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. Vapi supports two approaches:

  • Auto fallback — Vapi intelligently routes transcription to an alternative provider when your primary fails. No configuration required.
  • Manual fallback — You specify exact backup providers in priority order for full control over the failover sequence.

You can use both together. When combined, your manual fallbacks are tried first. If all of them fail, Vapi’s auto fallback takes over as a final safety net.

Key benefits:

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

Without any 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 follows this priority order:

  1. Manual fallbacks first — If you’ve configured explicit fallback transcribers, Vapi tries each one sequentially in the order you specified.
  2. Auto fallback as safety net — If all manual fallbacks fail (or none are configured), and auto fallback is enabled, Vapi intelligently selects an alternative provider and routes your transcription audio to it.
  3. Call termination — The call ends only if every fallback option has been exhausted.

Auto fallback

Auto fallback is the simplest way to add resilience. Toggle it on, and Vapi handles provider selection for you—automatically routing transcription audio to an alternative STT provider when your primary fails.

Enabling auto fallback may route audio to other providers. If your organization has strict compliance requirements, review your compliance settings to ensure this aligns with your needs.

To enable auto fallback via API, set transcriber.fallbackPlan.autoFallback.enabled to true:

1{
2  "transcriber": {
3    "provider": "deepgram",
4    "model": "nova-3",
5    "language": "en",
6    "fallbackPlan": {
7      "autoFallback": {
8        "enabled": true
9      }
10    }
11  }
12}

Manual fallbacks

Manual fallbacks give you full control over which providers Vapi tries, and in what order. This is useful when you need specific providers for compliance, language support, or cost reasons.

Configure via Dashboard

1

Open the assistant

Open the Dashboard. Select Assistants, then select an assistant.

2

Open the fallback settings

Select the Advanced tab. In the Fallbacks section, locate Fallback Transcriber.

3

Add a fallback transcriber

Select Add, then choose a Provider, Language, and Model.

4

Configure provider-specific settings (optional)

Expand Additional Configuration, then configure the available settings. Review the provider-specific settings for details about each option.

5

Add more fallbacks

Select Add to configure another fallback transcriber. Vapi attempts fallback transcribers in the order they are listed.

If HIPAA compliance is enabled, choose every fallback from the current list of HIPAA-compliant transcription providers.

Configure via API

Add the fallbackPlan property to your assistant’s transcriber configuration, and specify the fallback transcribers within the transcribers property. You can combine manual fallbacks with auto fallback for maximum resilience.

1{
2  "transcriber": {
3    "provider": "deepgram",
4    "model": "nova-3",
5    "language": "en",
6    "fallbackPlan": {
7      "autoFallback": {
8        "enabled": true
9      },
10      "transcribers": [
11        {
12          "provider": "assembly-ai",
13          "speechModel": "universal-streaming-multilingual",
14          "language": "en"
15        },
16        {
17          "provider": "azure",
18          "language": "en-US"
19        }
20      ]
21    }
22  }
23}

In this example, if Deepgram fails, Vapi tries AssemblyAI first, then Azure. If both manual fallbacks fail, auto fallback intelligently selects another available provider.

Provider-specific settings

Each transcriber provider supports different configuration options. Expand a provider below to review its available settings. For the complete schema, see the provider-specific transcriber fields in the Create Assistant API reference.

  • language: Language code (multi for multilingual or en for English).
  • speechModel: AssemblyAI streaming speech model.
  • mode: Accuracy and latency mode for universal-3-5-pro: max_accuracy, min_latency, or balanced.
  • prompt: Additional context for universal-3-5-pro, up to 1,750 characters.
  • agentContext: Voice-agent context for universal-3-5-pro, up to 1,750 characters.
  • languageCodes: Language hints for universal-3-5-pro automatic language detection.
  • realtimeUrl: WebSocket URL used for the realtime transcription connection.
  • wordBoost: Custom vocabulary array, up to 2,500 characters per entry.
  • keytermsPrompt: Up to 100 keyterms, with a maximum of 50 characters each.
  • endUtteranceSilenceThreshold: Duration of silence in milliseconds used to detect the end of an utterance.
  • disablePartialTranscripts: Set to true to disable partial transcripts.
  • confidenceThreshold: Minimum confidence for accepting transcripts, from 0 to 1.
  • formatTurns: Enables transcript formatting.
  • endOfTurnConfidenceThreshold: Minimum end-of-turn confidence, from 0 to 1.
  • minEndOfTurnSilenceWhenConfident: Minimum silence in milliseconds before ending a confident turn.
  • maxTurnSilence: Maximum turn silence in milliseconds.
  • vadAssistedEndpointingEnabled: Enables VAD-assisted endpoint detection.
  • language: Language code in BCP 47 format, such as en-US, es-MX, or fr-FR.
  • segmentationSilenceTimeoutMs: Duration of silence after which a phrase is finalized, from 100 to 5000 milliseconds.
  • segmentationMaximumTimeMs: Maximum segment duration, from 20000 to 70000 milliseconds.
  • segmentationStrategy: Phrase-boundary strategy: Default, Time, or Semantic.
  • model: Cartesia transcription model.
  • language: Language code. Language availability depends on the selected model.
  • model: Deepgram transcription model.
  • language: Language code for transcription.
  • languages: Language hints for Flux Multilingual. Omit this field to allow automatic detection.
  • keywords: Keywords with optional boost values, such as ["companyname", "productname:2"].
  • keyterm: Important words or phrases for keyterm prompting.
  • smartFormat: Enables formatting for values such as numbers and dates.
  • endpointing: Silence duration used for endpointing, from 10 to 500 milliseconds.
  • mipOptOut: Opts your account out of Deepgram’s Model Improvement Partnership when you use your own Deepgram API key.
  • numerals: Converts spoken numbers to numerals.
  • profanityFilter: Replaces profanity in transcripts with asterisks.
  • redaction: Redacts selected categories: pci, pii, phi, or numbers.
  • confidenceThreshold: Minimum confidence for accepting transcripts, from 0 to 1.
  • eotThreshold: End-of-turn confidence threshold for Flux models, from 0.5 to 0.9.
  • eotTimeoutMs: Maximum wait after speech before finalizing a turn with Flux, from 500 to 10000 milliseconds.

