Languages Voices Models And Transcribers In DialNexa

Choose the right DialNexa language, voice, voice model, LLM model, and transcriber stack for English, Hindi-English, low-latency, and cost-aware Voice AI agents.

Languages, voices, models, and transcribers define the conversation stack for one agent version. To a caller, that stack becomes a simple experience: did the agent understand me, reply naturally, use the right facts, and finish the task? This page helps you choose those controls as one working system instead of four disconnected dropdowns.

Do not choose the stack like a buffet. A great voice, wrong language, and impatient transcriber will still produce a strange call.

What Each Selector Owns

Selector	What it controls	Concrete dashboard behavior
Voice	Speaker identity, provider, sample audio, provider voice id, Nexa voice id, and voice language options.	The voice modal can search by name, provider voice id, raw voice id, voice_ id, or vel_ id.
Language	The language used by the agent version and compatibility rules.	In the newer builder, language is selected per voice row. In legacy layouts, a separate language selector can appear.
Voice model	Text to speech model allowed for the selected voice.	ElevenLabs selection currently standardizes on Flash v2.5 for supported dashboard choices.
LLM model	Reasoning, tool calls, structured output, fallback behavior, pipeline type, and model cost.	Cascaded agents default to GPT-4o Mini when available. Speech to Speech agents show only speech-to-speech models.
Transcriber	Speech to text provider, primary model, and optional fallback STT.	The current transcriber selector lists Deepgram and Soniox choices, excludes AssemblyAI from new primary and fallback choices, and shows INR per minute where available.

Selection Order That Avoids Rework

1. Pick the caller language first

   Start with English, Hindi, Hindi-English, or another enabled language. Language decides which voice rows are useful and which transcribers are valid.

2. Choose the transcriber

   For cascaded agents, use Deepgram Flux for English-only agents or Soniox for Hindi-English, multilingual, Indian English, or accent-heavy calls. If primary STT latency is a concern, open the transcriber settings and configure fallback STT.

3. Choose the voice

   Open the voice modal, filter by language, listen to samples, copy the Nexa voice ID if needed, and choose the row language before clicking Use Voice.

4. Choose the model

   Pick the LLM model after the listening and speaking layer is sensible. A strong model cannot reason over words the transcriber never heard.

5. Review pricing preview

   Check visible INR rates for LLM, transcriber, and voice model where available. Telephony is still separate.

Compatibility Rules The Builder Applies

Rule	What users experience	Why it matters
Deepgram Flux is English only in the UI.	If Flux is selected while a non-English language is active, the dashboard can move the language back to English where possible.	Flux is for English turn boundaries, not broad multilingual calls.
Speech to Speech agents use a direct speech model.	Voice model, transcriber, and Audio Cache controls are hidden or ignored for Speech to Speech agents.	The realtime model listens and speaks directly instead of using separate STT and TTS services.
Fallback STT is configured from the transcriber settings popover.	Users can enable fallback STT, choose a different fallback transcriber, and set a fallback wait in milliseconds.	This helps reduce failed or delayed recognition when the primary STT path is slow.
Audio Cache is on by default for cascaded agents.	Non-super-admin users can see Audio Cache, but disabling it requires DialNexa support.	Repeated phrases should stay fast by default, especially in campaigns.
Soniox exposes Response Eagerness.	The Speech Settings panel shows Response Eagerness only for supported Soniox paths.	This is where users tune whether the agent replies sooner or waits longer.
Hindi-English exposes Hinglish Map.	Hinglish Map appears when the language code is Hindi-English.	Users can replace formal Hindi wording with natural mixed-language phrasing.
Voice models depend on the selected voice.	The voice settings popover fetches models for the current voice.	Do not assume a voice model available for one voice is valid for another.
Published versions lock important controls.	Selectors and settings can be disabled after publish.	Create or edit a draft version when testing a stack change.

Voice Selector Details Users Should Know

The voice modal is more than a list of pretty names.

Control	Use it for
Provider filter	Switch between all public voice providers: ElevenLabs, Cartesia, SmallestAI, and Sarvam AI.
Language filter	Show voices that support one or more selected languages. The modal now starts with no language filter so users can see the full provider catalog before narrowing it.
Per-row language dropdown	Select the exact language for that voice before using it.
Gender and accent filters	Narrow a large voice library without opening every sample.
Search	Search by voice name, provider voice id, internal id, voice_ id, or Nexa voice id.
Copy Nexa voice ID	Copy the vel_ voice id for internal notes, support, or API-oriented setup discussions.
Sample playback	Listen before testing. Then still place a real test call, because sample audio does not include your prompt.

