Sign in
← AI PulseAug 23, 2026

Wire · news · Multi-source brief

Vercel AI SDK Deepgram Integration Updated with Transcription and Speech Enhancements

The Vercel AI SDK has released version 3.1.0 of its Deepgram integration, introducing fixes for transcription options, improved speech voice and language composition, and changes to speaker diarization defaults.

By Illumora Editorial

Source · Aug 23, 2026, 1:46 AM · On Illumora · Aug 23, 2026, 1:52 AM

Media from the primary source — shown here so you can stay on Illumora.

Synthesized from multiple allowlisted primaries on the same event. Lanes →

Brief drafted by Illumora’s editorial model from the linked primary source. Ops desk reviews flagged pieces. How we write →

Read the source →Vercel AI SDK Changelog — Release @ai-sdk/deepgram@3.1.0 · vercel/ai
Save

Vercel AI SDK has updated its Deepgram integration to version 3.1.0, addressing several features related to transcription and speech generation. This release includes fixes for how transcription options are handled and refines the composition of speech voice and language settings.

Key Points

  • The @ai-sdk/deepgram@3.1.0 release includes fixes for transcription options and speech voice/language composition.
  • Transcription options such as keyterm, paragraphs, intents, sentiment, and replace are now correctly sent as query parameters in /v1/listen requests.
  • The provider's callable signature for transcription has been widened to accept any transcription model ID, moving beyond the previous 'nova-3' limitation.
  • Speaker diarization (diarize) no longer defaults to true in pre-recorded requests; it is now only sent when explicitly set in providerOptions.deepgram.
  • Bare voice family IDs (e.g., aura-2, aura) now compose the upstream model ID from generateSpeech voice and language options, with language defaulting to en.
  • Full voice IDs (e.g., aura-2-helena-en) continue to pass through unchanged.
  • The DeepgramSpeechModelId union has been trimmed to family IDs plus string types.

Context

According to the Vercel AI SDK Changelog, the @ai-sdk/deepgram@3.1.0 update specifically addresses issues where certain transcription parameters were silently dropped from requests. The change in diarization default behavior is noted as a response to speaker diarization being a paid Deepgram add-on, preventing unintended charges for users who did not explicitly request it. The update also enhances the flexibility of speech generation by allowing broader model ID support and more precise voice composition.

Why It Matters

This update provides developers using the Vercel AI SDK with more control and accuracy when integrating Deepgram's transcription and speech synthesis capabilities. The changes ensure that specified transcription options are correctly applied and that speaker diarization costs are only incurred when explicitly desired, potentially impacting application design and operational expenses.

What To Do

  • Review your existing Deepgram integration code if you rely on speaker diarization for pre-recorded requests.
  • Explicitly pass providerOptions: { deepgram: { diarize: true } } if your application requires speaker diarization.
  • Note the expanded support for transcription model IDs and adjust your model selection if you use models other than 'nova-3'.
  • Test speech generation with bare voice family IDs to ensure correct composition with language options.