Deepgram Voice Agent Languages and Dialect Support

Quick answer: As at 20 October 2026, Deepgram’s Voice Agent API can listen in 63 Nova-3 language entries, 10 of them with Flux Multilingual turn detection, but its own agent model list names Deepgram voices only in English and Spanish. So a fully Deepgram-native agent is documented end to end in 2 languages. Aura-2 voices for German, French, Dutch, Italian and Japanese exist on the text-to-speech API but are not in that list, and every other language needs a third-party voice.

Where Zian fits when your callers need more than two languages

If you need a phone agent for callers outside English and Spanish, choose the agent platform first and the speech vendor second. Zian is the platform pick and LeadsNow.ai is the done-for-you pick. Deepgram’s own lists show how quickly a single-vendor stack runs out of native voices. The platform decides whether a caller in Hindi, Portuguese or Arabic actually gets answered, booked and followed 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. It also handles answering machines and voicemail on calls. It is in a partnership-application beta, and its home page names English (multi-accent), Mandarin, Arabic, French, German, Spanish and Japanese among its 30+ languages but does not publish a full per-language list like the matrix on this page, so test your own languages and dialects on real calls during the 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 and answers every lead in under 60 seconds, 24/7.

Which languages does Deepgram’s voice agent support? It hears, thinks and speaks separately

Deepgram’s Voice Agent API is one WebSocket that chains three providers: listen (speech-to-text), think (the LLM) and speak (text-to-speech). Each has its own language list, and the agent works in a language only when all three cover it. We read every page below on 20 October 2026 on developers.deepgram.com and deepgram.com, and counted the lists by script.

• Listen is Deepgram only. The Configure the Voice Agent reference says of agent.listen.provider.type: “Currently only Deepgram is supported”. The STT Models page names two families: Flux (version v2), with model-integrated end-of-turn detection, and Nova (version v1), such as nova-3 or nova-2. Nova uses application-level voice activity detection to decide when you have finished speaking.
• Think is open. The LLM Models page lists six provider types (open_ai, anthropic, aws_bedrock, google, groq and nvidia), with Deepgram-managed models for four of them and your own endpoint for the rest. Deepgram publishes no per-language list for this layer. Its multilingual guide steers language by prompt instead, with patterns such as “Match the language of each user message independently.”
• Speak is open, with a short native list. The default is Deepgram’s own Flux TTS (version v2) or Aura (version v1). You can also use Deepgram-managed Cartesia, or bring your own OpenAI, ElevenLabs, Cartesia or Amazon Polly account, per the Voice Agent TTS Models page.

So the honest unit is the language pair: what the agent can hear, against what it can say. The rest of this page counts both, language by language and then locale by locale.

What the Deepgram agent can hear: Flux, Flux Multilingual and Nova-3

A Deepgram agent can hear far more languages than it can speak. Three listen models matter for an agent, and the speech-to-text Models & Languages Overview gives the language list for each.

• flux-general-en: “English (all accents)”, code en. This is the turn-aware default for English agents.
• flux-general-multi (Flux Multilingual): 10 languages, namely English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. Deepgram’s Multilingual Voice Agents guide calls it “the default recommendation”. You bias it with a language_hints array, and without hints it auto-detects the language each turn.
• nova-3: 63 language entries carrying 104 language codes, plus multi for code-switching across the same 10 languages as Flux Multilingual. Deepgram’s model table describes Nova-3 as having no turn detection, so in an agent it relies on voice activity detection. Nova-2 is also accepted, but every language it lists is already on Nova-3.

The Nova-3 list is still growing. Deepgram’s changelog records Kazakh as a new Nova-3 language on 3 September 2026. Its other speech-to-text language entries since then improve accuracy for existing languages, and on 6 October 2026 Deepgram extended multi to nova-3-medical.

Two caveats from the same docs shape what “supports a dialect” means on the listening side. First, Flux treats locales loosely. The Flux Multilingual & Language Prompting page says: “Locale-level subtags (e.g., en-GB, pt-BR) are accepted. If no exact match exists, Flux treats them as the base language code.” A pt-BR hint on Flux is a Portuguese hint, not a Brazilian model. Second, spelling is American. Deepgram’s Languages Support page says English transcripts use “standardized American spelling” with en-US or en-GB alike. If an Australian or British prompt quotes the caller back, that matters.

