Transcription Settings

Telnyx AI Assistants support multiple speech-to-text (STT) models for transcribing caller audio. The model you choose affects transcription accuracy, supported languages, and response latency. You can also tune provider-specific transcription behavior, such as end-of-turn detection, formatting, keyterm boosting, and Azure region selection.

Available models

Model	Engine	Best for
`deepgram/flux`	Deepgram	Conversational AI, optimized for turn-taking with multilingual support
`deepgram/nova-3`	Deepgram	Fast multilingual transcription, recommended for multilingual assistants
`deepgram/nova-2`	Deepgram	Fast multilingual transcription on Deepgram’s previous-generation model
`azure/fast`	Azure	Fast multilingual transcription with optional Azure region and API key configuration
`assemblyai/universal-3-5-pro`	AssemblyAI	Conversational, multilingual streaming transcription with configurable turn detection, backed by Universal-3.5 Pro Realtime. The legacy alias `assemblyai/universal-streaming` resolves to the same model
`xai/grok-stt`	xAI	Multilingual transcription using Grok STT
`soniox/stt-rt-v4`, `soniox/stt-rt-v5`	Soniox	Multilingual streaming transcription with automatic language detection, term biasing, and endpoint detection
`nvidia/parakeet-v3`	Parakeet	Multilingual transcription with automatic language detection
`omi-health/omi-med-stt-v1`	Parakeet	English-only medical transcription (Omi Health, Parakeet-based)
`cohere/ar-stt`	Cohere	Self-hosted Arabic/English transcription. Requires an explicit `ar` or `en` language — no auto-detect
`telnyx/basira`	Telnyx	Non-streaming Arabic transcription
`reson8/turns`	Reson8	Turn-based transcription of 10 European languages with automatic language detection
`speechmatics/standard`	Speechmatics	Self-hosted transcription with 17 language packs, including bilingual packs. No auto-detect; unset or `auto` defaults to `en`, so set a supported code for other languages
`speechmatics/linden-1`	Speechmatics	Preview. Vendor-hosted transcription with 67 supported language codes. No auto-detect; defaults to `en`, so set a supported code for other languages
`humain/realtime`	Humain	Streaming transcription with native Arabic and Arabic/English code-switching support
`distil-whisper/distil-large-v2`	Whisper	Lower-latency English-only transcription

deepgram/flux supports English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, and Dutch. For broader language coverage, use deepgram/nova-3, deepgram/nova-2, azure/fast, assemblyai/universal-3-5-pro, xai/grok-stt, or nvidia/parakeet-v3.

Selecting a model

Portal

In the AI Assistants tab, edit your assistant and navigate to the Voice tab. Select your preferred STT model from the Transcription Model dropdown.

AI Assistant Transcription Model Selection

When you change models in the Portal, related settings are reset to the defaults for that provider. For example:

deepgram/flux supports explicit languages, auto, and multi for its supported languages, and applies Flux end-of-turn defaults.
Other Deepgram models enable smart_format and numerals by default.
assemblyai/universal-3-5-pro applies AssemblyAI turn detection defaults.
soniox/stt-rt-v4 and soniox/stt-rt-v5 disable endpoint detection by default; term biasing and language hints are left unset.
azure/fast defaults the Azure region to latency, which auto-selects the closest supported Telnyx-managed region.
nvidia/parakeet-v3 uses automatic multilingual transcription.
omi-health/omi-med-stt-v1 pins the language to en.
cohere/ar-stt defaults the language to ar. Set it to en explicitly if needed — unlike other models, it does not auto-detect and rejects auto.
reson8/turns defaults the language to auto for automatic detection. Setting an explicit language reduces latency.
speechmatics/standard defaults the language to en; omitting it, or setting auto, also resolves to en.
speechmatics/linden-1 (preview) has no auto-detect. Omitting language defaults to en, so set a code from its supported list for other languages; setting auto is rejected with a 400.

API

Set the transcription.model field when creating or updating an assistant:

curl -X POST https://api.telnyx.com/v2/ai/assistants \
  -H "Authorization: Bearer $TELNYX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "My Assistant",
    "model": "anthropic/claude-haiku-4-5",
    "instructions": "You are a helpful voice assistant.",
    "transcription": {
      "model": "deepgram/flux"
    }
  }'

You can also set the transcription language explicitly. If omitted or set to auto, supported models auto-detect the language:

"transcription": {
  "model": "deepgram/nova-3",
  "language": "es"
}

Languages

Supported language options depend on the selected model.