Speech To Speech Agent Stack

Speech to Speech agents use a realtime speech model for both listening and speaking. Choose this pipeline when low turn latency matters more than separate control over STT and TTS providers.

Control	Cascaded agent	Speech to Speech agent
Model selector	Shows text-to-text LLMs.	Shows speech-to-speech models only.
Voice selector	Requires voice and voice model compatibility.	Uses compatible realtime voice options and does not require a separate voice model.
Transcriber selector	Selects primary STT and optional fallback STT.	Not used.
Audio Cache	Available and enabled by default.	Not used because there is no separate TTS cache.
Max call duration slider	Up to 90 minutes.	Up to 60 minutes.
Pricing preview	LLM, voice engine, and transcriber lines can appear.	Shows realtime model pricing without separate voice engine or transcriber lines.

Speech to Speech agents currently start from a blank configuration in the create-agent dialog. Build and test the prompt directly instead of starting from a template.

Fallback STT

Fallback STT runs a backup transcriber alongside the primary transcriber for cascaded agents. Enable it when call quality, accents, or provider latency make recognition reliability more important than the lowest possible transcription cost.

Setting	What it does
Enable fallback STT	Turns on parallel backup recognition for the agent version.
Fallback transcriber	Chooses a different backup transcriber from the primary one.
Fallback wait	Sets how long, in milliseconds, DialNexa waits for the primary result after the fallback result finalizes first.

Fallback STT is version-specific. Published versions lock the setting, so create or edit a draft version before changing fallback STT in production.

Stack Recommendations

Use case	Start with	Test carefully
English support agent.	English, Deepgram Flux, GPT-4o Mini or another OpenAI option, ElevenLabs or Cartesia.	Names, ticket numbers, short answers, and interruptions.
English interruption-heavy agent.	English, Deepgram Flux (English only), concise prompt, fallback LLM if model latency is a real issue.	Early caller speech during the greeting.
Hindi-English sales or support agent.	Hindi-English, Soniox, Hinglish Map, SmallestAI or ElevenLabs voice tested on mixed phrases.	Locality names, English numbers, and casual Hindi-English replies.
Indian English agent.	Indian English, Soniox, GPT-4o Mini, Sarvam AI (en-IN).	Indian names, amounts in Indian numbering, locality names.
Appointment booking or payment reminder.	Stable transcriber, low LLM temperature, explicit functions, post-call fields.	Function arguments and extracted fields.
Repeated outbound campaign.	Short repeated lines, Audio Cache enabled, stable voice configuration.	Cache hit rate and first audio delay on repeat calls.
Latency-sensitive web call.	Speech to Speech agent, compatible realtime model, concise prompt.	Greeting timing, interruptions, and browser audio quality.

Add The Action Layer Only After The Stack Works

Once the caller can be heard and answered correctly, connect the call to the system that should receive the result.

Call result	Where it usually belongs	Useful docs
Meeting booked or rescheduled.	Calendar plus written confirmation.	Google Calendar, Gmail, email with Resend.
Qualified lead or renewal risk.	CRM and owner alert.	HubSpot, Salesforce, Slack.
Support issue.	Ticketing or support inbox.	Zendesk, Intercom.
Campaign recipient result.	Spreadsheet, workflow branch, or follow-up message.	Google Sheets, workflow integrations, WhatsApp with Wati.

What To Compare Before Publishing

Language

Ask callers to use the language mix you expect in production. Do not approve a Hindi-English agent from a polished English-only demo.

Transcriber

Compare transcript and recording together. If the transcript is wrong, model tuning is the wrong first fix.

Voice

Test the greeting, the longest compliance line, a name, a date, an amount, and a goodbye line. Voices often fail on numbers before they fail on vibes.

LLM

Keep temperature low for functions and structured results. Compare models with the same prompt and same caller script.

Cost

Use the pricing preview as a selection signal, but do not optimize cost before the call works. Cheap wrong calls are still wrong calls.

Related Reading

Choose Voice AI Providers

Decide what to use when.

Speech To Text

Compare Deepgram and Soniox.

Text To Speech

Tune ElevenLabs and Cartesia voices.

LLMs And Conversation Behavior

Tune model behavior and fallback.