The quotable version: a Deepgram voice agent can transcribe 63 Nova-3 language entries, but only 10 of them get Flux’s turn-aware listening, and the rest end turns on silence detection.

What the Deepgram agent can say with Deepgram’s own voices

Deepgram has three voice families. All three are usable in the agent, but they differ widely in language coverage.

• Flux TTS (version v2) is the agent default (flux-kit-en) when you omit the speak setting. The Flux TTS Voices & Languages page lists 36 voices, all English: 21 American, 8 British, 3 Indian, and 1 each for Australian, Irish, Singaporean and Filipino. Deepgram’s TTS overview says Flux TTS is English-only “until Flux TTS’s multilingual voices ship”. Deepgram’s Flux TTS product page says more languages are “on the roadmap”, with no date.
• Aura-2 (version v1) has 90 voices in seven languages in the Aura voices catalogue: English 41, Spanish 17, German 7, French 2, Dutch 9, Italian 9 and Japanese 5.
• Aura (first generation) has 12 English voices: 9 American, 2 British and 1 Irish.

Here is the finding that matters for agents. The Voice Agent API reference defines the Deepgram speak model as an enumerated list of 99 values: the 12 first-generation Aura voices, the 36 Flux TTS voices and 51 Aura-2 voices. Those 51 are all 41 English Aura-2 voices plus 10 of the 17 Spanish ones. It lists no German, French, Dutch, Italian or Japanese voice, which is 32 Aura-2 voices in total. The agent’s own TTS page describes Aura as “the broadest Deepgram voice catalog across English and Spanish”, but the same page says “For a complete list of Deepgram Aura TTS models see TTS Voice Selection” and links the seven-language catalogue. Deepgram’s text-to-speech product page answers its multilingual FAQ the same way, saying Deepgram offers “native English and Spanish code-switching voices” and that the Voice Agent API “supports broader multilingual deployments through integrated third-party text-to-speech providers”.

The evidence cuts both ways, and the page should say so. The same enumerated list omits 7 Spanish voices that the catalogue does include: three Mexican, two Peninsular, one Colombian and one Argentine. That suggests the list lags the catalogue rather than reflecting a hard limit, so “not listed” may mean “not yet documented”. The reference lags elsewhere too: its Cartesia model list holds only sonic-2 and sonic-multilingual, and its ElevenLabs list omits eleven_v3, which Deepgram’s own TTS Models page shows in an example. We did not send a German Aura-2 voice to a live agent. If you need one, send it in a Settings message and check whether you get a SettingsApplied event or an error.

One more rule shapes everything that follows: Deepgram voices carry their language in the model name, such as aura-2-sirio-es. The configure reference says they “reject a language field with UNPARSABLE_CLIENT_MESSAGE”, and the multi setting is accepted only for ElevenLabs, OpenAI and Cartesia voices.

The Deepgram voice agent language matrix

How we counted. We set these rules before filling any cell and applied them to every row.

• Rows are Deepgram’s own Nova-3 language entries, 63 in all, grouped where they share a verdict.
• Hears means the entry is on the Nova-3 list. Turn-aware means it is also one of Flux Multilingual’s 10 languages, or English for flux-general-en.
• Agent voices counts Deepgram voices for that language that appear in the Voice Agent API reference’s speak-model list. Catalogue only counts Deepgram voices in that language that exist on the TTS API but are not in the agent’s list.
• Full native agent is Yes when the agent can hear the language and lists a Deepgram voice for it. It is Unconfirmed when only catalogue voices exist, and No when Deepgram has no voice in that language at all.

Every count comes from pages saved and parsed on 20 October 2026.

Read the last row first. Two of 63 entries, English and Spanish, run as a fully Deepgram-native agent on what the agent documents today. Seven more have Deepgram voices that the agent’s API reference does not list, although its TTS page links them in the full Aura catalogue. The other 54 can be heard but not spoken by Deepgram, and 16 of those are on none of the Cartesia Sonic 3.6, ElevenLabs Turbo v2.5 or Amazon Polly lists. We did not check OpenAI’s voice list. Cartesia lists Chinese as one code, zh, which we matched to Deepgram’s two Mandarin entries, and Amazon lists Cantonese as yue-CN against Deepgram’s zh-HK. For Hindi, Russian and Portuguese, Flux Multilingual gives you turn-aware listening, but a Deepgram-only build has no voice to answer with.

