> ## Documentation Index
> Fetch the complete documentation index at: https://dialnexa.com/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Speech To Text And Transcription In DialNexa

> Compare DialNexa speech to text providers for realtime transcription, language support, response timing, transcript accuracy, and cost.

Speech to text in DialNexa is the part of a Voice AI call that decides what the system believes the caller said. If this layer is wrong, every downstream feature inherits the mistake: the model answers the wrong question, functions receive wrong arguments, summaries sound confident but false, and post-call fields become review work.

![DialNexa transcriber selector showing the Soniox option for the selected agent language.](https://mintcdn.com/dialnexa/O6bVvssz6DpTKOa0/images/documentation/screenshots/transcriber-selector-pricing.jpg?fit=max\&auto=format\&n=O6bVvssz6DpTKOa0\&q=85\&s=f8dd1990f9a314901d0b55a9b5467ceb)

![DialNexa call detail page showing summary, transcription tabs, live transcript, and accurate transcript controls.](https://mintcdn.com/dialnexa/0efoAN-6So4r-6NC/images/documentation/screenshots/call-detail-transcript-tabs.png?fit=max\&auto=format\&n=0efoAN-6So4r-6NC\&q=85\&s=c2e40167feeb2c2fdb7d935a89a53807)

> [!NOTE]
>
> If the transcript says the caller asked about carrots when they clearly asked about careers, do not rewrite the whole prompt yet. Start with listening.

## DialNexa Speech To Text And Transcription Options

| Display name | Provider | Best fit | Important limit |
| - | - | - | - |
| Deepgram Flux | Deepgram `flux-general-en` for English, or `flux-general-multi` for supported non-English and mixed-language calls. | Fast turn boundaries, English calls, and supported Hindi or Hinglish tests where Flux is enabled. | Test real caller audio. Multilingual Flux uses language hints where DialNexa can derive them, and broad multilingual choices may run auto-detect. |
| Soniox | Soniox `stt-rt-v5` | Hindi-English, Hinglish, multilingual, and Indian-accented calls. | Test noisy phone calls and local names before scaling. |
| Sarvam Saaras | Sarvam `saaras:v4` | Supported Indian-language and Hindi-English calls. | Compare short pauses, names, and mixed-language phrases. Recognition hints are best effort. |

The selector can show different options depending on the agent language. Pricing per minute for each option is shown next to the selector in the dashboard. Cascaded agents can also use fallback STT from the transcriber settings popover. The current dashboard selector does not expose AssemblyAI for new primary or fallback choices.

For help comparing transcription providers, see the [voice AI provider selection guide](https://dialnexa.com/docs/voice-ai/provider-selection-guide.md). For how transcription evidence appears after the call, see [transcripts, recordings, and summaries](https://dialnexa.com/docs/monitoring/transcripts-recordings-and-summaries.md).

## Deepgram Flux Versus Soniox

> [tab] Use Deepgram Flux
>
> Start with Deepgram Flux for English calls where fast turn boundaries matter. For supported non-English or mixed-language tests, DialNexa can use multilingual Flux with language hints derived from the agent language.

> [tab] Use Soniox
>
> Use Soniox when callers naturally mix Hindi and English, speak in Indian regional accents, or when Deepgram produces high error rates for your caller base. DialNexa sends English and Hindi language hints to Soniox. Response Eagerness adjusts Soniox maximum endpoint delay from 3 seconds (patient) to 2 seconds (eager), without an extra transcript hold.

## Sarvam Saaras Speech To Text

Sarvam Saaras is selectable across workspaces for supported agent languages, as a primary or compatible fallback transcriber. Open the transcriber selector in a draft, review the workspace rate preview, and choose **Sarvam Saaras**. Keep the fallback different from the primary. Sarvam fallback transcription is configured at ₹0 per minute; the primary Sarvam rate is separate. Confirm the workspace rate preview before saving.

Use **Response Eagerness** in Speech Settings to test how the agent handles pauses within an answer. A more patient setting can join speech that the transcriber split into separate final results. Use [Boosted Keywords](https://dialnexa.com/docs/agent-settings/speech-settings.md#sarvam-recognition-hints) for important names and terms, then verify a controlled call's recording and transcript before publishing.