Model	Language behavior
`deepgram/flux`	English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian, Dutch, plus `auto` and `multi` modes
`deepgram/nova-3`	Auto-detect plus supported Deepgram Nova 3 language codes
`deepgram/nova-2`	Auto-detect plus supported Deepgram Nova 2 language codes
`azure/fast`	Explicit Azure locale codes, such as `en-US`, `es-MX`, or `fr-FR`
`assemblyai/universal-3-5-pro`	`auto` for multilingual detection, or one of 18 supported languages: `en`, `es`, `de`, `fr`, `pt`, `it`, `tr`, `nl`, `sv`, `no`, `da`, `fi`, `hi`, `vi`, `ar`, `he`, `ja`, `zh`
`xai/grok-stt`	`auto` plus supported Grok STT language codes
`soniox/stt-rt-v4`, `soniox/stt-rt-v5`	`auto` (or unset) omits the language hint and lets Soniox auto-detect; ISO 639-1 codes (`en`, `es`, `nl`, `fr`, and others) bias detection toward that language. Use `language_hints` in `settings` to pin multiple languages at once.
`nvidia/parakeet-v3`	Automatic multilingual detection
`cohere/ar-stt`	`ar` (default) or `en` only — no auto-detect. `auto` is rejected.
`reson8/turns`	`auto` (default) for automatic detection, or one of `nl`, `en`, `fr`, `fy`, `de`, `it`, `pl`, `pt`, `es`, `sv`
`speechmatics/standard`	One of `en`, `ba`, `eu`, `gl`, `ga`, `mt`, `mn`, `sw`, `ug`, `cy`, `tl`, or a bilingual pack: `ar_en`, `cmn_en`, `en_ms`, `en_ta`, `es-bilingual-en`, `cmn_en_ms_ta` — no auto-detect. Unset or `auto` defaults to `en`.
`speechmatics/linden-1`	A code from Speechmatics’ Realtime list (67 language codes, e.g. `ar`, `cmn`, `es`, `es-bilingual-en`, `fr`, `de`, `en-GB`, `cmn-Hant`) — no auto-detect. Unset defaults to `en`, so set a code for other languages; explicit `auto` is rejected with a `400`.
`humain/realtime`	`ar`, `en`, `codeswitch` (Arabic/English code-switching), or `auto` (resolves to code-switching). Omitting `language` applies `en`, not `auto`.
`distil-whisper/distil-large-v2`	English only

If your assistant has a language filter set elsewhere in the Voice tab, the Portal only shows transcription models and language choices that are compatible with that language.

Deepgram settings

Deepgram Flux end-of-turn detection

deepgram/flux is optimized for live voice agents. It provides end-of-turn detection so the assistant can start responding as soon as the caller finishes speaking. It also supports eager end-of-turn, which starts large language model (LLM) processing before the caller fully stops speaking to reduce perceived response latency. When you select deepgram/flux in the Portal, these settings are applied by default:

Field	Type	Range	Portal default	Description
`eot_threshold`	number	0.5-0.9	0.8	Confidence required to trigger a final end of turn. Higher values require more confidence and may add latency.
`eot_timeout_ms`	integer	500-10000	5000	Maximum silence duration, in milliseconds, before forcing an end of turn.
`eager_eot_threshold`	number	0.3-0.9	0.4	Confidence required to start speculative LLM processing before final end-of-turn confirmation. Lower values trigger earlier.

eager_eot_threshold must be less than or equal to eot_threshold. Setting both thresholds to the same value effectively disables eager end-of-turn behavior because the system waits for final end-of-turn confirmation before starting LLM processing.

The eager_eot_threshold field is controlled by the FE-eager-eot-threshold Portal feature flag. When that flag is disabled, the Portal hides the field, but API payloads can still include it if your account supports the setting.

When using Flux, the Portal may also prompt you to lower the assistant’s start speaking plan timings. Flux works best with low start speaking delays, such as 0.1 seconds for wait time and endpointing plan thresholds.

Keyterm Boost

keyterm is a comma-separated list of terms to boost during recognition. Use it for product names, customer names, acronyms, or domain-specific vocabulary. It is supported by deepgram/flux, deepgram/nova-3, reson8/turns, speechmatics/linden-1, and assemblyai/universal-3-5-pro (and its legacy alias assemblyai/universal-streaming). The full keyterm value is limited to 1000 characters. AssemblyAI models additionally accept at most 100 distinct terms of up to 50 characters each. Keyterm Boost also supports dynamic variables. Use variables when boosted terms are caller-specific, such as a customer name, participant names, account name, or product names passed into the assistant at conversation start.