Regional dialects: which locales pair end to end

Language support is not the same as dialect support. For locales we use a stricter rule. A locale pair needs Nova-3 to list that exact locale code, and the agent’s list to hold at least one Deepgram voice tagged with that accent. A voice-only or code-only locale does not count. The table counts only voices in the agent’s list. Catalogue-only voices are shown in brackets.

That gives five locale pairs: en-US, en-GB, en-AU, en-IN and es-419. Our comparison of voice APIs with region-specific accents scores Deepgram at 4 of its 12 variants. That is the same four English pairs as here. The difference is that es-419 is not one of that page’s 12 variants. That page also counts Deepgram’s whole standalone TTS catalogue, which is why its Mexican, Peninsular and French voice counts are higher than the agent-listed counts above.

The weak spot is the voice side, and the regional gaps are uneven. Deepgram recognises Brazilian and European Portuguese, Canadian French and Swiss German by code, but has no agent voice for any of them. It has Filipino, Singaporean and Irish English voices, but no matching Nova-3 locale code. For those, the general en model or flux-general-en (“all accents”) does the listening.

Need the agent to answer in a language Deepgram cannot speak? Zian runs AI receptionist and appointment-setting agents in 30+ languages, with voice cloning supported, across calls, SMS, WhatsApp, Instagram and Facebook Messenger.

Apply For Partnership

Plugging in a third-party voice (and LLM) for other languages

For any language outside English and Spanish, Deepgram’s own guidance is to swap the voice. The Multilingual Voice Agents guide has two rows: for a bilingual English and Spanish agent, use a Deepgram Aura codeswitching voice; for “any other multilingual mix”, use Cartesia, OpenAI or ElevenLabs with language: "multi". Here is what each route documents, read on the providers’ own pages on 20 October 2026.

Three details from those pages will save you a failed test call.

• Check the model in Deepgram’s Cartesia examples. They still use sonic-2, and Cartesia’s older models page says sonic-2 snapshots “will stop working after October 20, 2026”. Its current Sonic 3.6 page lists 44 languages. Copy the model ID from Cartesia, not from the example, and confirm it in a Settings message: Deepgram’s API reference still lists only sonic-2 and sonic-multilingual as Cartesia model IDs.
• ElevenLabs labels the model Deepgram names as superseded. ElevenLabs’ Models page lists eleven_turbo_v2_5 in 32 languages as a “First generation low-latency model (outclassed by Flash models)”. Deepgram’s own example also shows eleven_v3 over HTTPS, because v3 has no WebSocket streaming.
• A fallback chain can change the language. The speak setting accepts an ordered array of providers, and the agent falls back to the next one on an error or timeout. If your primary is a Spanish Aura-2 voice and your fallback is an English-only voice, a failed request is answered in the wrong voice for that caller. Put a same-language voice second.

You can also change voice during a call. The UpdateSpeak message swaps the speak model mid-conversation, and on Flux TTS “the new voice takes effect on the agent’s next turn”. The UpdateListen message does the same for the language and the language_hints. Deepgram warns that omitted language_hints are cleared, not kept. Combined with the languages field that Flux Multilingual returns each turn, this gives you a detect-then-switch pattern: hear the language, then move both the hints and the voice. Deepgram documents both messages, but the switching logic is yours to write.

The think layer is rarely the constraint. All six LLM provider types are general-purpose models, and Deepgram’s documented control is the prompt. Test that the model actually answers in the caller’s language. Our language parity test for multilingual AI agents gives a set of probes you can run per language.

How to configure a Deepgram voice agent for your language

Pick the row that matches your callers. Each row follows from the two matrices above.

The last row is the boundary of this product. Because you cannot swap the ears, a language Nova-3 does not list has no documented listen model on the Deepgram Voice Agent API, whatever voice you bring: the agent docs name Flux, nova-3 and nova-2, and every Nova-2 language is already on Nova-3. For those callers you need a stack where you choose the recogniser. Our comparison of voice agent APIs for multilingual support covers that layer by layer.