## What DialNexa Configures Behind The Selector

These details explain the behavior users see in the dashboard and call records.

| DialNexa behavior | User-visible result |
| - | - |
| Deepgram Flux uses short endpointing for English. | English calls feel quicker because the system closes caller turns sooner. |
| Deepgram Flux can use the multilingual model for supported non-English languages. | A Hindi or Hindi-English agent can run through `flux-general-multi` with language hints such as `hi` and `en` when DialNexa can derive them. |
| Soniox uses STT RT v5 and receives valid language hints. | Supported single-language agents send that language's ISO hint in strict mode. Hindi-English calls send `en` and `hi` hints so mixed-language speech is treated as code-switched speech rather than one textbook language. Unsupported or broad language choices run without strict hints. |
| Soniox endpoint detection is enabled. | Soniox can finalize caller turns without an extra post-final hold, which can reduce response latency on supported calls. |
| Soniox, Sarvam and Deepgram support Response Eagerness. | Sarvam and Deepgram Nova use a short wait after a final transcript to join speech split by pauses. Flux adjusts its end-of-turn confidence instead. Soniox adjusts its maximum endpoint delay without an extra transcript hold. |
| Boosted Keywords are sent through the selected provider's supported path. | Deepgram Nova 3 and Flux receive keyterm hints, older Deepgram keyword paths receive keyword hints, Soniox receives context terms, and Sarvam receives a best-effort prompt hint. |
| Provider errors are handled during runtime. | Temporary connection drops, malformed frames, or provider restarts can be dropped, retried, or recovered instead of immediately ending the call. |
| Fallback STT runs a backup transcriber in parallel. | Users can choose a different fallback transcriber and a fallback wait time in milliseconds. After one engine wins a final result, finals from the other engine are suppressed for `3` seconds to avoid replying twice to the same utterance. |
| Transcriber pricing is fetched per selected option. | Users can see INR per minute beside primary and fallback transcriber rows when the workspace billing preview is available. |

## Boosted Keywords For Speech To Text

Boosted Keywords are recognition hints for supported cascaded transcribers. Add short terms that the caller is likely to say and that the transcriber might otherwise miss.

| Use Boosted Keywords for | Avoid using them for |
| - | - |
| Brand names, product names, people names, localities, acronyms, and uncommon phrases. | Long instructions, whole scripts, punctuation-heavy text, or terms the caller will never say. |
| Terms that appear in recordings but are misheard in transcripts. | Guessing before you have call evidence. |
| Small, reviewed lists that match the campaign or use case. | Large glossaries that make the hint list noisy. |

The Speech Settings panel validates Boosted Keywords before saving. Terms can contain letters, numbers, and spaces only. DialNexa preserves multi-word phrases, removes duplicate terms case-insensitively, and sends the resulting list to Deepgram, Soniox, or Sarvam when the selected transcriber supports it.

When **Boost dynamic variables** is enabled, DialNexa also adds usable per-call values from the test-call form or batch recipient row. The agent's configured terms keep priority. Dynamic values are cleaned and appended only while the combined list stays within the same term and length limits. See [Speech Settings](https://dialnexa.com/docs/agent-settings/speech-settings.md#boost-dynamic-variables) for the exact rules.

## Fallback STT

Fallback STT helps when the primary transcriber is slow or unreliable for a specific caller population. When enabled, DialNexa can use a backup transcriber if it finalizes first and the primary result does not arrive within the configured wait.

![DialNexa transcriber settings popover showing fallback STT enabled with a fallback transcriber selected and a fallback wait value.](https://mintcdn.com/dialnexa/hO53UXdd8cKV3leF/images/documentation/screenshots/transcriber-fallback-stt-settings.png?fit=max\&auto=format\&n=hO53UXdd8cKV3leF\&q=85\&s=93bcbec8718073339cbae6c15d9ad707)

| Setting | Recommendation |
| - | - |
| Primary transcriber | Keep this as your best-fit provider for the caller language. |
| Fallback transcriber | Pick a different provider or model from the primary option. |
| Fallback wait | Start with the `0 ms` default, then increase it only when call evidence shows the primary needs more time. |

> [!NOTE]
>
> Fallback STT applies to cascaded agents. Speech to Speech agents do not use a separate STT provider.

The fallback transcriber must have a different transcriber ID from the primary transcriber. The dashboard hides the primary option from the fallback list. Create or update API requests that send the same ID for both return `400 Bad Request`.

The fallback selector requests the applicable rate from the current workspace plan. If the preview is unavailable, review Billing transactions after a test call for a separate fallback-transcriber component before scaling the agent.

DialNexa processes a valid fallback final result only once. If the fallback result is queued again during timing coordination, it is not dropped just because the first pass was deferred. After either the primary or fallback engine wins, the other engine's finals are blocked for `3` seconds. Finals from the same winning engine remain eligible, so a new utterance recognized by that engine is not blocked by this cross-engine window.