"transcription": {
  "model": "deepgram/nova-3",
  "settings": {
    "keyterm": "Telnyx,VoIP,SIP,{{customer_name}},{{product_name}}"
  }
}

Smart Format and Numerals

For Deepgram models other than Flux, the Portal exposes these settings and enables both by default when you select the model:

Field	Type	Default	Description
`smart_format`	boolean	`true`	Automatically formats transcripts for readability, including punctuation and casing.
`numerals`	boolean	`true`	Converts spoken numbers to digits, for example “five hundred” to “500”.

AssemblyAI settings

assemblyai/universal-3-5-pro supports configurable turn detection. When you select it in the Portal, these defaults are applied:

Field	Type	Range	Portal default	Description
`end_of_turn_confidence_threshold`	number	0-1	—	Deprecated and ignored. Accepted for backwards compatibility — turn detection is configured with `min_turn_silence` and `max_turn_silence`.
`min_turn_silence`	integer	100-5000	400	Minimum silence duration, in milliseconds, before a turn can end.
`max_turn_silence`	integer	100-5000	1280	Maximum silence duration, in milliseconds, before forcing an end of turn.
`prompt`	string	≤1750 characters	—	Context about the conversation — the domain, names, or terms likely to come up — to improve transcription. Describe the expected content rather than giving instructions, and leave it empty unless needed.
`keyterm`	string	≤1000 characters in total; ≤100 distinct terms of ≤50 characters each	—	A comma-separated list of terms to boost during recognition, for product names, acronyms, or domain-specific vocabulary. AssemblyAI applies its own term limits on top of the total length cap.

min_turn_silence must be less than or equal to max_turn_silence.

"transcription": {
  "model": "assemblyai/universal-3-5-pro",
  "language": "auto",
  "settings": {
    "prompt": "A customer support call about SIP trunking and VoIP services.",
    "keyterm": "Telnyx,SIP trunking,{{customer_name}}",
    "min_turn_silence": 400,
    "max_turn_silence": 1280
  }
}

Soniox settings

Soniox offers two model versions: soniox/stt-rt-v4 and soniox/stt-rt-v5. Both support multilingual streaming transcription with automatic language detection.

Term Biasing

soniox/stt-rt-v4 and soniox/stt-rt-v5 support context, a comma-separated list of terms to boost during recognition. Use it for staff names, building names, product names, or other domain-specific vocabulary. Term Biasing also supports dynamic variables. Use variables when boosted terms are caller-specific, such as a customer name, participant names, account name, or product names passed into the assistant at conversation start.

"transcription": {
  "model": "soniox/stt-rt-v5",
  "settings": {
    "context": "Telnyx,VoIP,SIP,{{customer_name}},{{product_name}}"
  }
}

Language Hints

language_hints pins recognition to one or more languages at once, as an array of ISO 639-1 codes. Use it when an assistant regularly serves callers in a known set of languages, such as Dutch and French, instead of relying on full auto-detection.

"transcription": {
  "model": "soniox/stt-rt-v5",
  "settings": {
    "language_hints": ["nl", "fr"]
  }
}

Endpoint detection

Field	Type	Range	Default	Description
`interim_results`	boolean	—	`false`	Stream interim (non-final) transcripts in addition to final ones. Useful for live captions or low-latency UI feedback.
`enable_endpoint_detection`	boolean	—	`false`	When enabled, Soniox emits end-of-utterance events at the cadence configured by `max_endpoint_delay_ms`.
`max_endpoint_delay_ms`	integer	500-3000	—	Maximum silence, in milliseconds, before Soniox emits an end-of-utterance event. Only honored when `enable_endpoint_detection` is `true`.

"transcription": {
  "model": "soniox/stt-rt-v5",
  "settings": {
    "enable_endpoint_detection": true,
    "max_endpoint_delay_ms": 1200
  }
}

Parakeet settings

nvidia/parakeet-v3 supports multilingual transcription with automatic language detection. It does not require provider-specific transcription settings.

"transcription": {
  "model": "nvidia/parakeet-v3",
  "language": "auto"
}

Reson8 settings

reson8/turns delivers transcripts per turn of speech: the assistant receives the full transcript of a turn when the caller finishes speaking. Language defaults to auto for automatic detection and can be fixed to one of the 10 supported languages. It does not require provider-specific transcription settings.

"transcription": {
  "model": "reson8/turns",
  "language": "auto"
}

Cohere settings

cohere/ar-stt is a self-hosted model for Arabic and English transcription. Unlike other models, it does not auto-detect the language and requires an explicit ar or en — auto (and any other value) is rejected. Omitting language defaults to ar. It does not require provider-specific transcription settings.

