Transcriber Provider Comparison

Overview

Sulus supports many speech-to-text providers, but three come up most often when teams need multilingual coverage or maximum accuracy: Deepgram, AssemblyAI, and Gladia. The Transcriber Config page covers general provider setup and fallback plans, and the Multilingual Assistants page covers building a multilingual assistant end-to-end. This page is a focused, feature-by-feature comparison of these three providers so you can pick the right one for your use case.

Deepgram is generally the recommended provider for multilingual voice agents, offering the best balance of speed and multilingual accuracy. To enable multilingual support, set language to "multi" on either the Nova 2 or Nova 3 model:

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

Nova 3 also adds support for Hebrew, Urdu, Tagalog, and Arabic bilingual transcription. See the Transcriber Config page for the full list of Deepgram-specific options (keywords, smart formatting, end-of-turn tuning for Flux models, and more).

AssemblyAI's Universal-Streaming model delivers ultra-fast, immutable transcripts in roughly 300ms with intelligent endpointing, which makes it an excellent choice for natural-feeling voice interactions. Its key limitation: it currently supports English only, so it isn't a fit for multilingual deployments.

{
  "transcriber": {
    "provider": "assembly-ai",
    "speechModel": "universal-streaming-english",
    "language": "en"
  }
}

Choose AssemblyAI when your callers are exclusively English-speaking and you want the lowest possible transcription latency.

Gladia's Solaria model offers the broadest language coverage of the three — 110+ languages and dialects — and handles code-switching (a caller mixing languages mid-conversation) natively. Unlike Deepgram, you don't need to set a specific "multi" language flag; Gladia detects the spoken language automatically:

{
  "transcriber": {
    "provider": "gladia",
    "model": "solaria"
  }
}

Gladia also offers additional audio intelligence features beyond transcription, such as translation, summarization, sentiment analysis, and named-entity recognition, if your use case can take advantage of them.

As a starting point: choose Deepgram for the best all-around balance of speed and multilingual accuracy, choose AssemblyAI when your deployment is English-only and latency is the top priority, and choose Gladia when you need the widest possible language coverage or want built-in audio intelligence features. For general configuration (including fallback plans across providers) see Transcriber Config; for assembling a full multilingual assistant (transcriber, voice, and system prompt working together) see Multilingual Assistants. Note that if HIPAA compliance mode is enabled on your organization, only Deepgram (along with Azure and Soniox) is available among these three as a HIPAA-compliant transcriber option — see the HIPAA & PCI compliance guide for details.