Transcriber Config

Overview

Sulus provides several ways to configure transcribers for your assistants, from built-in providers to custom integrations and fallback plans.

Sulus supports a wide range of speech-to-text providers that can be configured directly on your assistant. Supported providers include:

• Deepgram (e.g., nova-3, nova-2, Flux models)
• AssemblyAI (Universal-Streaming)
• Azure
• Gladia
• Speechmatics
• Soniox
• Google
• OpenAI
• ElevenLabs (scribe_v1)
• Cartesia (ink-whisper)

Set the transcriber field on your assistant with a provider, model, and language:

{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-3",
    "language": "en"
  }
}

Each provider supports its own configuration options:

Deepgram

• model: nova-3, nova-2, flux-general-en, flux-general-multi, etc.
• language: Language code
• keywords / keyterm: Boost specific words for better recognition
• smartFormat: Enable smart formatting for numbers/dates
• eotThreshold (0.5-0.9): End-of-turn confidence (Flux models only)
• eotTimeoutMs (500-10000): Max wait after speech before finalizing turn (Flux models only)

AssemblyAI

• speechModel: universal-streaming-english (English only — AssemblyAI does not currently offer a multilingual model)
• wordBoost: Custom vocabulary array
• confidenceThreshold (0-1): Minimum confidence for accepting transcriptions (default: 0.4)
• endUtteranceSilenceThreshold: Silence duration (ms) to detect end of utterance
• vadAssistedEndpointingEnabled: Enable VAD-based endpoint detection

Azure

• language: BCP-47 format (e.g., en-US, es-MX)
• segmentationSilenceTimeoutMs (100-5000): Silence duration before phrase finalization
• segmentationStrategy: Default, Time, or Semantic

Soniox

• model: stt-rt-v5
• languageHintsStrict: Restrict to selected language (default: true)
• maxEndpointDelayMs (500-3000): Delay between end of speech and endpoint detection
• customVocabulary: Domain-specific terms array

You can configure fallback transcribers so calls continue if your primary provider fails. Sulus supports two approaches:

• Auto fallback: Sulus automatically routes to an alternative provider, no configuration needed.
• Manual fallback: You specify exact backup providers in priority order.

{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-3",
    "language": "en",
    "fallbackPlan": {
      "autoFallback": {
        "enabled": true
      },
      "transcribers": [
        {
          "provider": "assembly-ai",
          "speechModel": "universal-streaming-english",
          "language": "en"
        },
        {
          "provider": "azure",
          "language": "en-US"
        }
      ]
    }
  }
}

In this example, if Deepgram fails, Sulus tries AssemblyAI first, then Azure, and finally falls back automatically if both fail. Note that AssemblyAI is English-only, so it's only a suitable fallback when your primary language is English — for multilingual fallback coverage, consider Gladia or Speechmatics instead.

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

If you need a transcription service not natively supported, you can integrate your own via WebSocket. Sulus connects to your endpoint, streams binary PCM audio, and expects responses in this format:

{
  "type": "transcriber-response",
  "transcription": "The transcribed text",
  "channel": "customer",
  "transcriptType": "final"
}

The transcriptType field can be "final" (definitive) or "partial" (provisional, superseded by later messages). Configure it in your assistant like this:

{
  "assistant": {
    "transcriber": {
      "provider": "custom-transcriber",
      "server": {
        "url": "wss://your-server.example.com/api/custom-transcriber",
        "credentialId": "your-credential-id"
      }
    }
  }
}

If HIPAA compliance is enabled, only Deepgram, Azure, and Soniox are available as transcriber options (including fallbacks).

In summary, Sulus's transcriber configuration is flexible. You can use built-in providers with rich per-provider settings, set up automatic or manual fallbacks for resilience, or integrate a fully custom transcription service via WebSocket.