AI SDK @ai-sdk/deepgram 3.1.46: live streaming transcription for Nova models (November 5, 2026)

What changed

@ai-sdk/deepgram 3.1.46, released November 5, 2026, adds live streaming transcription for Deepgram Nova models. You call experimental_streamTranscribe from ai and read partial and final transcripts, word timings and usage as they arrive. It runs on Deepgram's /v1/listen WebSocket API. Flux models are not supported by this integration.

Prerequisites

A Deepgram API key in DEEPGRAM_API_KEY
ai 7 and @ai-sdk/deepgram 3.1.46 or later
Audio as a ReadableStream of 16 kHz PCM

Steps

Install: pnpm add ai @ai-sdk/deepgram.
Stream your audio in:

TypeScript

import { deepgram } from '@ai-sdk/deepgram';
import { experimental_streamTranscribe as streamTranscribe } from 'ai';

const result = streamTranscribe({
  model: deepgram.transcription('nova-3'),
  audio: audioStream, // ReadableStream<Uint8Array | string>
  inputAudioFormat: { type: 'audio/pcm', rate: 16000 },
  providerOptions: {
    deepgram: {
      language: 'en',
      streaming: { interimResults: true, endpointing: 300 },
    },
  },
});

for await (const part of result.fullStream) {
  if (part.type === 'transcript-partial' || part.type === 'transcript-final') {
    console.log(part.type, part.text);
  }
}

console.log(await result.text);
console.log(await result.segments);
console.log(await result.durationInSeconds);

Tune providerOptions.deepgram.streaming if needed: endpointing is the silence in milliseconds before an utterance is finalized, and utteranceEndMs (1000 or more) works with interimResults.

Test it

Pipe a short recording of yourself counting to ten. You should see transcript-partial parts change as you speak and a transcript-final part after each pause.

Not confirmed

The function is marked experimental, so its name and options may change in a later release.
The docs do not say whether this path bills differently from normal Deepgram streaming.