"transcription": {
  "model": "cohere/ar-stt",
  "language": "ar"
}

Speechmatics settings

speechmatics/linden-1 is a preview, vendor-hosted model with broad language coverage (67 language codes, compared with 17 for speechmatics/standard). It has no automatic language detection: omitting language defaults to en, and setting auto is rejected with a 400 (unlike speechmatics/standard, which resolves auto to en). Set the caller’s language explicitly. es-bilingual-en selects the Spanish and English bilingual pack. To choose a spelling or script variant, set language to en-US, en-GB, or en-AU (English), or cmn-Hans (Simplified) or cmn-Hant (Traditional) for Mandarin. en alone uses the default spelling. These codes are accepted for speechmatics/linden-1 only, not speechmatics/standard.

Field	Type	Description
`keyterm`	string	Comma-separated terms to boost during recognition (see Keyterm Boost).
`remove_disfluencies`	boolean	When `true`, filler words such as “um” and “uh” are removed from the transcript. Defaults to `false`.

remove_disfluencies applies to speechmatics/linden-1 only, and only for the languages Speechmatics supports for disfluency removal: ar, cmn, da, de, el, en, es, fr, he, hi, hu, it, ja, nl, pl, pt, and ru. Support follows the base language, so the variants en-US, en-GB, en-AU, cmn-Hans, and cmn-Hant are covered. For other languages, filler words remain in the transcript. See Speechmatics formatting.

"transcription": {
  "model": "speechmatics/linden-1",
  "language": "en-GB",
  "settings": {
    "keyterm": "Telnyx,VoIP,SIP,{{customer_name}},{{product_name}}",
    "remove_disfluencies": true
  }
}

Azure settings

azure/fast supports region selection and an optional Azure API key reference.

Field	Type	Description
`region`	string	Azure transcription region. The Portal defaults to `latency`, which auto-selects the closest supported Telnyx-managed region.
`api_key_ref`	string	Optional integration secret reference for your Azure API key. When provided, the Portal only shows regions that support custom API keys.

Common Telnyx-managed regions include latency, australiaeast, centralindia, eastus, northcentralus, westeurope, and westus2. Additional Azure regions are available when using your own Azure API key.

"transcription": {
  "model": "azure/fast",
  "language": "en-US",
  "region": "latency"
}

Fallback models and language booster

An assistant can run more speech-to-text models next to its transcription model, transcription.model, in one of two modes:

Field	Mode	Behavior
`fallback_models`	Fallback	Up to 3 models, in order. The next model takes over when the model in use fails, at the start of a call or mid-call.
`challenger`	Language booster	One model that transcribes alongside the transcription model. Its `rule` decides which transcript the assistant uses.

An assistant uses one mode at a time. Each fallback model and the language booster take an optional language and settings:

language: omit it to use the transcription model’s language. If the model doesn’t support that language, the request is rejected; set a language for that model.
settings: omit it to use the transcription model’s settings where they apply to the model. A settings object takes the same fields and limits as transcription.settings and replaces the transcription model’s settings for that model. Fields for another provider are dropped, and the provider’s defaults fill in the rest.

Fallback models

The transcription model and every fallback model must be streaming models: deepgram/flux, deepgram/nova-3, deepgram/nova-2, assemblyai/universal-3-5-pro (or its legacy alias assemblyai/universal-streaming), xai/grok-stt, soniox/stt-rt-v4, soniox/stt-rt-v5, humain/realtime, or reson8/turns. Each fallback model must differ from the transcription model and from the other fallback models.

"transcription": {
  "model": "deepgram/flux",
  "language": "en",
  "fallback_models": [
    { "model": "deepgram/nova-3" },
    { "model": "soniox/stt-rt-v5", "settings": { "context": "Telnyx,VoIP,SIP" } }
  ]
}

Language booster

The language booster is set with transcription.challenger. It and the transcription model must both be streaming models, listed above, or both non-streaming models: azure/fast, nvidia/parakeet-v3, omi-health/omi-med-stt-v1, cohere/ar-stt, distil-whisper/distil-large-v2, openai/whisper-large-v3-turbo, or telnyx/basira. The language booster can also be the same model as the transcription model, on another language. Set rule to choose how the assistant picks the transcript:

Rule	Models	Behavior
`best_turn` (default)	Any pair	Both models transcribe the whole call. Each turn uses the language booster’s transcript only when it scores higher than the transcription model’s (clearly higher with non-streaming models).
`best_engine`	Streaming	Both models transcribe the first turns, then the call continues alone on the model whose transcripts scored higher. If neither clearly leads, the transcription model continues.
`merge_words`	`telnyx/basira` with `cohere/ar-stt`	Both models transcribe each utterance, and their words are merged so Arabic and English spoken in the same sentence are both kept. The pair runs on the language that applies to `telnyx/basira`, which must be Arabic (`ar` or an `ar-` locale such as `ar-SA`), `multi`, or `auto`.

best_turn and best_engine score transcripts on how complete and confident they are, not on their language, so they work best when both models understand the callers’ language. Set the model that is more accurate for the callers as transcription.model: the language booster must outscore it to win. Under best_turn with streaming models, the transcription model also decides when each turn ends. While both models transcribe, the end of each turn can wait up to about 1.5 seconds for the other model, and eager end-of-turn, such as Deepgram Flux eager_eot_threshold, does not apply.

"transcription": {
  "model": "deepgram/flux",
  "language": "en",
  "challenger": {
    "model": "deepgram/nova-3",
    "rule": "best_turn"
  }
}

If the language booster can’t run, the call runs on the transcription model alone. If either model fails mid-call, the call continues on the other.

Updates and validation

Omit fallback_models or challenger in an update to keep the stored value. Send null to remove either one; [] also removes fallback_models. A new fallback_models list replaces the stored list.
To switch modes, remove the other mode in the same request. For example, send "fallback_models": [] with a new challenger.
When an update changes transcription.model or transcription.language, the assistant’s stored fallback models or language booster that no longer fit are removed without an error. The response shows what remains.
With fallback models or a language booster other than merge_words, the transcription model must support transcription.language alongside other models. For example, cohere/ar-stt supports only ar there.
A request that sets both modes, more than 3 fallback models, or a model that doesn’t fit is rejected with 400. The error names the model, for example fallback 2 (deepgram/nova-2) doesn't support 'ar' — set a language for it.
In a workflow node, a transcription override that sets model, fallback_models, or challenger doesn’t inherit the assistant’s fallback models or language booster; it uses only the ones it sets. Otherwise the node inherits them, and they must fit the node’s language: a node or assistant change they no longer fit is rejected with 400, naming the node.

Configure advanced settings via API

Create an assistant with tuned transcription settings:

curl -X POST https://api.telnyx.com/v2/ai/assistants \
  -H "Authorization: Bearer $TELNYX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Low Latency Assistant",
    "model": "anthropic/claude-haiku-4-5",
    "instructions": "You are a helpful voice assistant.",
    "transcription": {
      "model": "deepgram/flux",
      "language": "en",
      "settings": {
        "eot_threshold": 0.8,
        "eot_timeout_ms": 5000,
        "eager_eot_threshold": 0.4,
        "keyterm": "Telnyx,VoIP,SIP,{{customer_name}},{{product_name}}"
      }
    }
  }'

Guides

Missions

AI Insights

Assistants API

Missions API

AI Tests API

Conversations API

For AI Agents

Transcription Settings

Available models

Selecting a model

Portal

API

Languages

Deepgram settings

Deepgram Flux end-of-turn detection

Keyterm Boost

Smart Format and Numerals

AssemblyAI settings

Soniox settings

Term Biasing

Language Hints

Endpoint detection

Parakeet settings

Reson8 settings

Cohere settings

Speechmatics settings

Azure settings

Fallback models and language booster

Fallback models

Language booster

Updates and validation

Configure advanced settings via API

​Available models

​Selecting a model

​Portal

​API

​Languages

​Deepgram settings

​Deepgram Flux end-of-turn detection

​Keyterm Boost

​Smart Format and Numerals

​AssemblyAI settings

​Soniox settings

​Term Biasing

​Language Hints

​Endpoint detection

​Parakeet settings

​Reson8 settings

​Cohere settings

​Speechmatics settings

​Azure settings

​Fallback models and language booster

​Fallback models

​Language booster

​Updates and validation

​Configure advanced settings via API

​Related resources

Available models

Selecting a model

Portal

API

Languages

Deepgram settings

Deepgram Flux end-of-turn detection

Keyterm Boost

Smart Format and Numerals

AssemblyAI settings

Soniox settings

Term Biasing

Language Hints

Endpoint detection

Parakeet settings

Reson8 settings

Cohere settings

Speechmatics settings

Azure settings

Fallback models and language booster

Fallback models

Language booster

Updates and validation

Configure advanced settings via API

Related resources