Skip to main content
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 fine-tune transcription behavior through settings like eager end-of-turn detection.

​
Available models

ModelEngineBest for
distil-whisper/distil-large-v2TelnyxDefault — good accuracy, runs on Telnyx infrastructure
openai/whisper-large-v3-turboTelnyxHigh accuracy, larger model, multilingual support
deepgram/fluxDeepgramConversational AI — optimized for turn-taking, English only
deepgram/nova-3DeepgramMultilingual with automatic language detection, slightly higher latency
deepgram/nova-2DeepgramLegacy — stable, cost-effective
azure/fastAzureFast transcription with multilingual support
deepgram/flux is English-only. For multilingual use cases, use deepgram/nova-3, openai/whisper-large-v3-turbo, or azure/fast.

​
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

​
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, the model auto-detects the language: 
"transcription": {
  "model": "deepgram/nova-3",
  "language": "es"
}

​
Eager end-of-turn (Deepgram Flux)

Eager end-of-turn detection is a latency optimization available exclusively with deepgram/flux. It starts large language model (LLM) processing before the caller fully stops speaking, reducing perceived response time by ~150 ms on average. Eager end-of-turn is enabled by default when using deepgram/flux. The eager_eot_threshold defaults to 0.3, meaning the system aggressively begins speculative processing at the earliest sign of a turn ending.

​
How it works

  1. EagerEndOfTurn — The transcription model detects a likely pause. The system begins generating an LLM response speculatively.
  2. TurnResumed — If the caller continues speaking, the speculative response is discarded and transcription continues.
  3. EndOfTurn — When the caller definitively finishes, the final (or already-prepared) response plays back.
Because the LLM starts processing early, the response is often ready the moment the caller stops — eliminating the wait that normally occurs between end-of-speech and response playback.

​
Parameters

These settings are available under transcription.settings and apply only to deepgram/flux:
FieldTypeRangeDefaultDescription
eager_eot_thresholdnumber0.3–0.90.3Confidence level to start speculative LLM processing. Lower values trigger earlier.
eot_thresholdnumber——Confidence level for final end-of-turn confirmation. Must be greater than or equal to eager_eot_threshold.
eot_timeout_msinteger——Maximum silence duration (ms) before forcing an end-of-turn.
The following settings apply to all Deepgram models:
FieldTypeDefaultDescription
smart_formatboolean—Enables automatic punctuation, casing, and formatting.
numeralsboolean—Converts spoken numbers to digits (e.g., “five hundred” → “500”).

​
Trade-offs

Lower eager_eot_threshold (e.g., 0.3)
  • Faster perceived responses — the LLM starts earlier
  • More speculative LLM calls (50–70% increase) — some will be discarded when the caller continues speaking
  • Best for: latency-sensitive applications like customer service bots
Higher eager_eot_threshold (e.g., 0.7)
  • Fewer wasted LLM calls — triggers only on high-confidence pauses
  • Slightly higher response latency
  • Best for: cost-sensitive applications or conversations with frequent mid-sentence pauses
Disabling eager end-of-turn: Set eager_eot_threshold equal to eot_threshold. When both thresholds match, the system waits for full end-of-turn confirmation before starting LLM processing.

​
Tuning recommendations

Use the defaults — eager_eot_threshold: 0.3 provides the fastest response times. Accept the higher LLM call volume as a trade-off for responsiveness.
"transcription": {
  "model": "deepgram/flux",
  "settings": {
    "eager_eot_threshold": 0.3
  }
}
Raise the eager threshold slightly to reduce speculative calls while still getting a latency benefit.
"transcription": {
  "model": "deepgram/flux",
  "settings": {
    "eager_eot_threshold": 0.5
  }
}
Disable eager end-of-turn by matching both thresholds. This avoids any speculative processing.
"transcription": {
  "model": "deepgram/flux",
  "settings": {
    "eager_eot_threshold": 0.7,
    "eot_threshold": 0.7
  }
}

​
Configure 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": {
        "eager_eot_threshold": 0.3,
        "eot_threshold": 0.7,
        "eot_timeout_ms": 5000,
        "smart_format": true,
        "numerals": true
      }
    }
  }'