gpt-realtime-whisper

1
import { experimental_streamTranscribe as streamTranscribe } from 'ai';
2
import { createGateway, gateway } from '@ai-sdk/gateway';
3
import { readFile } from 'node:fs/promises';
4

5
const modelId = 'openai/gpt-realtime-whisper';
6

7
// Mint this on your server, then send only the short-lived token to the client.
8
const { token } = await gateway.experimental_transcription.getToken({
9
  model: modelId,
10
});
11
const clientGateway = createGateway({ apiKey: token });
12

13
// Raw 24 kHz, 16-bit signed little-endian mono PCM audio.
14
const bytes = await readFile('audio.pcm');
15
const audio = new ReadableStream<Uint8Array>({
16
  start(controller) {
17
    controller.enqueue(new Uint8Array(bytes));
18
    controller.close();
19
  },
20
});
21

22
const result = streamTranscribe({
23
  model: clientGateway.transcription(modelId),
24
  audio,
25
  inputAudioFormat: { type: 'audio/pcm', rate: 24000 },
26
});
27

28
for await (const part of result.fullStream) {
29
  if (part.type === 'transcript-delta') {
30
    process.stdout.write(part.delta);
31
  }
32
}
33

34
console.log('\nFinal:', await result.text);