New cascaded agents start with Soniox STT RT v5 as the primary transcriber and Deepgram Flux as the fallback where the selected language supports that combination. If Flux does not support the selected language, DialNexa can substitute a compatible Deepgram model or leave fallback disabled. Always confirm the saved draft because language compatibility can change the default pair.

## Transcript Types

DialNexa can show more than one transcript view for the same call.

| Transcript | When to use it |
| - | - |
| Live transcript | Debug turn-taking, interruptions, and what the agent heard before replying. |
| Accurate transcript | Review cleaner post-call text when available. |
| Recording | Settle disputes about what was said, noise, overlap, silence, pronunciation, and phone quality. |

> [!TIP]
>
> Text tells you what the system believed. Audio tells you what happened. When they disagree, trust the recording first.

## How To Test Speech To Text

**1. Use the same caller script**

Compare transcribers with the same greeting, caller answers, interruption, name, city, number, and final outcome.

**2. Test names and places**

Use real customer names, locality names, company names, product terms, and common abbreviations.

**3. Test interruptions**

Ask a caller to speak during the greeting, correct themselves mid-answer, pause, and give one-word replies.

**4. Test language switching**

For Hindi-English calls, include English numbers, Hindi phrases, and mixed casual replies in the same call.

**5. Compare transcript with recording**

Mark the exact point where the transcript diverges from audio.

**6. Review downstream effects**

Check function arguments, workflow branches, summaries, and post-call fields. Bad listening quietly becomes bad automation.

## How Transcription Affects Integrations

Integrations usually receive data that started as caller speech. If the transcript is wrong, a CRM update, ticket note, WhatsApp message, or spreadsheet row can be wrong too.

| Integration result | Transcription risk | What to inspect first |
| - | - | - |
| CRM field update in [HubSpot](https://dialnexa.com/docs/integrations/hubspot.md) or [Salesforce](https://dialnexa.com/docs/integrations/salesforce.md). | Name, company, lead score, or intent was misheard. | Recording plus accurate transcript. |
| Ticket note in [Zendesk](https://dialnexa.com/docs/integrations/zendesk.md) or [Intercom](https://dialnexa.com/docs/integrations/intercom.md). | Issue category or callback detail was captured incorrectly. | Transcript segment near the caller's problem statement. |
| WhatsApp or email follow-up. | The agent repeats the wrong date, amount, or next step. | Post-call fields and the exact generated message. |
| Review queue in [Google Sheets](https://dialnexa.com/docs/integrations/googlesheets.md). | Bad extraction looks like a bad lead record. | Transcript, summary, and field definitions together. |

## Common Speech To Text Mistakes

> [accordion] Assuming one transcriber fits every language
>
> Deepgram Flux and Soniox can both be useful, but they behave differently by language, accent, and noise. Compare them on real calls before standardizing a stack.

> [accordion] Passing invalid language hints to Soniox
>
> The runtime filters language hints to valid Soniox values. For Hindi-English agents, DialNexa sends English and Hindi hints rather than an invalid combined language code. For unsupported broad choices, it avoids strict hints instead of sending a value Soniox cannot use.

> [accordion] Treating Boosted Keywords as a transcript guarantee
>
> Boosted Keywords bias recognition toward important terms. They do not force exact output, so recordings and transcript review are still required.

> [accordion] Changing the model before checking the transcript
>
> If the LLM received the wrong caller words, the LLM is not the first problem.

> [accordion] Ignoring the phone path
>
> Plivo, SIP trunking, and web calls can produce different audio with the same agent. Compare through the same route when testing transcribers.

> [accordion] Testing only quiet calls
>
> Real callers use speaker mode, traffic, low signal, and short answers. Test those before a large campaign does it for you.

## Recap

Choose a transcriber using real caller language, accents, noise, and timing requirements. Compare transcript accuracy and response behavior with the same test calls before publishing the selected stack.

## Related Reading

> [card] [Supported Transcribers](https://dialnexa.com/docs/reference/supported-transcribers.md)
>
> Read the reference table for selectable transcribers.

> [card] [Speech Settings](https://dialnexa.com/docs/agent-settings/speech-settings.md)
>
> Tune Response Eagerness and Denoising Mode.

> [card] [Latency And Turn Taking](https://dialnexa.com/docs/voice-ai/latency-turn-taking-and-interruptions.md)
>
> Understand how turn boundaries affect response timing.

> [card] [Transcripts And Recordings](https://dialnexa.com/docs/monitoring/transcripts-recordings-and-summaries.md)
>
> Review call evidence after live calls.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.