Link	Date read
Azure TTS voice entries per locale (en-US 118, en-GB 20, en-AU 19, en-IN 20, es-ES 24, es-MX 23, es-US 2, pt-PT 5, pt-BR 28, fr-FR 23, fr-CA 6, Cantonese 5); Neural HD Flash counts; 16 Arabic TTS locales	Microsoft	learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=tts	18 October 2026
Azure STT locales for all 12 variants; 18 Arabic STT locales; multilingual-voice default locale and <lang xml:lang> accent sentence	Microsoft	learn.microsoft.com/en-us/azure/ai-services/speech-service/language-support?tabs=stt	18 October 2026
Amazon Polly voices per locale (en-US 15, en-GB 4, en-AU 3, en-IN 3, es-ES 6, es-MX 2, es-US 4, pt-PT 2, pt-BR 4, fr-FR 6, fr-CA 3, Cantonese 1); Arabic arb and ar-AE	Amazon Web Services	docs.aws.amazon.com/polly/latest/dg/available-voices.html	18 October 2026
Amazon Transcribe locale codes for all 12 variants; Arabic ar-SA and ar-AE; es-MX streaming Region footnote	Amazon Web Services	docs.aws.amazon.com/transcribe/latest/dg/supported-languages.html	18 October 2026
Google TTS voices per locale (en-US 99, en-GB 63, en-AU 49, en-IN 49, es-ES 49, es-US 48, pt-PT 4, pt-BR 43, fr-FR 42, fr-CA 45, yue-HK 34); no es-MX; Arabic ar-XA	Google Cloud	docs.cloud.google.com/text-to-speech/docs/list-voices-and-types	18 October 2026
Gemini-TTS language table (es-MX, pt-PT, ar-001 Preview; ar-EG GA; Pre-GA terms)	Google Cloud	docs.cloud.google.com/text-to-speech/docs/gemini-tts	18 October 2026
Google STT V2 locale codes for all 12 variants; 19 Arabic codes	Google Cloud	docs.cloud.google.com/speech-to-text/docs/speech-to-text-supported-languages	18 October 2026
Deepgram Aura-2 voices by locale (en-US 36, en-GB 2, en-AU 2, es-MX 6, es-ES 6, es-419 2, es-CO 2, es-AR 1, fr-FR 2)	Deepgram	developers.deepgram.com/docs/tts-models	18 October 2026
Deepgram Flux TTS: 36 English voices (21 American, 8 British, 3 Indian, 1 Australian, 1 each Irish, Singaporean, Filipino)	Deepgram	developers.deepgram.com/docs/flux-tts/voices	18 October 2026
Deepgram Nova-3 recognition codes (en-US, en-AU, en-GB, en-IN, es, es-419, pt-BR, pt-PT, fr, fr-CA, zh-HK; ar plus 16 country codes); Flux “English (all accents)”	Deepgram	developers.deepgram.com/docs/models-languages-overview	18 October 2026
Cartesia Sonic 3.6: 44 languages, “fastest”; named en-US and en-GB voices	Cartesia	docs.cartesia.ai/build-with-cartesia/tts-models/latest	18 October 2026
Cartesia regional locale codes and default-accent fallback sentence	Cartesia	docs.cartesia.ai/build-with-cartesia/capability-guides/advanced-capabilities	18 October 2026
Cartesia per-voice accents field and en-GB fallback example	Cartesia	docs.cartesia.ai/build-with-cartesia/capability-guides/multilingual-voices	18 October 2026
Cartesia Ink 2 languages (en, fr, hi, ja, es)	Cartesia	docs.cartesia.ai/build-with-cartesia/stt/latest	18 October 2026
Cartesia ink-whisper language list including yue	Cartesia	docs.cartesia.ai/build-with-cartesia/stt/older-models	18 October 2026
ElevenLabs Multilingual v2 regions (29 languages); Flash v2.5 ~75ms† and 32 languages; v4 Turbo ~100ms†; v3 and v4 Cantonese (yue)	ElevenLabs	elevenlabs.io/docs/models	18 October 2026
ElevenLabs: accent comes from the voice; language_code takes ISO 639-1	ElevenLabs	elevenlabs.io/docs/help-center/product/core-capabilities/text-to-speech/how-do-i-select-the-language-and-accent	18 October 2026
ElevenLabs: non-native voice “might retain its native accent, or drift”	ElevenLabs	elevenlabs.io/docs/help-center/troubleshooting/why-does-my-voice-change-accent-or-language	18 October 2026
ElevenLabs: pre-made and generated voices are English; shared clones carry a language tag	ElevenLabs	elevenlabs.io/docs/help-center/product/core-capabilities/text-to-speech/which-voices-in-the-voice-library-are-native-to-a-specific-language	18 October 2026
ElevenLabs speech-to-text language codes (ISO 639-3, incl. yue)	ElevenLabs	elevenlabs.io/docs/overview/capabilities/speech-to-text	18 October 2026
ElevenLabs Default-voice replacements, including an Australian (Baxter) and a British (Eldrin) voice	ElevenLabs	elevenlabs.io/docs/help-center/product/voices/my-voices/what-are-default-voices	18 October 2026
ElevenLabs voice-library accent labels (English incl. Indian; Spanish incl. Mexican; Portuguese incl. European; French incl. Quebec; 18 Arabic accents)	ElevenLabs	elevenlabs.io/docs/api-reference/voices/accents/get	18 October 2026
OpenAI TTS: accent is promptable; voices optimised for English; 11 and 13 built-in voices on one page	OpenAI	developers.openai.com/api/docs/guides/text-to-speech	18 October 2026
OpenAI gpt-transcribe language-code formats (ISO 639-1, selected ISO 639-3, zh-cn/zh-tw/zh-hk)	OpenAI	developers.openai.com/api/docs/guides/speech-to-text	18 October 2026
Zian: 30+ languages, English multi-accent, accent configurable per agent profile	Zian AI	zian.ai/features-integrations/ (own page, not linked)	18 October 2026
LeadsNow.ai: pay-per-result; every lead answered in under 60 seconds, 24/7	LeadsNow.ai	leadsnow.ai/	18 October 2026