See Deepgram for current models and model-specific language support.

  • model: ElevenLabs transcription model.
  • language: ISO 639-1 language code.
  • silenceThresholdSeconds: Silence before VAD commits, from 0.3 to 3 seconds.
  • confidenceThreshold: VAD sensitivity, from 0.1 to 0.9; lower values are more sensitive.
  • minSpeechDurationMs: Minimum speech duration for VAD, from 50 to 2000 milliseconds.
  • minSilenceDurationMs: Minimum silence duration for VAD, from 50 to 2000 milliseconds.
  • model: Gladia transcription model.
  • languageBehaviour: Controls automatic or manual language detection.
  • language: Language code.
  • languages: Language codes used when manual multilingual detection is configured.
  • confidenceThreshold: Minimum confidence for accepting transcripts, from 0 to 1.
  • endpointing: Time to wait before considering speech complete, from 0.01 to 10 seconds.
  • speechThreshold: Speech-detection sensitivity, from 0 to 1.
  • prosody: Enables detection of non-speech events such as laughter and music.
  • audioEnhancer: Preprocesses audio to improve accuracy, with additional latency.
  • transcriptionHint: Context that guides transcription, up to 600 characters.
  • customVocabularyEnabled: Enables custom vocabulary.
  • customVocabularyConfig: Custom vocabulary and intensity configuration.
  • region: Processing region, us-west or eu-west.
  • receivePartialTranscripts: Enables partial transcripts.
  • model: Gemini transcription model.
  • language: Language selection, such as Multilingual, English, Spanish, or French.
  • model: OpenAI realtime transcription model.
  • language: Language code for transcription.
  • model: Soniox transcription model.
  • language: ISO 639-1 language code.
  • languages: Expected-language hints. An empty array enables automatic detection without hints.
  • languageHintsStrict: Restricts transcription to selected language hints when enabled.
  • maxEndpointDelayMs: Maximum endpoint delay, from 500 to 3000 milliseconds.
  • customVocabulary: Brand names, product names, or other domain-specific terms.
  • contextGeneral: Key-value pairs that provide domain or topic context.

See Soniox for current model and language details.

  • model: Speechmatics transcription model.
  • language: Language code or auto for automatic detection.
  • operatingPoint: standard for faster transcription or enhanced for higher accuracy. Defaults to enhanced.
  • region: Processing region, eu or us. Defaults to eu.
  • enableDiarization: Enables speaker identification for multi-speaker conversations.
  • maxDelay: Maximum delay for partial transcripts, from 500 to 10000 milliseconds. Defaults to 3000.
  • customVocabulary: Words or phrases to improve recognition, with optional soundsLike variations.
  • numeralStyle: Number formatting style: written or spoken. Defaults to written.
  • endOfTurnSensitivity: End-of-turn sensitivity, from 0 to 1. Defaults to 0.5.
  • removeDisfluencies: Removes disfluencies from English transcripts when enabled.
  • minimumSpeechDuration: Minimum speech-segment duration, from 0 to 5 seconds.
  • model: Talkscriber transcription model.
  • language: Language code for transcription.

See Talkscriber for the supported model and languages.

  • model: xAI transcription model.
  • language: Language code for transcription.

See xAI for the supported model and languages.

Best practices

  • Start with auto fallback for quick, zero-config resilience—it works well for most use cases.
  • Add manual fallbacks when you need control over specific providers for compliance, language, or cost reasons.
  • Combine both for maximum reliability—manual fallbacks run first, auto fallback catches anything they miss.
  • Use different providers for manual fallbacks to protect against provider-wide outages.
  • Consider language compatibility when selecting fallbacks—ensure all fallback transcribers support your required languages.
  • For HIPAA compliance, choose every fallback from the current list of HIPAA-compliant transcription providers and review data-routing implications before enabling auto fallback.

FAQ

Auto fallback lets Vapi intelligently select an alternative provider for you—no configuration needed. Manual fallback lets you specify exact providers in a specific priority order. You can use both together: manual fallbacks are tried first, and auto fallback acts as a safety net if they all fail.

Review Transcriber providers for the current provider list and the fallbackPlan.transcribers schema for accepted fallback configurations.

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. When auto fallback activates, Vapi may route transcription audio to a different cloud provider than your primary. If you have data residency or compliance requirements, we recommend reviewing your organization’s policies to ensure this aligns with your needs.

If your primary transcriber supports native endpointing (like Deepgram Flux), that capability won’t carry over to the fallback provider. Vapi will automatically switch to heuristic-based endpointing to continue detecting turn boundaries.

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.