What each extra language costs you to build

For one or two languages, building on Deepgram is entirely workable, and English plus Spanish is the easiest pair it supports. Each language beyond that adds work the language tables do not show.

• Voice choice. Pick a second voice vendor, and usually a second key, contract and latency budget for it.
• Configuration. Add the hints or code-switching settings, with a same-language fallback voice.
• Content. Translate and review the greeting, the required disclosures and the prompt.
• Vocabulary. Add keyterms for local names, suburbs and products.
• Turn-taking. Retune end-of-turn thresholds, and for Nova-3 languages accept silence-based turn ending. Our page on why Deepgram Flux cuts callers off covers the Flux settings.
• Testing. Record real calls in each dialect, because a catalogue is a promise and a recording is evidence.

Multiply that by the number of languages your callers actually use, and by every change Deepgram makes to its lists, such as a new Nova-3 language, a voice added to the agent list or a third-party model sunset. That maintenance is the real cost of running a multilingual agent yourself.

Want the multilingual agent run for you? Zian answers inbound calls 24/7, books appointments and follows up in 30+ languages. Apply to the partnership-application beta and test your own languages on real calls.

Apply For Partnership

Deepgram voice agent languages: frequently asked questions

Which languages does Deepgram’s voice agent support?

It depends on the direction. For listening, the Voice Agent API uses only Deepgram speech-to-text, and Nova-3 lists 63 language entries, 10 of which also run on the turn-aware Flux Multilingual model. For speaking, the agent’s own model list names Deepgram voices only in English and Spanish, as at 20 October 2026, per the Voice Agent API reference. German, French, Dutch, Italian and Japanese have Aura-2 voices that the reference does not list, so test those first. Every other language needs a third-party voice such as Cartesia, ElevenLabs, OpenAI or Amazon Polly.

Can a Deepgram voice agent speak German, French or Japanese with Deepgram’s own voices?

Deepgram’s Aura-2 catalogue has 7 German, 2 French, 9 Dutch, 9 Italian and 5 Japanese voices on its text-to-speech API. None of those 32 voices appears in the Voice Agent API reference’s list of speak models, as at 20 October 2026, and Deepgram’s Voice Agent TTS Models page describes Aura as its catalogue “across English and Spanish”. The same TTS page links that seven-language catalogue as its complete list of Aura models, and the reference list also omits 7 of the 17 Spanish voices, so it may simply lag the catalogue. Test the exact voice in a Settings message before you design around it.

Does Deepgram support Australian, Indian and other English accents in a voice agent?

Yes, on both sides. Nova-3 lists en-US, en-AU, en-GB, en-IN and en-NZ, and Flux English is described as “English (all accents)”. The agent’s voice list includes 2 Australian Aura-2 voices and 1 Australian Flux TTS voice, 3 Indian Flux TTS voices and 12 British voices across Aura-2, Flux TTS and first-generation Aura. There is no New Zealand voice, and transcripts use American spelling even when you set en-GB.

Which Spanish dialects can a Deepgram voice agent handle?

Nova-3 lists two Spanish codes, es and es-419 (Latin America). Flux Multilingual takes es and treats any regional subtag it does not match as plain Spanish. On the voice side the agent’s list holds 10 Spanish Aura-2 voices: 4 Peninsular, 3 Mexican, 2 Latin American and 1 Colombian. Five of them (Aquila, Carina, Diana, Javier and Selena) switch between English and Spanish inside one reply.

Can I use a non-Deepgram voice or LLM with the Deepgram Voice Agent API?

Yes for the voice and the LLM, no for the ears. The agent’s speak setting accepts Deepgram, Deepgram-managed Cartesia, and your own OpenAI, ElevenLabs, Cartesia or Amazon Polly credentials, and the think setting accepts six LLM provider types. The listen provider is Deepgram only. Deepgram’s Configure the Voice Agent page states “Currently only Deepgram is supported” for speech-to-text.

Does a Deepgram voice agent switch language automatically mid-call?

It can hear a switch. Flux Multilingual detects the language each turn and returns a languages field, and Nova-3 with language set to multi code-switches across 10 languages. Speaking the switch is narrower: with Deepgram voices, only the five English and Spanish codeswitching voices do it, and the multi setting is accepted only for ElevenLabs, OpenAI and Cartesia voices.

Where every figure on this page comes from