Transcription models

Transcription models convert audio bytes into normalized text. They fit upload processors, meeting notes, support-call review, accessibility, and media pipelines.

1. Create a transcription model

Construct the provider model in server-side configuration.

import { OpenAIClient } from '@anvia/openai'

const client = new OpenAIClient({ apiKey })

export const transcriptionModel = client.transcriptionModel({ modelId: 'whisper-1' })

OpenAI, Gemini, and Grok provide v1 transcription adapters. Supported formats, size limits, languages, and optional parameters vary by provider.

2. Read the audio bytes

Validate authorization, media type, and size before loading an upload into memory.

import { readFile } from 'node:fs/promises'

const audioPath = 'support-call.wav'
const audio = await readFile(audioPath)

transcribe() accepts audio bytes as audio.data, a Uint8Array or ArrayBuffer in its options object. Empty audio is rejected before the provider is called.

3. Transcribe the audio

Pass one options object containing audio: { data, filename, mediaType? } and model.

import { transcribe } from '@anvia/core/transcription'
import { transcriptionModel } from './models'

const transcript = await transcribe({
    audio: {
        data: audio,
        filename: audioPath,
        mediaType: 'audio/wav'
    },
    model: transcriptionModel,
    language: 'en',
    prompt: 'Transcribe the customer support call exactly.',
    temperature: 0
})

console.log(transcript.text)

A useful filename helps the provider infer the media type; pass an explicit mediaType on the audio object when the filename is ambiguous. Language, prompt, temperature, provider options, retries, and abortSignal are optional.

4. Handle the result safely

The normalized result contains text and the raw provider response.

await transcripts.save({
  recordingId,
  text: transcript.text,
})

Audio and transcripts may contain private or regulated information. Define access, retention, deletion, redaction, and downstream-use rules before sending transcripts to agents, indexes, analytics, or logs.

For document and image extraction, continue with OCR models.