AssistantsModel configurations

Transcriber fallback configuration

Configure fallback transcribers that activate automatically if your primary transcriber fails.

Overview

Transcriber fallback configuration ensures your calls continue even if your primary speech-to-text provider experiences issues. Your assistant will sequentially fallback to the transcribers you configure, in the exact order you specify.

Key benefits:

  • Call continuity during provider outages
  • Automatic failover with no user intervention required
  • Provider diversity to protect against single points of failure

Without a fallback plan configured, your call will end with an error if your chosen transcription provider fails.

How it works

When a transcriber failure occurs, Vapi will:

  1. Detect the failure of the primary transcriber
  2. Switch to the first fallback transcriber in your plan
  3. Continue through your specified list if subsequent failures occur
  4. Terminate only if all transcribers in your plan have failed

Configure via Dashboard

1

Open Transcriber tab

Navigate to your assistant and select the Transcriber tab.

2

Expand Fallback Transcribers section

Scroll down to find the Fallback Transcribers collapsible section. A warning indicator appears if no fallback transcribers are configured.

3

Add a fallback transcriber

Click Add Fallback Transcriber to configure your first fallback:

  • Select a provider from the dropdown
  • Choose a model (if the provider offers multiple models)
  • Select a language for transcription
4

Configure provider-specific settings (optional)

Expand Additional Configuration to access provider-specific settings like numerals formatting, VAD settings, and confidence thresholds.

5

Add more fallbacks

Repeat to add additional fallback transcribers. Order matters—the first fallback in your list is tried first.

If HIPAA or PCI compliance is enabled on your account or assistant, only Deepgram and Azure transcribers will be available as fallback options.

Configure via API

Add the fallbackPlan property to your assistant’s transcriber configuration, and specify the fallback transcribers within the transcribers property.

1{
2  "transcriber": {
3    "provider": "deepgram",
4    "model": "nova-3",
5    "language": "en",
6    "fallbackPlan": {
7      "transcribers": [
8        {
9          "provider": "assembly-ai",
10          "speechModel": "universal-streaming-multilingual",
11          "language": "en"
12        },
13        {
14          "provider": "azure",
15          "language": "en-US"
16        }
17      ]
18    }
19  }
20}

Provider-specific settings

Each transcriber provider supports different configuration options. Expand the accordion below to see available settings for each provider.

  • model: Model selection (nova-3, nova-3-general, nova-3-medical, nova-2, flux-general-en, etc.).
  • language: Language code for transcription.
  • keywords: Keywords with optional boost values for improved recognition (e.g., ["companyname", "productname:2"]).
  • keyterm: Keyterm prompting for up to 90% keyword recall rate improvement.
  • smartFormat (boolean): Enable smart formatting for numbers and dates.
  • eotThreshold (0.5-0.9): End-of-turn confidence threshold. Only available with Flux models.
  • eotTimeoutMs (500-10000): Maximum time to wait after speech before finalizing turn. Only available with Flux models. Default is 5000ms.
  • language: Language code (multi for multilingual, en for English).
  • speechModel: Streaming speech model (universal-streaming-english or universal-streaming-multilingual).
  • wordBoost: Custom vocabulary array (up to 2500 characters total).
  • keytermsPrompt: Array of keyterms for improved recognition (up to 100 terms, 50 characters each). Costs additional $0.04/hour.
  • endUtteranceSilenceThreshold: Duration of silence in milliseconds to detect end of utterance.
  • disablePartialTranscripts (boolean): Set to true to disable partial transcripts.
  • confidenceThreshold (0-1): Minimum confidence threshold for accepting transcriptions. Default is 0.4.
  • vadAssistedEndpointingEnabled (boolean): Enable VAD-based endpoint detection.
  • language: Language code in BCP-47 format (e.g., en-US, es-MX, fr-FR).
  • segmentationSilenceTimeoutMs (100-5000): Duration of silence after which a phrase is finalized. Configure to adjust sensitivity to pauses.
  • segmentationMaximumTimeMs (20000-70000): Maximum duration a segment can reach before being cut off.
  • segmentationStrategy: Controls phrase boundary detection. Options: Default, Time, or Semantic.
  • model: Model selection (fast, accurate, or solaria-1).
  • language: Language code.
  • confidenceThreshold (0-1): Minimum confidence for transcription acceptance. Default is 0.4.
  • endpointing (0.01-10): Time in seconds to wait before considering speech ended.
  • speechThreshold (0-1): Speech detection sensitivity (0.0 to 1.0).
  • prosody (boolean): Enable prosody detection (laugh, giggle, music, etc.).
  • audioEnhancer (boolean): Pre-process audio for improved accuracy (increases latency).
  • transcriptionHint: Hint text to guide transcription.
  • customVocabularyEnabled (boolean): Enable custom vocabulary.
  • customVocabularyConfig: Custom vocabulary configuration with vocabulary array and default intensity.
  • region: Processing region (us-west or eu-west).
  • receivePartialTranscripts (boolean): Enable partial transcript delivery.
  • model: Model selection (currently only default).
  • language: Language code.
  • operatingPoint: Accuracy level. standard for faster turnaround, enhanced for highest accuracy. Default is enhanced.
  • region: Processing region (eu for Europe, us for United States). Default is eu.
  • enableDiarization (boolean): Enable speaker identification for multi-speaker conversations.
  • maxDelayMs: Maximum delay in milliseconds for partial transcripts. Balances latency and accuracy.
  • model: Gemini model selection.
  • language: Language selection (e.g., Multilingual, English, Spanish, French).
  • model: OpenAI Realtime STT model selection (required).
  • language: Language code for transcription.
  • model: Model selection (currently only scribe_v1).
  • language: ISO 639-1 language code.
  • model: Model selection (currently only ink-whisper).
  • language: ISO 639-1 language code.

Best practices

  • Use different providers for fallbacks to protect against provider-wide outages.
  • Consider language compatibility when selecting fallbacks—ensure all fallback transcribers support your required languages.
  • Test your fallback configuration to ensure smooth transitions between transcribers.
  • For HIPAA/PCI compliance, ensure all fallbacks are compliant providers (Deepgram or Azure).

FAQ

All major transcriber providers are supported: Deepgram, AssemblyAI, Azure, Gladia, Google, Speechmatics, Cartesia, ElevenLabs, and OpenAI.

No additional fees for using fallback transcribers. You are only billed for the transcriber that processes the audio.

Failover typically occurs within milliseconds of detecting a failure, ensuring minimal disruption to the call.

Yes, each fallback transcriber can have its own language configuration. However, for the best user experience, we recommend using the same or similar languages across all fallbacks.